Broad Street Scientific 2018-2019
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
BROAD<br />
STREET<br />
SCIENTIFIC<br />
VOLUME 8 | <strong>2018</strong>-<strong>2019</strong><br />
The North Carolina School of Science and Mathematics Journal of<br />
Student STEM Research
Front Cover<br />
This map segment of Durham features<br />
the North Carolina School of Science and<br />
Mathematics campus (upper-center) as well as<br />
the nearby Duke University campus (lowerleft)<br />
and several historic neighborhoods. Map<br />
data © <strong>2019</strong> Google; Image created by Kathleen<br />
Hablutzel.<br />
Approximate Scale: 6.5 kilometers<br />
Biology Section<br />
This image of the dorsal raphe nucleus labels<br />
dopamine neurons in green, red, and yellow.<br />
This region of the brain is critical in generating<br />
the increased sociability that typically occurs<br />
after a period of social isolation. Image credit<br />
Gillian Matthews, Ungless Lab, Imperial<br />
College London.<br />
Approximate Scale: 450 micrometers<br />
Chemistry Section<br />
This is a transmission electron microscope image<br />
of a graphene lattice. Graphene is a periodic<br />
structure entirely composed of carbon atoms. At<br />
this scale, individual atoms can be observed at<br />
the corners of the hexagons. Image credit Ethan<br />
Minot, Department of Physics, Oregon State<br />
University (Original grayscale image colorized).<br />
Approximate Scale: 4 nanometers
Engineering Section<br />
Visualizing patterns of air traffic over the<br />
contiguous United States reveals major airports<br />
and commonly flown-over regions. The darkest<br />
areas receive little-to-no flyovers. Image credit<br />
Aaron Koblin, Scott Hessels, and Gabriel<br />
Dunne, UCLA.<br />
Approximate Scale: 4500 kilometers<br />
Mathematics and Computer Science Section<br />
The Opte Project aims to visualize the internet<br />
by mapping routing paths from all over the<br />
world. Each color represents computers from a<br />
different region of the world. This visualization<br />
is from 2015. Image credit Barrett Lyon/The<br />
Opte Project.<br />
Approximate Scale: One zettabyte for the<br />
year 2015<br />
Physics Section<br />
The Baryon Oscillation Spectroscopic Survey<br />
(BOSS) Great Wall, a galaxy supercluster, is<br />
one of the largest structures in the observable<br />
universe. This image shows a simulation of how<br />
galaxy clusters form. The filaments are regions<br />
where galaxies are more likely to be found.<br />
Image credit Volker Springel, Max Planck<br />
Institute for Astrophysics.<br />
Approximate Scale: 1 billion light years<br />
Back Cover<br />
<strong>Scientific</strong> collaborations span the globe.<br />
This map depicts collaboration networks<br />
between researchers in different locations.<br />
Collaborations often - but not always - seem<br />
to follow linguisic and cultural connections.<br />
Image computed by Oliver H. Beauchesne and<br />
SCImago Lab, data by Elsevier Scopus.<br />
Approximate Scale: 39,000 kilometers
TABLE of CONTENTS<br />
4 Letter from the Chancellor<br />
5 Words from the Editors<br />
6 <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> Staff<br />
7 Essay: The AI We Haven't Considered<br />
JACKSON MEADE, 2020<br />
Biology<br />
10 Overexpression of a Heat Shock Protein in Cyanobacteria to Increase Growth Rate<br />
ROBERT LANDRY, <strong>2019</strong><br />
18 Hypoglycemic Effect of Momordica charantia Against Type 2 Diabetes Modeled in Bombyx mori<br />
AARUSHI VENKATAKRISHNAN, <strong>2019</strong><br />
Chemistry<br />
26 Tetraethyl Orthosilicate-Polyacrylonitrile Hybrid Membranes and their Application in Redox<br />
Flow Batteries<br />
ETHAN FREY, <strong>2019</strong><br />
32 Novel Synergistic Antioxidative Interactions Between Soy Lecithin and Cyclodextrin-<br />
Encapsulated Quercetin in a Lipid Matrix<br />
ANIRUDH HARI, <strong>2019</strong><br />
37 Utilization of Atomic Layer Deposition to Create Novel Metal Oxide Photoanodes for Solar-<br />
Driven Water Splitting<br />
ANNIE WANG, <strong>2019</strong>
Engineering<br />
44 Using a Hybrid Machine Learning Approach for Test Cost Optimization in Scan Chain Testing<br />
LUKE DUAN, <strong>2019</strong><br />
49 Novel Water Desalination Filter Utilizing Granular Activated Carbon<br />
GEOFFREY FYLAK, <strong>2019</strong><br />
Mathematics and Computer Science<br />
59 Long Prime Juggling Patterns<br />
DANIEL CARTER AND ZACH HUNTER, <strong>2019</strong><br />
67 An Analysis of a Novel Neural Network Architecture<br />
VATSAL VARMA, <strong>2019</strong> ONLINE<br />
Physics<br />
75 Effects of Relativity on Quadrupole Oscillations of Compact Stars<br />
ABHIJIT GUPTA, <strong>2019</strong><br />
84 Effect of Elliptic Flow Fluctuations on the Two- and Four-Particle Azimuthal Cumulant<br />
BRIAN LIN, <strong>2019</strong><br />
Featured Article<br />
89 An Interview with Dr. Valerie Ashby
LETTER from the CHANCELLOR<br />
"Science is a cooperative enterprise, spanning the generations. It's the passing of a torch from teacher, to student, to<br />
teacher. A community of minds reaching back to antiquity and forward to the stars."<br />
~ Dr. Neil deGrasse Tyson<br />
I am proud to introduce the eighth edition of the<br />
North Carolina School of Science and Mathematics’<br />
(NCSSM) scientific journal, <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>. Each<br />
year students at NCSSM conduct significant scientific<br />
research, and <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> is a student-led and<br />
student-produced showcase of some of the impressive<br />
research being done by students.<br />
Excellence in scientific research has a deep and<br />
far-reaching impact on nearly every aspect of daily life,<br />
including (among other areas) health care, food safety,<br />
space travel, national security, and the environment.<br />
When NCSSM students are given opportunities to apply<br />
their learning through research, they are doing more than<br />
increasing their individual knowledge; their valuable<br />
work is increasing our collective body of knowledge<br />
and strengthening our ability to address current global<br />
challenges and prepare for those to come.<br />
Opened in 1980, NCSSM was the nation’s first public<br />
residential high school where students study a specialized<br />
curriculum emphasizing science and mathematics.<br />
Teaching students to do research and providing them with<br />
opportunities to conduct high-level research in biology,<br />
chemistry, physics, computational science, engineering<br />
and computer science, math, humanities, and the social<br />
sciences are critical components of NCSSM’s mission<br />
to educate academically talented students to become<br />
state, national, and global leaders in science, technology,<br />
engineering, and mathematics. I am thrilled that each year<br />
we continue to increase the outstanding opportunities<br />
NCSSM students have to participate in research.<br />
This publication serves to highlight some of the high<br />
quality research students conduct each year at NCSSM<br />
under the direction of our outstanding faculty and in<br />
collaboration with researchers at major universities. For<br />
thirty-four years, NCSSM has showcased student research<br />
through our annual Research Symposium each spring and<br />
at major research competitions such as the Regeneron<br />
Science Talent Search and the International Science and<br />
Engineering Fair. The publication of <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong><br />
provides another opportunity to share with the broader<br />
community the outstanding research being conducted by<br />
NCSSM students each year.<br />
I would like to thank all of the students and faculty<br />
involved in producing <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>, particularly<br />
faculty sponsor Dr. Jonathan Bennett, and senior editors<br />
Emily Wang, Navami Jain, and Kathleen Hablutzel.<br />
Explore and enjoy!<br />
Dr. Todd Roberts, Chancellor<br />
4 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>
WORDS from the EDITORS<br />
Welcome to the <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>, NCSSM’s journal<br />
of student research in science, technology, engineering<br />
and mathematics. In this eighth edition of <strong>Broad</strong> <strong>Street</strong><br />
<strong>Scientific</strong>, we hope to inspire readers to get involved in the<br />
scientific community by sharing the innovative research<br />
conducted by our students. We hope you enjoy this year’s<br />
edition!<br />
This year’s theme is networks: the connections we<br />
find within and between groups throughout our world.<br />
Connectivity is an integral component of modern life,<br />
and studying people or objects interacting in networks<br />
allows us to describe collective behavior of groups.<br />
Billions of interconnected neurons comprise the human<br />
brain, yet a brain is more than a bunch of cells. Brains can<br />
think, feel, and act both consciously and unconsciously.<br />
Thus, networks do not simply behave as the sum of their<br />
parts. Networks are powerful in predicting the complex<br />
behavior of a dynamic group without needing complex<br />
information on each individual in a network. For example,<br />
networks can predict the spread of an infectious disease<br />
without needing information on each individual in the<br />
network. Networks are powerful tools in describing our<br />
interconnected world.<br />
In the featured images of this journal, we explore the<br />
scales of networks, from the atomic to astronomical levels.<br />
The featured image for the Chemistry section displays<br />
a network of carbon atoms on the scale of fractions of<br />
a nanometer, while the featured image for the Physics<br />
section displays a network of superclusters of galaxies on<br />
the scale of approximately one billion light years – one of<br />
the largest known structures in the universe. On any scale,<br />
our world is built on interactions, and these interactions<br />
organize our world into networks.<br />
We would like to thank the faculty, staff and<br />
administration of NCSSM for their continued support<br />
towards our student researchers. It is this unmatched<br />
encouragement that prepares us to use our interests and<br />
skills in STEM to address problems in our community,<br />
both locally and beyond. For 39 years, NCSSM has<br />
fostered an environment conducive to learning through<br />
encouraging students to take risks and take ownership of<br />
their academic path. We would especially like to thank our<br />
faculty advisor, Dr. Jonathan Bennett, for his support and<br />
guidance throughout the publication process. We would<br />
also like to thank Chancellor Dr. Todd Roberts, Dean of<br />
Science Dr. Amy Sheck, and Director of Mentorship and<br />
Research Dr. Sarah Shoemaker. Lastly, the <strong>Broad</strong> <strong>Street</strong><br />
<strong>Scientific</strong> would like to acknowledge Dr. Valerie Ashby,<br />
chemistry professor and Dean of Trinity College of Arts<br />
and Sciences at Duke University, for speaking with us<br />
about her inspiring journey in STEM and offering advice<br />
to young prospective scientists.<br />
Kathleen Hablutzel, Navami Jain, and Emily Wang<br />
Editors-in-Chief<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 5
BROAD STREET SCIENTIFIC STAFF<br />
Editors-in-Chief<br />
Kathleen Hablutzel, <strong>2019</strong><br />
Navami Jain, <strong>2019</strong><br />
Emily Wang, <strong>2019</strong><br />
Publication Editors<br />
Rohit Jagga, 2020<br />
Grishma Patel, <strong>2019</strong><br />
Sanjana Pothugunta, 2020<br />
Eleanor Xiao, 2020<br />
Biology Editors<br />
Megan Wu, <strong>2019</strong><br />
Ishaan Maitra, 2020<br />
Joseph Wang, 2020<br />
Chemistry Editors<br />
Melody Wen, 2020<br />
Varun Varanasi, 2020<br />
Engineering Editors<br />
Aakash Kothapally, 2020<br />
Jason Li, 2020<br />
Mathematics and<br />
Computer Science Editors<br />
Hahn Lheem, <strong>2019</strong><br />
Olivia Fugikawa, 2020<br />
Physics Editors<br />
Will Staples, 2020<br />
Ben Wu, 2020<br />
Faculty Advisor<br />
Dr. Jonathan Bennett<br />
6 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>
THE AI WE HAVEN'T CONSIDERED<br />
Jackson Meade<br />
Jackson Meade was selected as the winner of the <strong>2018</strong>-<strong>2019</strong> <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> Essay Contest. His award included the<br />
opportunity to interview Dr. Valerie Ashby, distinguished chemist and professor and Dean of Trinity College of Arts and<br />
Sciences at Duke University. This interview can be found in the Featured Article section of the journal.<br />
“People worry that computers will get too smart and take over the world, but the real problem is that they’re too stupid and they’ve<br />
already taken over the world.”<br />
~ Pedro Domingos<br />
When we bring up artificial intelligence in<br />
conversation, the rhetoric is relatively future-oriented.<br />
Discussions about the “possibilities” AI possesses – and<br />
the dangers it poses – abound, all in the context of what<br />
our technological future holds. But peel back the layer of<br />
speculation, and you may find something surprising. It<br />
might not be obvious, but artificial intelligence is already<br />
here – in fact, it’s everywhere.<br />
Though that statement sounds concerning, there isn’t<br />
a conspiracy of shadowy Artificial Intelligences operating<br />
behind the backs of the public. We’ve simply grown<br />
accustomed to its cohabitation in our systems. Artificial<br />
Intelligence, through “Machine Learning,” started<br />
accelerating in 1957, when Frank Rosenblatt designed<br />
the first Neural Network, called a perceptron (Lewis &<br />
Denning), to model the structure of the human brain<br />
(Marr). By 1985, Professor Terry Sejnowski had created<br />
NetTalk, which could pronounce 20,000 English words<br />
with just a week of training (New York Times).<br />
When you flip through a stuffed email inbox, machine<br />
learning keeps it from exploding by marking most of the<br />
spam and trashing it, arguably with impressive precision<br />
(Aski & Sourati). Go to your search engine and type your<br />
query, and the “suggested search” bar that appears at the<br />
bottom, as well as the results your query generates, are the<br />
product of a well-trained, personalized machine learning<br />
algorithm (Schachinger). When you purchase something<br />
on Amazon or scroll through your recommended videos<br />
page on YouTube, a machine learning system makes sure<br />
you see the kinds of things you might want to watch or<br />
buy, even if you couldn’t articulate it yourself. If you are<br />
looking at a screen, it is likely that machine learning had<br />
its hands (for lack of a more computerized term) in it.<br />
Since the early days of computing, computers have<br />
required painstaking algorithms – increasing by orders of<br />
magnitude in complexity – to do anything from displaying<br />
text to managing Google’s 40,000 search queries per<br />
second (Alphabet, Inc). This creates a ceiling of capabilities<br />
for our systems that grows infinitely harder to raise. But<br />
computer systems are, after all, human systems, and we<br />
should model them that way. This is exactly what machine<br />
learning algorithms do. Based on inferences from the data<br />
we give them, they teach themselves how to analyze and<br />
manipulate it, and the more data we give them, the better<br />
they get at doing their jobs (Faggella) (Lewis & Denning).<br />
This is significantly “human” – barring willful ignorance,<br />
we get better at analyzing and understanding our world<br />
given new information.<br />
Despite possible concerns, you are kept safe because of<br />
machine learning. In 2014, Kaspersky Lab's Anti-Malware<br />
Research Team processed between 200,000 and 315,000<br />
malicious files per day (Kaspersky Lab). But malicious files<br />
aren’t so different from each other, so machine learning<br />
algorithms can very easily identify the code for files with<br />
malicious intent far faster than any human actors could. In<br />
a country and world growing ever more concerned with<br />
data security, these algorithms provide a necessary wall<br />
between us and the actions of evil people.<br />
In our finances, we’re relinquishing control to the<br />
machines as well. Micromanagement of our funds is a<br />
multibillion-dollar business, and artificial intelligence<br />
completely disrupts it. While humans are good at predicting<br />
what the stock market can do over large spans of time<br />
because of noticeable trends, on smaller and smaller time<br />
scales and in more volatile markets, our grand spending<br />
schemes are fundamentally nothing short of guesswork.<br />
And while machine learning algorithms are admittedly<br />
built on guesswork, they can achieve super-human levels<br />
of accuracy during training on multitudes of data that are<br />
simply unattainable for even the most dedicated human.<br />
Predicting stocks is not the only artificial-intelligenceguided<br />
moneymaker around. Advertising is one of the most<br />
lucrative businesses of the modern world, having generated<br />
about $32.66 billion dollars in revenue for Google’s parent<br />
company, Alphabet, in Quarter 2 of <strong>2018</strong> (D’Onfro). This<br />
comes from thousands of paying customers, all of them<br />
companies hoping their product appeals to the right niche,<br />
and only works because of machine learning.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 7
In this realm, one cannot avoid the topic of driverless<br />
cars. Artificial intelligences are crucial to computer vision<br />
algorithms (Khirodkar, Yoo, & Kitani), though other<br />
hard-coded solutions can aid them. 74% of automotive<br />
company executives expect that these smart cars will be on<br />
the road by 2025, according to a report from IBM (IBM).<br />
The menial tasks of our lives – our driving, our purchases<br />
– will be automated if they can be.<br />
We have been exploring the risks of developing artificial<br />
intelligences prior to the day we could make them. In 1942,<br />
science fiction author Isaac Asimov published his nowfamous<br />
laws of robotics in a short story, “Runaround.”<br />
They stated:<br />
First, “A robot may not injure a human being or,<br />
through inaction, allow a human being to come to harm.”<br />
Second, “A robot must obey the orders given to it by<br />
human beings, except where such orders would conflict<br />
with the First Law.<br />
Third, “A robot must protect its own existence as long<br />
as such protection does not conflict with the First or<br />
Second Law.”<br />
But restricting our concerns about Artificial Intelligence<br />
to this view is too narrow. It comes from an assumption<br />
about the types of intelligences we intend to create. It<br />
assumes that we will “build ourselves” – that we will build<br />
copies of humans, in humanoid robot bodies with human<br />
emotions and human capabilities.<br />
We are a species that changes its environment to fit its<br />
needs instead of adapting to its surroundings. Machine<br />
learning and artificial intelligence are the newest evolution<br />
of this pattern – just another way that the world and the<br />
patterns within it can be adjusted according to our wishes.<br />
The patterns of our world once influenced us to a degree<br />
we could not control, but artificial intelligence will allow<br />
us to take full control and then completely relinquish it. All<br />
this works because machine learning is based in prediction<br />
– on understanding the once-unintelligible patterns that<br />
comprise the fabric of our world. That is a flaw that could<br />
spell the end of our humanity.<br />
It seems unlikely a malicious AI will attempt to literally<br />
end life on Earth. At the least, we have Asimov’s three laws<br />
to thank for that. But in a world where everything can be<br />
predicted, where everything we want to see can be shown<br />
to us, and where things that are “unpopular” or “troubling”<br />
never reach our eyes, it feels like a part of our humanity<br />
is lost. An artificial intelligence could operate in plain<br />
sight, tailoring our world to the patterns that dictate us.<br />
As mentioned, artificial intelligences are human systems,<br />
so they will follow the human model of changing the world<br />
to fit their needs. It is reasonable that if our needs rely on a<br />
series of predictable patterns, then an artificial intelligence<br />
with benevolent intentions could inadvertently neutralize<br />
the world’s ideological diversity and the differences that<br />
give us the human condition.<br />
This isn’t to say that we shouldn’t create artificial<br />
intelligences – in fact, it seems clear that our modern<br />
world couldn’t operate without them. There are proactive<br />
steps we must take to be stewards of our humanity. We<br />
must make an active choice to diversify our interests<br />
and the viewpoints to which we expose ourselves, even<br />
when they aren’t completely satisfying. We should model<br />
another fundamental element of our humanity into our<br />
artificial intelligences: variation. Our machine learning<br />
algorithms cannot rely on optimizing patterns alone – they<br />
must contain anomalies in their paradoxically predictive,<br />
average-based algorithms. If we do this, we can ensure<br />
that our artificial intelligences will enhance us instead of<br />
dictating conformity.<br />
References<br />
A Q & A with Pedro Domingos: Author of ‘The Master<br />
Algorithm’ [Interview by J. Langston]. (2015, September<br />
17). Retrieved January 4, <strong>2019</strong>, from https://www.<br />
washington.edu/news/2015/09/17/a-q-a-with-pedrodomingos-author-of-the-master-algorithm/<br />
Aski, A. S., & Sourati, N. K. (2016). Proposed efficient<br />
algorithm to filter spam using machine learning<br />
techniques. Pacific Science Review A: Natural Science<br />
and Engineering, 18(2), 145-149. doi:10.1016/j.<br />
psra.2016.09.017<br />
D’Onfro, J. (<strong>2018</strong>, July 23). Alphabet jumps after big<br />
earnings beat. Retrieved January 7, <strong>2019</strong>, from https://<br />
www.cnbc.com/<strong>2018</strong>/07/23/alphabet-earnings-q2-<strong>2018</strong>.<br />
html<br />
Faggella, D. (<strong>2018</strong>, December 21). What is Machine<br />
Learning? | Emerj - Artificial Intelligence Research and<br />
Insight. Retrieved January 9, <strong>2019</strong>, from https://emerj.<br />
com/ai-glossary-terms/what-is-machine-learning/<br />
Google Search Trends, Search Per Second. (n.d.).<br />
Retrieved January 12, <strong>2019</strong>, from https://trends.google.<br />
com/trends/?geo=US<br />
IBM (2015). Automotive 2025: Industry without Borders.<br />
IBM Institute for Business Value. Retrieved January<br />
9, <strong>2019</strong>, from http://www-935.ibm.com/services/<br />
multimedia/GBE03640USEN.pdf<br />
Kaspersky Lab is Detecting 325,000 New Malicious<br />
Files Every Day. (n.d.). Retrieved January 5, <strong>2019</strong>, from<br />
https://www.kaspersky.com/about/press-releases/2014_<br />
kaspersky-lab-is-detecting-325000-new-malicious-filesevery-day<br />
8 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong>
Khirodkar, R., Yoo, D., & Kitani, K. M. (<strong>2018</strong>).<br />
VADRA: Visual Adversarial Domain Randomization<br />
and Augmentation. Carnegie Mellon University.<br />
Retrieved December 10, <strong>2018</strong>, from https://arxiv.org/<br />
pdf/1812.00491.pdf<br />
Learning, Then Talking. (1988, August 16). Retrieved<br />
January 6, <strong>2019</strong>, from https://www.nytimes.<br />
com/1988/08/16/science/learning-then-talking.html<br />
Lewis, T. G., & Denning, P. J. (<strong>2018</strong>). The Profession of<br />
IT: Learning Machine Learning. Communications of the<br />
ACM, 61(12), 24-27. Retrieved December 26, <strong>2018</strong>, from<br />
https://calhoun.nps.edu/bitstream/handle/10945/60898/<br />
Denning_Learning_Machine_Learning_ACM_<strong>2018</strong>-12.<br />
pdf?sequence=1&isAllowed=y.<br />
Levenson, E. (2014, January 31). The TSA is in the<br />
Business of ‘Security Theater,’ Not Security. Retrieved<br />
January 7, <strong>2019</strong>, from https://www.theatlantic.com/<br />
national/archive/2014/01/tsa-business-security-theaternot-security/357599/<br />
Marr, B. (2016, March 08). A Short History of Machine<br />
Learning -- Every Manager Should Read. Retrieved<br />
January 5, <strong>2019</strong>, from https://www.forbes.com/sites/<br />
bernardmarr/2016/02/19/a-short-history-of-machinelearning-every-manager-should-read/#2493077c15e7<br />
Matney, L. (2017, May 17). Google has 2 billion users on<br />
Android, 500M on Google Photos. Retrieved January 5,<br />
<strong>2019</strong>, from https://techcrunch.com/2017/05/17/googlehas-2-billion-users-on-android-500m-on-google-photos/<br />
Schachinger, K. (<strong>2018</strong>, December 06). A Complete Guide<br />
to the Google RankBrain Algorithm. Retrieved January<br />
4, <strong>2019</strong>, from https://www.searchenginejournal.com/<br />
google-algorithm-history/rankbrain/<br />
Scott, T. (<strong>2018</strong>, December 06). Retrieved January 13,<br />
<strong>2019</strong>, from https://www.youtube.com/watch?v=-<br />
JlxuQ7tPgQ<br />
Wu, J., Zhang, C., Xue, T., Freeman, W. T., &<br />
Tenenbaum, J. B. (2016). Learning a Probabilistic Latent<br />
Space of Object Shapes via 3D Generative-Adversarial<br />
Modeling. Advances In Neural Information Processing<br />
Systems, 29. Retrieved November 11, <strong>2018</strong>, from https://<br />
arxiv.org/abs/1610.07584.<br />
Yoganarasimhan, H. (2017). Search Personalization<br />
Using Machine Learning. SSRN Electronic Journal.<br />
doi:10.2139/ssrn.2590020<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 9
OVEREXPRESSION OF A HEAT SHOCK PROTEIN IN<br />
CYANOBACTERIA TO INCREASE GROWTH RATE<br />
Robert Landry<br />
Abstract<br />
To increase earth’s capacity to support human population growth, methods of growing food more efficiently, especially<br />
in warmer environments as climate change progresses, must be developed. This project sought to increase the growth<br />
rate of one population of photosynthetic organisms, cyanobacteria, through genetic engineering. Synechococcus elongatus<br />
UTEX 2973 cultures were transformed to overexpress dnaJ, a heat shock protein, in normal and heat-stressed conditions<br />
to determine the gene’s effects on growth rates. The growth rates of the dnaJ overexpressing strain were related to the<br />
control--wild-type Synechococcus elongatus UTEX 2973 transformed with a plasmid without dnaJ--through comparisons<br />
of optical density measurements at 745 nanometers (OD745), which can accurately quantify growth rates. The change<br />
in OD745 in the dnaJ overexpressing strain was significantly greater than the OD745 measurements for the control in<br />
normal conditions. When the temperature was increased to 42˚C, the dnaJ overexpressing strain continued to grow,<br />
while the control strain’s OD745 measurements decreased. From this data, it appeared that the overexpression of a heat<br />
shock protein in the genome of cyanobacteria significantly increased their growth rates and provided heat resistance.<br />
Researching the effects of overexpressing a heat shock protein could be furthered in organisms such as corn, rice, soybeans,<br />
and other photosynthetic species.<br />
1. Introduction<br />
Cyanobacteria, bacteria that conduct photosynthesis,<br />
have the potential to revolutionize both agricultural<br />
practices and the food industry, if higher yields of target<br />
materials are attained (Chow et al., 2015). Cyanobacteria,<br />
capable of utilizing 10% of the sun’s energy, are nearly 10<br />
times more efficient at fixing carbon found in CO2 than<br />
other energy plants such as sugar cane or corn, which<br />
harness only 1% of the sun’s energy (Hunt, 2003). This<br />
efficiency drives cyanobacteria into the energy industry’s<br />
spotlight as a possible, influential source of energy for<br />
humanity. Moreover, their increased photosynthetic rates<br />
decrease the amount of CO2 in the atmosphere, which<br />
benefits the global environment. Five other aspects of<br />
these photosynthetic bacteria that interest scientists are<br />
that they: grow in high densities, use water as an electron<br />
donor, utilize infertile land, require non-food-based<br />
feedstock, and thrive in many different water conditions<br />
(brackish, fresh, or saltwater) (Parmar et al., 2015).<br />
Although all of these benefits already apply to<br />
cyanobacteria, it is still expensive to culture, grow, and<br />
eventually utilize the products of the bacteria in an<br />
efficient way. In order for cyanobacteria to be widely used,<br />
a sharp increase in target yields and decrease in expense<br />
must occur in order to compete with the simplicity and<br />
economic benefits of plants.<br />
Coupled with being more cost-effective when producing<br />
target materials, plants have also been genetically modified<br />
with genes originating from cyanobacteria to increase<br />
efficiency. For example, carbon fixation rates in transgenic<br />
tobacco were increased significantly after transforming<br />
cyanobacterial Rubisco into the tobacco’s genome.<br />
Photosynthetic efficiency was increased as a result of<br />
cyanobacteria’s efficiency, which serves as a precedent for<br />
future research (Occhialini et al., 2015).<br />
This transgenic tobacco demonstrates the viability of<br />
increasing the efficiency of plant growth with cyanobacteria<br />
research. This pursuit is important because scientists of<br />
the Global Harvest Institute estimate that the world could<br />
face a food crisis by 2030 (Martin, 2017). Developing new<br />
methods of growing crops is paramount to mitigating this<br />
impending humanitarian need.<br />
In recent decades, knowledge regarding cyanobacteria<br />
has increased exponentially, stemming first from the<br />
genome-mapping of Synechocystis sp. 6803, one species<br />
of cyanobacteria. Now there are more than 128 different<br />
strains of cyanobacteria fully sequenced, which provides<br />
many opportunities in genetic engineering to study<br />
the properties of the bacteria. This developing field<br />
of genetic engineering allows researchers to utilize<br />
various transformation techniques in order to optimize<br />
photosynthetic rates within cyanobacteria, and ultimately<br />
in other organisms as well (Al-Haj et al., 2016).<br />
The species Synechococcus elongatus PCC7942 is one<br />
species of cyanobacteria that has had its entire genome<br />
sequenced and therefore is a candidate for many genetic<br />
engineering projects that study photosynthetic processes,<br />
regulation of nitrogen-containing compounds, and<br />
acclimation to stressed conditions (Home - Synechococcus<br />
elongatus PCC 7942). Synechococcus elongatus PCC7942,<br />
previously known as Anacystis nidulans R2, is a freshwater<br />
cyanobacteria that was the first cyanobacteria to be<br />
successfully transformed using exogenous DNA (Shestakov<br />
10 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
Table 1. The list of forward and reverse primers used for isolating dnaJ.<br />
Species Gene Direction Melting<br />
Temperature<br />
(°C)<br />
Synechococcus elongatus<br />
UTEX L 2973<br />
Sequence (5’-3’)<br />
dnaJ Forward 69.2 GAGAATTCATGGGTC-<br />
GTCGCTGGA<br />
Purpose<br />
Transformation<br />
Synechococcus elongatus<br />
UTEX L 2973<br />
dnaJ Reverse 68.19 GAGGATCCCTAGCATG-<br />
CAAGCTCTCCTG<br />
Transformation<br />
Synechococcus elongatus<br />
UTEX L 2973<br />
Synechococcus elongatus<br />
UTEX L 2973<br />
dnaJ Forward 68.16 ATGCAAAATTTTCGC-<br />
GACTACTATGCC<br />
dnaJ Reverse 67.47 TCAACGCGATTGTTC-<br />
GAGCGAT<br />
RT-PCR<br />
RT-PCR<br />
& Khyen, 1970). Synechococcus elongatus PCC7942 are<br />
obligate photoautotrophs, which means that they only<br />
rely on their photosynthetic ability to produce nutrients<br />
instead of being able to break down and use nutrients found<br />
in their environment (Minda et al., 2008). Due to this<br />
attribute, Synechococcus elongatus PCC7942’s photosynthetic<br />
efficiency must be optimized for any condition, including<br />
stress, to outlast their natural competition. One such way<br />
that Synechococcus elongatus PCC7942 has been shown<br />
to adapt to extreme heat and high light conditions is the<br />
induction of the dnaK and dnaJ genes (Hihara et al., 2001).<br />
The gene dnaK has three different homologues found<br />
in the genome of Synechococcus elongatus PCC7942,<br />
designated dnaK1, dnaK2, and dnaK3. DnaK1’s function is<br />
unknown in the Synechococcus elongatus PCC7942, although<br />
it is known to be found in the cytosol of the cyanobacteria.<br />
Both dnaK2 and dnaK3 are essential for the growth of<br />
Synechococcus elongatus PCC7942 (Watanabe, 2007). Similar<br />
to dnaK, dnaJ has 4 homologues within the Synechococcus<br />
elongatus PCC7942 genome, referred to as dnaJ1, dnaJ2,<br />
dnaJ3, and dnaJ4. DnaJ3 has been found to be located in<br />
the membrane of the cyanobacteria. DnaJ2 is shown to be<br />
induced in extreme heat and high light conditions. Apart<br />
from these two homologues, dnaJ2 and dnaJ3, most of<br />
dnaJ roles in the cell have not been discovered (Shestakov<br />
& Khyen, 1970). The substitute for Synechococcus elongatus<br />
PCC7492 that will be used in this experiment due to<br />
budget constraints is Synechococcus elongatus UTEX L 2973.<br />
Within Synechococcus elongatus UTEX L 2973, there are<br />
10 homologues of dnaJ (Genome). Their respective roles<br />
within the cell beyond molecular chaperones are largely<br />
unknown, apart from dnaJ3, which is a known heat<br />
shock protein (Genome). The third homologue of dnaJ<br />
was isolated and overexpressed in a transformed strain of<br />
cyanobacteria in this research project.<br />
The goal of this project is to determine the effects of<br />
dnaJ on the photosynthetic rate of Synechococcus elongatus<br />
UTEX L 2973 and explore the correlation between the<br />
genes’ overexpression and growth rates in various heat<br />
conditions. This research could lead to new advancements<br />
in industry and agriculture through the higher production<br />
rates of glucose and target materials.<br />
2. Methods<br />
2.1 – Culturing Synechococcus elongatus UTEX L 2973<br />
Synechococcus elongatus UTEX L 2973 thrive in BG-11<br />
liquid medium at 30°C under 12-hour light cycles from a<br />
Percival Incubator. Once the cyanobacteria showed initial<br />
growth in the medium, the bacteria were aliquoted to more<br />
containers to protect the Synechococcus elongatus UTEX L<br />
2973 from contamination that could ruin the whole strain<br />
(Kufryk et al., 2002).<br />
2.2 – DNA extraction, PCR and RT-PCR<br />
The DNA from Synechococcus elongatus UTEX L 2973<br />
was extracted using the QIAamp DNA Mini Kit and its<br />
corresponding protocol (QIAGEN). Using the primers<br />
listed in Table 1, dnaJ was isolated including the restriction<br />
enzyme cut sites necessary for ligation. The PCR was run<br />
according to the OneTaq Hot Start protocol (Biolabs). The<br />
extension phase lasted for 2 minutes and the annealing<br />
temperature was 61°C.<br />
2.3 – Cloning<br />
Using BamHI and EcoRI restriction enzymes, dnaJ<br />
was ligated into the plasmid pSyn_6 from a GeneArt<br />
Synechococcus Engineering Kit.<br />
2.4 – Transformation of E. Coli<br />
A 5-alpha E. coli strain was transformed using the<br />
heat shock method to replicate the desired plasmid. Two<br />
different plasmids were used to transform the E. coli, one<br />
vector without dnaJ and one plasmid including dnaJ. After<br />
transformation, the E. coli grew in SOC medium, which<br />
was then spread on LB plates with spectinomycin at 50<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 11
μg/mL concentration. After growing overnight, colonies<br />
were labeled and were inoculated into tubes corresponding<br />
to their label to grow overnight.<br />
2.5 – Transformation of Synechococcus Elongatus UTEX L<br />
2973<br />
The plasmid DNA from the E. Coli was extracted using<br />
a Spin Miniprep Kit and its corresponding protocol. This<br />
plasmid DNA was then used to transform Synechococcus<br />
Elongatus UTEX L 2973 following the protocol provided<br />
by GeneArt Synechococcus Engineering Kit. This vector<br />
has not been used to transform pSyn_6 before.<br />
2.6 – Statistical Analysis<br />
To analyze the OD745 data, error bars were calculated<br />
by multiplying the standard error of the mean by two. To<br />
test significance, a t-test calculator for the comparison of<br />
means was used to determine a p-value. One asterisk (*)<br />
represents significance at a p-value of < .05; two asterisks<br />
(**) concludes significance at a p-value of < .01; three<br />
asterisks (***) demonstrates that the data are significant at<br />
a p-value of < .005<br />
show the successful isolation of dnaJ, a gene of length<br />
1.8kb (Fig. 2a).<br />
Figure 2a. Successful PCR amplification of dnaJ. The<br />
band at 1.8kb is dnaJ.<br />
Figure 2b. The cutout portion of the gel isolated the<br />
vector that was used in gel extraction and ligation.<br />
Figure 2c. Cutouts from the gel isolated dnaJ.<br />
These bands were ligated into the plasmid after gel<br />
extraction.<br />
Figure 1. dnaJ will be inserted in between EcoRI and<br />
BAMHI<br />
3. Results<br />
3.1 – Cloning Strategy<br />
A cloning strategy was used (Fig. 1). DnaJ was isolated<br />
including the restriction enzyme cut sites necessary for<br />
ligation using the aforementioned primers (Table 1). The<br />
enzymes cut the target gene at the lines on either side of<br />
dnaJ (Fig. 1). The vector for GeneArt also had the same<br />
two restriction enzyme cut sites, BamHI and EcoRI, as the<br />
isolated gene. Utilizing DNA Ligase, the dnaJ was inserted<br />
into the plasmid in the 5’-3’ direction with a constitutively<br />
active promoter, PpsaA (NEB).<br />
The bands around 1.8kb surrounded by the red boxes<br />
Figure 2d. Gel electrophoresis of restriction enzyme<br />
digested plasmid from transformed E. coli. The bands<br />
at 1.8kb and 4.5 kb in lane 7 demonstrate correct<br />
ligation and transformation of the E. coli colony.<br />
Plasmid from this colony was used to transform<br />
Synechococcus elongatus UTEX L 2973.<br />
12 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
Following the preliminary PCR, another PCR reaction<br />
was run, and its products were cut with restriction enzymes<br />
before being exposed to ultraviolet light. Ultraviolet is a<br />
known DNA mutagen and hence, exposing dnaJ to this<br />
light before transformation in the cyanobacteria could<br />
alter its natural use in the cell. The vector was also cut<br />
with the restriction enzymes, BamHI and EcoRI. These<br />
two cut products were run through gels to purify the<br />
digested DNA (Fig 2b & 2c). Once the products were<br />
cut out, gel extraction was run to purify the DNA from<br />
the gel, so that ligation could be run (QIAquick). After<br />
the ligated plasmid was formed using DNA Ligase and a<br />
ligation buffer, the E. coli strain 5-alpha from New England<br />
Biolabs was transformed using heat-shock method (Fig.<br />
1). 1, 3, and 6 μL of extracted DNA solution were added<br />
into separate vials of transformation-competent E. coli<br />
cells and were mixed gently. This mixture was put on ice<br />
for 30 minutes then heat-shocked at 42°C for 30 seconds<br />
without shaking. The transformed E. coli was put on ice for<br />
2 minutes. 250 μL of room temperature SOC medium was<br />
added to the vial of E. coli. This vial was incubated shaking<br />
horizontally at 55 rpms at 37°C for 1 hour. Following<br />
the incubation, the various tubes of transformed E. coli<br />
were plated on separate solid LB medium plates with<br />
spectinomycin at a concentration of 50 μg/mL. The plates<br />
were incubated overnight at 37°C. Because of the presence<br />
of spectinomycin on the plates and the plasmid’s resistance<br />
to spectinomycin, the colonies that grew on the plate<br />
overnight had to contain the target plasmid.<br />
Once the colonies formed, 12 colonies were isolated<br />
and grown individually in 3.0 mL of LB medium with<br />
spectinomycin at a concentration of 50 μg/mL overnight.<br />
The plasmid was isolated from these vials of transformed<br />
E. coli using the Spin Miniprep Kit (QIAprep). The plasmid<br />
was digested by EcoRI and BamHI. Gel electrophoresis<br />
was conducted to determine whether or not the plasmid<br />
incorporated the target gene properly (Fig. 2d). Culture<br />
6 replicated the desired plasmid as seen by the bands at<br />
4.5kb and 1.8kb, so the remaining plasmid that was not<br />
run through the gel was used to transform Synechococcus<br />
elongatus UTEX L 2973.<br />
Figure 3a. Two colonies transformed with only<br />
vector in the presence of 10 μg/mL spectinomycin.<br />
Figure 3b. Six colonies overexpressing dnaJ in the<br />
selective presence of spectinomycin.<br />
3.2 – Transformation<br />
The cyanobacteria were transformed using the protocol<br />
corresponding to the GeneArt Synechococcus Engineering<br />
Kit. Following transformation, the cyanobacteria<br />
were plated on solid BG-11 media with 10 μg/mL<br />
spectinomycin under normal conditions. Colonies formed<br />
and overexpressed dnaJ, and those that were transformed<br />
with only pSyn_6 plasmid grew (Fig. 3a & 3b). All of<br />
these colonies were numbered and then inoculated into<br />
flasks containing liquid BG-11 media with 10 µg/mL<br />
spectinomycin.<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 13
3.3 – Growth Assays<br />
tested, it appeared as if the overexpression resulted in<br />
increased rates in both conditions (Fig. 5a & 5b).<br />
Figure 4a. Flasks with varying optical densities. The<br />
three flasks on the left were cultured for the longest<br />
time and thus, had the highest optical densities. The<br />
flasks on the right had grown more recently and<br />
were not as dark.<br />
Figure 4b. Optical density values of corollary flasks.<br />
Higher optical densities correspond to darker flasks.<br />
In order to determine the effects of dnaJ’s overexpression<br />
on growth rates within cyanobacteria, optical density<br />
measurements were taken from different cultures at 745<br />
nanometers (nm) at varying temperatures. This is an<br />
appropriate wavelength because optical density measures<br />
turbidity instead of absorbance. The absorbance of the<br />
selected wavelength should be negligible in order for the<br />
measurements to strictly account for the reflection of light<br />
off of the cells in the solution (Martin, 2014).<br />
There were 7 flasks of cyanobacteria before<br />
transformation with varying optical densities (Fig. 4a).<br />
In order to demonstrate what color and darkness of flasks<br />
correlates to OD745 values, optical density measurements<br />
corresponding to cyanobacteria culture were graphed (Fig.<br />
4b). From the left to right, the optical density values were:<br />
.250, .133, .292, .119, .144, .022, .014. The darkest flasks<br />
evidently have the highest optical density measurements.<br />
This growth assay measures the total increase in optical<br />
density over time. The higher the change in optical density<br />
is, the higher the rate of growth for the colony is. Thus,<br />
when testing the two different transformed strains of<br />
Synechococcus elongatus UTEX L 2973, the transformed<br />
strain overexpressing dnaJ should have the higher change<br />
in optical density if the heat shock protein overexpressed<br />
through dnaJ truly increases photosynthetic and growth<br />
rates.<br />
Because dnaJ codes for a heat shock protein, it<br />
was suspected that the growth rates of the strain<br />
overexpressing this gene would be significantly greater<br />
in only heat stressed conditions in comparison to the<br />
cyanobacteria only transformed with the vector. It was<br />
believed that the growth rate in normal conditions would<br />
not be affected greatly by the heat shock protein because<br />
the overexpression would not be necessary to withstand<br />
high temperatures. However, once the growth rates were<br />
Figure 5a. Graph of optical density (745nm) of control<br />
stain and dnaJ overexpressing strain 30° Celsius. The<br />
data suggest dnaJ significantly increases growth<br />
rates.<br />
Figure 5b. The colonies were grown at 40° for 9 days.<br />
There was a trend in the data that indicates dnaJ<br />
promotes faster growth, but after the colony was<br />
exposed to a higher temperature, 42°, the control<br />
stain did not grow whereas the dnaJ strain continued<br />
normal growth. Data became significant two days<br />
after increased temperature.<br />
The overexpression of dnaJ in normal conditions of<br />
30°C and 12-hour light cycles increased the growth rate<br />
of the cyanobacteria significantly in comparison to the<br />
control. This significance was seen as early as day 8. The<br />
average OD745 of the dnaJ strain after 12 days was .39 in<br />
comparison to the vector strain that had an average OD745<br />
of .25. This increase in optical density is attributed to the<br />
overexpression of dnaJ.<br />
DnaJ’s overexpression within cyanobacteria in heat<br />
stressed conditions of 40°C and 12-hour light cycles also<br />
tended to increase growth rates. After 9 days, the average<br />
14 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
OD745 of the overexpressing strain was .58 while the other<br />
strain had an average of only .52; however, the standard<br />
deviation within each of the sample groups was too high<br />
to conclude significance at 40°. When the temperature<br />
in the Percival Incubator was increased to 42°, the dnaJ<br />
overexpressing strain grew normally, whereas the control<br />
strain’s average optical density decreased. The average<br />
optical density of the dnaJ overexpressing strain after<br />
19 days was 1.35, and the control strain had an average<br />
optical density of .35. After just two days being exposed<br />
to the higher temperature, the difference between the two<br />
strains was significant, suggesting that the overexpression<br />
of dnaJ provided heat resistance to the transformed<br />
strain of cyanobacteria. There is a visual difference in<br />
optical density at day 19 in comparison to day 5, which<br />
demonstrates dnaJ’s potential to increase growth rates and<br />
provide heat resistance (Fig. 6).<br />
Figure 6. The photo of the flasks in the top panel was<br />
taken on day 5. The flasks have similar tints of green.<br />
The last photo was taken on day 19. In the eight flasks<br />
on the left, the dnaJ overexpressing cultures have a<br />
much darker color than the control flasks.<br />
4. Discussion<br />
Essentially, this research sought to create a unique<br />
strain of cyanobacteria through genetic transformation.<br />
The specific plasmid utilized in the experimentation<br />
had not been used to transform Synechococcus elongatus<br />
UTEX L 2973 previously. The successful transformation<br />
as seen from the colony growth in the selective presence<br />
of spectinomycin demonstrates the competence of the<br />
plasmid pSyn_6 in transforming the experiment’s specific<br />
strain.<br />
Despite the successful outcome of the research,<br />
there were several limitations in the experiment due to<br />
equipment and budget restrictions. One such limitation<br />
was the inability to determine the difference between<br />
the rates of oxygen evolution in the two strains. This<br />
would have led to a more precise measurement of the<br />
photosynthetic rate because oxygen is directly produced in<br />
photosynthesis. Optical density is a less direct measurement<br />
of this rate, but accurate, nonetheless. Without the<br />
generation of sugars through photosynthesis, the strains<br />
BIOLOGY<br />
could not grow. Because of this, higher photosynthetic<br />
rates should correspond to higher growth rates. Another<br />
limitation of this experiment was the inability to confirm<br />
gene expression in the transformed strains. However,<br />
confirming the correct plasmid makes it reasonable to<br />
assume that the growth rates increased on account of dnaJ<br />
overexpression.<br />
This beneficial genetic overexpression has many<br />
potential applications in both the agriculture and energy<br />
industries. Because cyanobacteria are currently the most<br />
photosynthetically efficient organisms on the planet,<br />
this modification could lead to future applications in<br />
agriculture or more economic biofuel production that will<br />
capitalize on their efficiency (Hunt, 2003). One possible<br />
application could be the production and secretion of<br />
sugars for consumption. Because cyanobacteria are not<br />
seasonal like sugar cane, they could produce sugars more<br />
consistently and more efficiently, especially following<br />
genetic engineering. Clearly, isolation of sugar from a<br />
cyanobacteria solution would have to be much cheaper<br />
for this to be a viable contender with sugar cane, but<br />
nonetheless, this could be a potential application of<br />
genetically engineered cyanobacteria. Beyond sugars,<br />
cyanobacteria’s products have been manipulated to<br />
produce ethanol (Chow et al., 2015). Producing ethanol<br />
could prove to be a disruptive application of cyanobacteria<br />
in the energy industry, especially when paired with dnaJ<br />
overexpression.<br />
Another possible application could be overexpressing<br />
heat shock proteins in other photosynthetic organisms to<br />
determine their effect on growth and photosynthetic rates.<br />
Overexpressing either dnaJ or corollary proteins specific<br />
to certain species within corn, rice, or soybeans could<br />
lead to increased production of these crops both in fertile<br />
geographies and in regions that are currently considered<br />
arid. Because heat shock proteins increased growth in<br />
Synechococcus elongatus UTEX L 2973 even in heat stressed<br />
conditions, it could be possible to genetically engineer<br />
cash crops to make them resistant to higher temperatures.<br />
This resistance could lead to the cultivation of previously<br />
infertile land, feeding millions more people worldwide.<br />
Further experimentation must be done to conclude the<br />
viability of any of these applications.<br />
5. Acknowledgements<br />
I would like to thank Dr. Monahan for teaching<br />
me the research process and guiding me through the<br />
fickle experimentation that is molecular biology. Thanks<br />
to her instruction and her patience with my stubborn<br />
commitment to this project, I was able to persevere<br />
through obstacles and accomplish my dream of genetically<br />
engineering cyanobacteria. I would like to thank Dr.<br />
Sheck for supervising me while I spent hours in the sterile<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 15
hood working with my cyanobacteria. I would also like to<br />
thank the rest of my Research in Biology colleagues for<br />
encouraging me throughout my time researching. I would<br />
like to thank Kevin Zhang and Tyler Edwards who were<br />
lab assistants during the Glaxo Summer Research Fellows<br />
Program. Finally, I want to thank the North Carolina School<br />
of Science and Mathematics and the Glaxo Endowment for<br />
blessing me with the opportunity to experience research in<br />
high school. I have learned many lessons that I will carry<br />
with me through the rest of my career in both research and<br />
other fields.<br />
6. References<br />
Al-Haj, L., Lui, Y. T., Abed, R. M. M., Gomaa, M. A.<br />
& Purton, S. Cyanobacteria as Chassis for Industrial<br />
Biotechnology: Progress and Prospects. Life (Basel) 6,<br />
(2016).<br />
Algae, U. C. C. of. UTEX L 2973 Synechococcus elongatus.<br />
UTEX Culture Collection of Algae Available at: https://<br />
utex.org/products/utex-l-2973. (Accessed: 25th January<br />
<strong>2018</strong>)<br />
Biolabs, N. E. Protocol for OneTaq Hot Start DNA<br />
Polymerase (M0481). New England Biolabs: Reagents<br />
for the Life Sciences Industry Available at: https://www.<br />
neb.com/protocols/2012/09/05/one-taq-hot-start-dnapolymerase-m0481.<br />
(Accessed: 27th October <strong>2018</strong>)<br />
Biolabs, N. E. Taq 2X Master Mix. New England Biolabs:<br />
Reagents for the Life Sciences Industry Available at:<br />
https://www.neb.com/products/m0270-taq-2x-mastermix#Product<br />
Information. (Accessed: 7th October <strong>2018</strong>)<br />
Chow, T.J. et al. Using recombinant cyanobacterium<br />
(Synechococcus elongatus) with increased carbohydrate<br />
productivity as feedstock for bioethanol production via<br />
separate hydrolysis and fermentation process. Bioresource<br />
Technology 184, 33–41 (2015).<br />
GeneArt Synechococcus Protein Expression Vector.<br />
Thermo Fisher <strong>Scientific</strong> Available at: https://www.<br />
thermofisher.com/order/catalog/product/A24230.<br />
(Accessed: 7th October <strong>2018</strong>)<br />
Hihara, Y., Kamei, A., Kanehisa, M., Kaplan, A. & Ikeuchi,<br />
M. DNA Microarray Analysis of Cyanobacterial Gene<br />
Expression during Acclimation to High Light. Plant Cell<br />
13, 793–806 (2001).<br />
Home - Synechococcus elongatus PCC 7942. Available<br />
at:https://genome.jgi.doe.gov/portal/synel/synel.home.<br />
html. (Accessed: 21st January <strong>2018</strong>)<br />
Hunt, S. Measurements of photosynthesis and respiration<br />
in plants. Physiol Plant 117, 314–325 (2003).<br />
Kufryk, G. I., Sachet, M., Schmetterer, G. & Vermaas, W.<br />
F. J. Transformation of the cyanobacterium Synechocystis<br />
sp. PCC 6803 as a tool for genetic mapping: optimization<br />
of efficiency. FEMS Microbiology Letters 206, 215–219<br />
(2002).<br />
Martin, A., Researchgate. Available at: https://www.<br />
researchgate.net/post/When_measuring_cyanobacterial_<br />
growth_when_do_I_use_which_wavelength.<br />
Martin, S. World will run out of food by 2050 thanks<br />
to population boom. Express.co.uk (2017). Available<br />
at: https://www.express.co.uk/news/science/803791/<br />
World-will-run-out-of-food-by-2050-population-boom.<br />
Minda, Renu, et al. “The Evolutionary Significance of<br />
‘Obligate’ Photoautotrophy of Cyanobacteria.” Current<br />
Science, vol. 94, no. 7, 10 April 2008, pp. 850-852.<br />
Occhialini, A., Lin, M. T., Andralojc, P. J., Hanson, M. R.<br />
& Parry, M. A. J. Transgenic tobacco plants with improved<br />
cyanobacterial Rubisco expression but no extra assembly<br />
factors grow at near wild-type rates if provided with<br />
elevated CO2. The Plant Journal 85, 148–160 (2015).<br />
Parmar, A., Singh, N. K., Pandey, A., Gnansounou, E. &<br />
Madamwar, D. Cyanobacteria and microalgae: A positive<br />
prospect for biofuels. Bioresource Technology 102, 10163–<br />
10172 (2011).<br />
QIAGEN. Quick-Start Protocol: QIAamp DNA Mini<br />
Kit. Confidence in Your PCR Results - The Certainty of<br />
Internal Controls - QIAGEN Available at: https://www.<br />
qiagen.com/us/resources/resourcedetail?id=566f1cb1-<br />
4ffe-4225-a6de-6bd3261dc920&lang=en.<br />
QIAprep Spin Miniprep Kit. Confidence in Your PCR<br />
Results - The Certainty of Internal Controls - QIAGEN<br />
Available at: https://www.qiagen.com/us/shop/sampletechnologies/dna/plasmid-dna/qiaprep-spin-miniprepkit/#orderinginformation.<br />
QIAquick Gel Extraction Kit. Confidence in Your PCR<br />
Results - The Certainty of Internal Controls - QIAGEN<br />
Available at: https://www.qiagen.com/us/shop/sampletechnologies/dna/dna-clean-up/qiaquick-gel-extractionkit/#orderinginformation.<br />
Restriction Endonuclease Products | NEB. Available<br />
at: https://www.neb.com/products/restrictionendonucleases.<br />
(Accessed: 2nd February <strong>2018</strong>)<br />
16 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
Shestakov, S. V. & Khyen, N. T. Evidence for genetic<br />
transformation in blue-green alga Anacystis nidulans.<br />
Molec. Gen. Genet. 107, 372–375 (1970).<br />
Synechococcus sp. UTEX 2973, complete genome. (2015).<br />
Watanabe, S., Sato, M., Nimura-Matsune, K., Chibazakura,<br />
T. & Yoshikawa, H. Protection of psbAII transcript from<br />
ribonuclease degradation in vitro by DnaK2 and DnaJ2<br />
chaperones of the cyanobacterium Synechococcus elongatus<br />
PCC 7942. Biosci. Biotechnol. Biochem. 71, 279–282<br />
(2007).<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 17
HYPOGLYCEMIC EFFECT OF Momordica charantia<br />
AGAINST TYPE 2 DIABETES MODELED IN Bombyx mori<br />
Aarushi Venkatakrishnan<br />
Abstract<br />
Diabetes is a disease that affects millions across the world, occurring when there are high levels of glucose in the blood.<br />
Currently, treatments for Type 2 Diabetes include lifestyle and diet changes, medication, and insulin injections; however,<br />
natural treatments, such as the vegetable bitter melon, have become more popular in recent years. As it is abundantly<br />
grown in Asia, which houses 60% of the world’s diabetics, this finding can be very effective. Using a silkworm model, the<br />
hypoglycemic effect of bitter melon was quantified by measuring the silkworm’s hemolymph glucose concentration with<br />
the phenol sulfuric acid method. Injections of saline, insulin, and bitter melon solutions were made at the first proleg of<br />
the silkworms. Hyperglycemia was induced after two days of a 10% high glucose diet, and human insulin significantly<br />
counteracted the effect. There were no changes to mass or length between the hyperglycemic and normal silkworms.<br />
After comparing its hypoglycemic effect to insulin, a known hypoglycemic agent, the most effective tested dose of bitter<br />
melon was found to be 175 µg/mL, 5 times greater than the corresponding insulin dose, 35 µg/mL. With further trials to<br />
determine the symptoms and overall effects to human health, bitter melon can potentially be recommended as an addition<br />
to the diet for diabetes treatment.<br />
1. Background<br />
1.1 Introduction<br />
Diabetes mellitus is a disease which is characterized by<br />
high levels of sugar in the blood, hyperglycemia, resulting<br />
from the body unable to use blood glucose for energy<br />
(Drive, n.d.). Typical symptoms include increased thirst,<br />
unexplained weight loss, and frequent infections (Drive,<br />
n.d.). This disease occurs when the body is unable to<br />
effectively use insulin, a hormone made by the pancreas,<br />
to process glucose, and thus causing an increase in blood<br />
sugar (“Insulin, Medicines, & Other Diabetes Treatments,”<br />
2016). There are two types ranging in severity: Type 1 and<br />
Type 2. Type 1 diabetes is an autoimmune disease in which<br />
the immune system destroys islet cells resulting in the body<br />
being unable to make insulin (Jin Yang & Mook Choi,<br />
2015). Type 2 diabetes is a chronic condition that changes<br />
how the body is able to metabolize glucose, caused by the<br />
pancreas either not producing enough insulin or the body<br />
becoming resistant to insulin (Matsumoto et al., 2011).<br />
The current treatment for Type 1 is administering insulin<br />
exogenically with numerous insulin treatments available,<br />
such as an insulin pump, pen, or an inhaler. These vary in<br />
terms of how fast they act, the quickest being 15 minutes<br />
after injection and the longest being several hours, but<br />
correspondingly the duration of the effect differs (“Insulin,<br />
Medicines, & Other Diabetes Treatments,” 2016). Type 2<br />
diabetes treatment includes maintaining a healthy lifestyle<br />
and monitoring blood glucose levels (Matsumoto et al.,<br />
2011). However, Type 2 diabetic patients can even take<br />
insulin treatments to make up for that not produced in the<br />
body; metformin is a commonly prescribed medicine first<br />
given to diabetic patients to lower the amount of glucose<br />
produced by the liver and help the body process insulin<br />
better (“Insulin, Medicines, & Other Diabetes Treatments,”<br />
2016).<br />
The number of people being affected by this condition<br />
has been increasing rapidly. In 2015, 1.5 million new cases<br />
were diagnosed, and in the United States alone, there were<br />
30.2 million Americans with some form of diabetes in 2017<br />
(“CDC Press Releases,” 2016). With this rise, the demand<br />
for diabetes treatments has increased. The development<br />
of new ways to introduce or stimulate insulin secretion<br />
is necessary as it has the potential to help the millions of<br />
people afflicted with diabetes live healthier lives.<br />
Although there are a variety of drugs in the market<br />
for Type 2 Diabetes, the desire for herbal medicines<br />
has increased as they can often be more accessible than<br />
traditional forms. In addition, these natural remedies are<br />
often more available and accepted than Western Medicine.<br />
There has always been a sector of herbal medicines called<br />
Complementary and Alternative Medicine (CAM); one<br />
of the oldest and most-well known practices is Ayurvedic<br />
medicine, which originates in India. Many herbs, fruits,<br />
and vegetables used in Ayurvedic medicine have shown<br />
promising results against diseases involving high blood<br />
pressure, anxiety, cancer, and more (Axe, 2015). One<br />
notable vegetable used is Momordica charantia, otherwise<br />
known as bitter melon. Numerous studies have been<br />
conducted that suggest bitter melon has hypoglycemic<br />
effects (Fuangchan, 2011; Jin Yan, 2015).<br />
Jin Yang et al. treated diabetic rats with bitter melon<br />
and found three functional components of bitter melon<br />
that were likely causing a hypoglycemic effect: charantin,<br />
vicine, and polypeptide-p (2015). By using three groups:<br />
a high fat control, high fat and 1% bitter melon, and high<br />
fat and 3% bitter melon, bitter melon had significantly<br />
18 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
improved glucose tolerance and insulin sensitivity. In the<br />
3% bitter melon, they found that it increased the levels<br />
of two insulin receptors, (phosphor-insulin receptor<br />
substreate-1 (Tyr612) and phosphor-Akt (Ser473), likely<br />
stimulating the hypoglycemic effect (Jin Yang & Mook<br />
Choi, 2015).<br />
Furthermore, bitter melon has been used in clinical<br />
studies using human patients with Type 2 Diabetes.<br />
Fuangchan et al. investigated the effects of varying doses<br />
of bitter melon (500 mg/day, 1000 mg/day, 2000 mg/day)<br />
when comparing it to metformin (1000 mg/day), a current<br />
diabetes medication (2011). By measuring the fructosamine<br />
concentrations over a 2-week time period from baseline to<br />
endpoint, they found that the 500 mg/day and 1000 mg/<br />
day doses did not significantly impact glucose levels, while<br />
the 2000 mg/day dose did. When compared to metformin,<br />
the effects of bitter melon were still less (Fuangchan, 2011).<br />
Both groups did not experience extreme adverse effects;<br />
only mild headaches, dizziness, and increased hunger<br />
were experienced in the 2000 mg/day bitter melon group<br />
(Fuangchan, 2011). The drawbacks of this study include<br />
the limited time as it was only conducted for 4 weeks, and<br />
because effects were only seen for the 2000 mg/day dose,<br />
higher dose levels would need to be tested.<br />
1.2 – Silkworm Model<br />
Although bitter melon is known to have hypoglycemic<br />
effects, research surrounding the topic is not standardized<br />
and hard to compare. With very minimal clinical trials,<br />
the side effects of bitter melon are difficult to determine<br />
as well. In Matsumoto et al., scientists established the<br />
silkworm as a reliable model of diabetes (Matsumoto et al.,<br />
2011). While silkworms do not have blood like humans,<br />
they have hemolymph which is a fluid “equivalent” to<br />
blood. In this study, glucose levels after treatment with a<br />
high glucose diet were higher than that of the silkworms<br />
fed a normal diet. By treating the hyperglycemic silkworms<br />
with insulin, glucose levels returned to normal. Moreover,<br />
they also tested an herbal extract, jiou, and found that it<br />
could mimic the effects of insulin by reducing glucose<br />
concentrations.<br />
Based on this research, the silkworm model could be<br />
used to determine the hypoglycemic effects of bitter melon.<br />
Here we show hyperglycemic silkworms that are treated<br />
with bitter melon extract, to study if their hemolymph<br />
sugar levels will go down without adverse reactions in<br />
terms of body size, body mass, or lifespan because bitter<br />
melon has been known to have hypoglycemic properties<br />
as used in ayurvedic medicine. This hypoglycemic effect<br />
is likely as bitter melon has the identified components<br />
charantin, vicine, and polypeptide-p and has been a cultural<br />
remedy as used in Ayurvedic medicine.<br />
BIOLOGY<br />
Figure 1. Anatomy of a silkworm. Length<br />
measurements were made from the thorax to the<br />
caudal leg. Injections were made in between the first<br />
proleg and the second proleg from the head capsule.<br />
2. Methods<br />
This study consisted of two preliminary experiments and<br />
two main experiments. The two preliminary experiments<br />
determined the equation for the Beer’s Law Plot to be<br />
used when calculating the D-glucose concentration of<br />
silkworms and established that a high glucose diet raised<br />
the D-glucose levels of silkworms. The main experiments<br />
tested the effect of insulin when compared to the same dose<br />
of bitter melon and evaluated the optimal concentration<br />
of bitter melon. The experiment unit was the addition<br />
and kind of hypoglycemic agent, measuring the change in<br />
average D-glucose levels. For the main experiments, the<br />
positive control was the hyperglycemic silkworm treated<br />
with insulin, a known hypoglycemic agent. The negative<br />
control was the silkworm fed a normal diet and injected<br />
with saline, to mimic the effect of an injection without the<br />
addition of a chemical agent.<br />
2.1 – Silkworm Diet<br />
The essentials of the silkworm diet consist of mulberry<br />
leaves. To create the silkworm diet, Carolina® Silkworm<br />
Diet was purchased from Carolina Biological. In a 2000<br />
mL glass beaker, ½ pound of the mulberry powdered diet<br />
was added to 720 mL (roughly 3 cups) of tap water. Using<br />
a stirring rod, the substances were thoroughly mixed to a<br />
uniform consistency. It was then covered with plastic wrap<br />
and secured with a rubber band. The beaker was placed in<br />
the microwave at high heat until the mixture came to a<br />
boil, usually after 1-2 minutes. This caused the mixture to<br />
rise and bubbles appeared on the surface. Once the mixture<br />
boiled, the beaker was removed from the microwave and<br />
stirred to again ensure uniform consistency. It was then<br />
placed back in the microwave to repeat the process. After<br />
the second boil and mixture, plastic wrap was tightly placed<br />
against the surface of the mixture to ensure no moisture<br />
escaped. Time was allowed for the beaker and substances<br />
to cool down. Then, the top of the beaker was secured<br />
with plastic wrap and a rubber band. It was placed in the<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 19
efrigerator to store.<br />
2.2 – Silkworm Maintenance<br />
Silkworm eggs were purchased from Carolina<br />
Biological. They were placed in petri dishes and incubated<br />
at 29 °C. After roughly a week, the eggs hatched, and they<br />
were disposed of. The larva was transferred to a fresh plate<br />
with mulberry powdered diet placed on a paper towel.<br />
Feedings were made every other day to clean out feces and<br />
remove dried food. The experiment was performed during<br />
the fifth instar (around 4 weeks after hatching). Raising<br />
the temperature increased growth, whereas lowering the<br />
temperature delayed growth.<br />
2.3 – High Glucose Diet<br />
To induce hyperglycemic conditions, a high glucose diet<br />
of the Mulberry Chow was created by mixing appropriate<br />
amounts of D-Glucose and the Mulberry Chow. D-Glucose<br />
was added to the Mulberry Chow in a beaker, and then<br />
mixed until all contents were dissolved. A 10% and 15%<br />
D-Glucose diet were created.<br />
2.4 – Injection<br />
50 µl of each solution was injected into the hemolymph<br />
at the second abdominal segment of the larva after the<br />
first proleg using 1 mL syringes (Fig. 2). Injections were<br />
done on 12-hour cycles for 2 days after 2 days of the high<br />
glucose diet for the preliminary experiments. Injections<br />
were performed once 24 hours before extraction for the<br />
main experiments. The total treatment lasted 4 days with<br />
measurements taken on Day 5.<br />
was made with 0.9% NaCl and 0.1% acetic acid. A stock<br />
solution of bitter melon was created by combining 0.50<br />
grams of powdered bitter melon with 50 mL of distilled<br />
water to make a 0.1 g/mL solution. It was heated at 40<br />
°C, for 15 minutes and left overnight for 2 days. Then,<br />
vacuum filtration was performed 3 times to filter out<br />
particulates. It was then appropriately diluted to the<br />
varying concentrations used in the experiment.<br />
2.6 – Glucose Quantification<br />
Hemolymph was collected from the larvae through a cut<br />
on the first proleg after they had developed to the fifth instar.<br />
Precipitated proteins were removed by centrifugation at<br />
3000 rpm for 10 min. 175 µl of the supernatant was diluted<br />
with 175 µl distilled water for sugar quantification (350 µl<br />
total). The total sugar in the hemolymph was determined<br />
using the 0.05 % phenol-sulfuric acid (PSA) method.<br />
Hemolymph extract (350 µl) was mixed vigorously with<br />
1050 µl 70% sulfuric acid. Immediately, 210 µl of 5% phenol<br />
aqueous solution were added and mixed. The test tubes<br />
were held in a water bath at 90 °C. The samples were then<br />
cooled to room temperature. The absorbance at 490 nm<br />
was measured using a spectrophotometer. Serially diluted<br />
glucose solution was used as a standard.<br />
2.7 – Statistical Measurements<br />
Measurements of the mass (g) and length (cm) of<br />
the silkworms were taken prior to experimentation.<br />
They were then monitored throughout the trial days<br />
and recorded until extraction. Data were analyzed using<br />
unpaired student t-tests with unequal variance, as sample<br />
size differed across the trials. Error bars were calculated<br />
using standard error of the mean (SEM).<br />
2.8 – Comparing the Effects of Insulin and Bitter Melon<br />
Figure 2. Injection site after the first proleg.<br />
Hemolymph was extracted from this same site as<br />
well.<br />
To determine whether bitter melon has hypoglycemic<br />
effects, a silkworm model was used as it has been previously<br />
identified as a workable model for diabetes research<br />
(Fuangchan, 2011). Three characteristics were measured to<br />
determine the effectiveness of the bitter melon treatment:<br />
body mass, body length, and hemolymph sugar levels. Each<br />
trial consisted of 6 treatments, separated into high-glucose<br />
and normal diet models. In addition to these treatments,<br />
the effect of insulin on both the hyperglycemic and normal<br />
silkworm was used to provide a standard of comparison to<br />
see the success of the bitter melon extract (Table 1.1).<br />
2.5 – Injection Solutions<br />
A 35 µg/mL solution of insulin was created by diluting<br />
a 20 mg/mL solution to 15 mL. The dilution solution<br />
20 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
Table 1.1. Experimental design comparing effect of<br />
equal dose bitter melon<br />
Diet<br />
Treatment Normal Diet High Glucose Diet<br />
Saline “Normal” with<br />
Saline (NS)<br />
“High” with<br />
Saline (HS)<br />
Insulin “Normal” with<br />
Insulin (NI)<br />
“High” with<br />
Insulin (HI)<br />
Bitter<br />
Melon<br />
“Normal” with<br />
Bitter Melon<br />
(NB)<br />
“High” with<br />
Bitter Melon<br />
(HB)<br />
A.<br />
B.<br />
2.9 – Determining the Ideal Concentration of Bitter Melon<br />
Following the experimental model from the Injections,<br />
varying concentrations of bitter melon were tested in the<br />
silkworms to determine which concentration of bitter<br />
melon has the largest effect on the sugar concentration in<br />
silkworms (Table 1.2).<br />
Table 1.2. Experimental design comparing effect of<br />
varying doses of bitter melon<br />
Diet<br />
Treatment Normal Diet High Glucose Diet<br />
Saline “Normal” with<br />
Saline (NS)<br />
“High” with Saline<br />
(HS)<br />
Insulin - “High” with<br />
Insulin (HI)<br />
Bitter<br />
Melon 1<br />
- “High” with Bitter<br />
Melon 1 (HB1)<br />
Bitter<br />
Melon 2<br />
- “High” with Bitter<br />
Melon 2 (HB2)<br />
Bitter<br />
Melon 3<br />
- “High” with Bitter<br />
Melon 3 (HB3)<br />
Figure 3. (A, B) Glucose standards undergoing phenolaqueous<br />
protocol and depicted visually with a<br />
gradient of yellow colors, shown from left to right as<br />
(1) Blank, (2) 1.00 M D-Glucose, (3) 0.055 M D-Glucose,<br />
(4) 0.0275 M D-Glucose<br />
3.2 – High Glucose Diet<br />
Before proceeding with the experiment, a baseline of a<br />
high glucose diet needed to be tested. In this preliminary<br />
experiment, three treatments were tested: a normal diet<br />
of mulberry chow, a 10% glucose diet after 24 hours, and<br />
the same diet after 48 hours. A sample size of 9 silkworms<br />
was used for the normal diet. Five silkworms were used for<br />
both of the 10% glucose diet treatments. Average glucose<br />
levels across the hemolymphs of the silkworms are shown<br />
(Fig. 4).<br />
3. Results<br />
3.1 – Glucose Quantification<br />
To understand and best utilize the phenol aqueous<br />
method, a series of D-glucose standards were used to<br />
create a Beer’s Law plot (Fig. 3A,B). Different D-Glucose<br />
concentrations were used to generate a standard. These<br />
included 0.0275 M, 0.055 M, and 1.000 M.<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 21
A.<br />
B.<br />
3.3 – Comparing the Effects of Insulin and Bitter Melon<br />
After establishing that a hyperglycemic diet could<br />
induce high glucose concentrations in silkworms, the<br />
effect of insulin was tested on a normal and high glucose<br />
diet. This standard concentration would also be replicated<br />
by an equal dose of bitter melon extract. Based on the<br />
results of the previous experiments, silkworms were fed a<br />
high glucose diet for 2 days to induce hyperglycemia (Fig.<br />
4). On Day 3 of the diet, injections were performed with<br />
the various treatments. Then, on Day 4, their hemolymphs<br />
were extracted and quantified for glucose concentration.<br />
C.<br />
Figure 4. (A) Average D-glucose concentration. (B)<br />
Average mass. (C) Average length. Three different<br />
treatment groups were measured: Normal (mulberry<br />
diet); High Day 1 (mulberry + 10% glucose for 24<br />
hours); High Day 2 (mulberry to 10% glucose for 48<br />
hours). (p < 0.05= *, p < 0.01= **, p < 0.001 = ***, p > 0.05=<br />
ns). Error bars ± SEM.<br />
The average glucose concentration of a normal<br />
silkworm is about 30.0 mg/mL (Fig. 4A). With the addition<br />
of a high glucose diet, the average glucose concentration is<br />
52.4 mg/mL after one day and 79.1 mg/mL after two days.<br />
This means that after 48 hours, there is almost a 2.5-fold<br />
increase. After conducting a t-test for further statistical<br />
analysis, p-values were calculated. The only statistically<br />
significant difference is between the normal diet and two<br />
days of the high glucose diet, indicating that 48 hours at<br />
a minimal of 10% D-glucose diet is required to increase<br />
hemolymph glucose levels significantly. There was no<br />
statistically significant difference between the mass and<br />
length of the normal silkworms and the two tested trials<br />
(Fig. 4B and 4C).<br />
Figure 5. Average glucose concentrations across<br />
various treatments. Six different treatment groups<br />
were measured: Normal with Saline (mulberry diet<br />
+ insect saline, n=8), Normal with Insulin (mulberry<br />
diet + 35 µg/mL insulin, n=7), Normal with Bitter<br />
Melon (mulberry diet + 35 µg/mL bitter melon<br />
extract, n=9), High with Saline (mulberry diet + 15%<br />
glucose for 84 hours + insect saline, n=9), High with<br />
Insulin (mulberry diet + 15% glucose for 84 hours +<br />
35 µg/mL insulin, n=10), and High with Bitter Melon<br />
(mulberry diet + 15% glucose for 84 hours + 35 µg/mL<br />
bitter melon extract, n=9). (p < 0.05= *, p < 0.01= **, p <<br />
0.001 = ***, p > 0.05 = ns). Error bars ± SEM.<br />
The average glucose concentration for each of the<br />
treatments are as follows (Fig. 5): 40.595 mg/mL (Normal<br />
diet with Saline), 26.929 mg/mL (Normal diet with 35<br />
µg/mL Insulin), 34.366 mg/mL (Normal diet with 35 µg/<br />
mL Bitter Melon), 62.905 mg/mL (15% High glucose diet<br />
with Saline), 33.287 mg/mL (15% High glucose diet with<br />
Insulin), and 50.579 mg/mL (15% High glucose diet with<br />
bitter melon). As can be seen from the p-values, there was<br />
a statistically significant difference between the normal<br />
with saline and high with saline, and one between the<br />
high with saline and high with insulin. This corresponds<br />
to the predicted control values. There was no statistically<br />
significant difference between the high with saline and<br />
high with bitter melon.<br />
22 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
3.4 – Revised Standard Curve<br />
Using a different series of D-glucose solutions and a<br />
new spectrophotometer, a new standard curve was created<br />
(Fig. 6).<br />
A.<br />
B.<br />
Figure 6. New Glucose Standard Curve. Using<br />
concentrations of 12.5 mg/mL and 25 mg/mL, the<br />
standard curve for glucose was created.<br />
The linear fit was used to evaluate the following<br />
experiment in terms of finding the ideal dose of bitter<br />
melon for hyperglycemic silkworms.<br />
3.5 – Determining the Ideal Concentration of Bitter Melon<br />
Ultimately, a 35 µg/mL concentration of bitter melon<br />
was not effective, but it did show a decrease from the<br />
hyperglycemic silkworms (Fig. 5). By adjusting the<br />
concentration of bitter melon, the downward trend could<br />
be quantified further.<br />
The average values for each treatment were: 26.046<br />
mg/mL (Normal with Saline), 37.111 mg/mL (15% High<br />
glucose diet with Saline), 25.043 mg/mL (15% High glucose<br />
diet with 35 µg/mL Insulin), 24.029 mg/mL (15% High<br />
glucose diet with 35 µg/mL Bitter melon), 33.459 mg/mL<br />
(15% High glucose diet with 87.5 µg/mL Bitter melon),<br />
20.006 mg/mL (15% High glucose diet with 175 µg/mL<br />
Bitter melon), and 18.265 mg/mL (15% High glucose diet<br />
with 350 µg/mL bitter melon) (Fig. 7A).<br />
Insulin significantly reduced the glucose concentrations<br />
of the high glucose diet to the levels of the normal diet. Two<br />
bitter melon treatments yielded statistically significant<br />
results, with the 175 µg/mL bitter melon extract having a<br />
p-value that most closely resembled insulin (0.00303 and<br />
0.002954).<br />
Figure 7. (A) Determining a Statistically Significant<br />
Bitter Melon dose. Seven different treatment groups<br />
were measured: Normal with Saline (mulberry diet<br />
+ insect saline, n=8), High with Saline (mulberry diet<br />
+ 15% glucose for 84 hours + insect saline, n=9), High<br />
with Insulin (mulberry diet + 15% glucose for 84 hours<br />
+ 35 µg/mL insulin, n=7, High with 35 µg/mL Bitter<br />
Melon (mulberry diet + 15% glucose for 84 hours +<br />
35 µg/mL bitter melon extract, n=5), High with 87.5<br />
µg/mL Bitter Melon (mulberry diet + 15% glucose<br />
for 84 hours + 87.5 µg/mL bitter melon extract, n=8),<br />
High with 175 µg/mL Bitter Melon (mulberry diet<br />
+ 15% glucose for 84 hours + 175 µg/mL bitter melon<br />
extract, n=6), and High with 350 µg/mL Bitter Melon<br />
(mulberry diet + 15% glucose for 84 hours + 350 µg/<br />
mL bitter melon extract, n=5). (p < 0.05= *, p < 0.01=<br />
**, p > 0.05 = ns). Error bars ± SEM. (B) (Left) 350 µg/<br />
mL bitter melon diet fed silkworm exhibiting a<br />
yellowish color and less rigid than (Right) Normal<br />
diet fed silkworm.<br />
4. Discussion<br />
4.1 – Bitter Melon’s Effect on Hyperglycemia<br />
The results confirm Matsumoto et al.’s conclusion that<br />
a silkworm model can exhibit hyperglycemia (Fig. 4). By<br />
feeding a 10% glucose mulberry diet, the hemolymph sugar<br />
levels increased significantly. For the convenience of the<br />
model, the addition of glucose was increased to 15% to<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 23
ensure that the intake across samples would be the same<br />
and that the effect was the greatest. Additionally, this<br />
figure also examines the mass and length across treatments.<br />
Matsumoto et al. claimed that mass and length differed with<br />
the addition of glucose to the diet, however, no significant<br />
variation in these metrics were found, indicating these<br />
metrics were not good indicators of health (2011).<br />
From there, insulin was added as a control treatment<br />
to compare the effect of bitter melon. The 35 µg/mL was<br />
appropriately scaled by the average insulin treatment for<br />
humans, with guidance from Matsumoto et al. (2011).<br />
There was a significant decrease in glucose levels, again<br />
confirming the silkworm model. New treatments were<br />
tested with a 15% glucose mulberry diet instead of the<br />
10% glucose mulberry diet (Fig. 4 & Fig. 5). With an equal<br />
concentration of bitter melon, 35 µg/mL, there was no<br />
significance in the data (p = 0.12). However, bitter melon<br />
did reduce the glucose levels from the high glucose and<br />
saline sample. From this, it could be concluded that bitter<br />
melon requires a larger dose than insulin does to perform<br />
a hypoglycemic effect.<br />
This prediction was tested (Fig. 7). With four varying<br />
concentrations of bitter melon – 35 µg/mL, 87.5 µg/<br />
mL, 175 µg/mL, and 350 µg/mL – the relationship of<br />
bitter melon doses and the hypoglycemic effect were<br />
discovered. The most effective dose out of the trials was<br />
175 µg/mL bitter melon, as it produced a similar p-value<br />
to the insulin treatment. However, the trend created<br />
with increasing doses did not suggest a linear trend, like<br />
predicted. Instead, it presented with a curved shape that<br />
is likely because of small sample size. The sample sizes of<br />
the various treatments ranged from 5 to 9, indicating that<br />
more samples would be necessary to determine if a linear<br />
trend exists.<br />
4.2 – Limitations<br />
Although the metrics of mass and length were not<br />
appropriate in quantifying the effect of bitter melon,<br />
qualitative observations of behavior helped define the<br />
health when compared to normal silkworms. With a high<br />
glucose mulberry diet, silkworms appeared lethargic and<br />
did not move as fast or eat as much as the normal diet<br />
silkworms. This could speak to food aversion or a toxicity<br />
of glucose in the diet. Silkworms have been frequently<br />
studied as a model of toxicity as they lack an adaptive<br />
immune system (Chen & Lu, <strong>2018</strong>). They possess PGs and<br />
LPS, immune stimulators that silkworms have developed<br />
based on their cell walls (Chen & Lu, <strong>2018</strong>). This allows<br />
silkworms to defend against pathogens and infections. If<br />
they had a similar response to the insulin or glucose diet,<br />
the model would not be ideal to study. With the addition of<br />
insulin, the worms were still slightly affected by this effect.<br />
In addition, silkworms in the bitter melon trials exhibited<br />
a yellowish tinge (Fig. 7B). The higher the dose, the higher<br />
the mortality rate. The study consisted of trials with 10<br />
initial worms, however, the sample sizes are significantly<br />
lower for the 35 µg/mL trial and the 350 µg/mL trial, as<br />
only 50% of the silkworms in those trials survived (Fig.<br />
7A).<br />
In addition, the administration of treatment also<br />
impacted the survival rate of the silkworms. As they<br />
have a limited hemolymph volume, injections caused<br />
severe hemolymph loss. Bruises and strain along the head<br />
and thorax were very visible. With extra pressure from<br />
the injections, silkworms often refrained from eating<br />
as they could not perform movement. This method of<br />
administering was rather ineffective as many samples<br />
could not be used for analysis.<br />
Lastly, the spectrophotometer used for the first half<br />
of the experiment was heavily used and therefore yielded<br />
unpredictable results. Therefore, results from the first<br />
part (Fig. 2 and 3) cannot be compared to results from the<br />
second part (Fig. 5), as there were two new standard curves<br />
created. Overall, similar results were yielded throughout<br />
the trials, which allows the general effect of the treatments<br />
to be compared.<br />
5. Conclusion and Future Work<br />
By using a silkworm model, diabetes could be effectively<br />
modeled. The most effective dose of bitter melon was<br />
determined to be 175 µg/mL, which had effects that closely<br />
resembled those of the insulin treatment. The data do not<br />
confirm a linear relationship between dose of bitter melon<br />
and hypoglycemic effect, as there was variation in the data.<br />
With this knowledge, future work is necessary before<br />
bitter melon can be marketed as a hypoglycemic agent. At<br />
this dose, side effects and symptoms should be generated to<br />
understand how it would impact human health. This can be<br />
done through mice and human clinical trials. Additionally,<br />
different methods of preparing the extract should be<br />
performed. A liquid extract was prepared with distilled<br />
water and a powdered form of bitter melon to make the<br />
injection solutions in the previously mentioned trials.<br />
The difference between powdered and fresh bitter melon<br />
should be studied, with the different parts of bitter melon<br />
as well (the core and the exterior). With the combination<br />
of all of these factors, the doses may vary depending on<br />
what is optimal.<br />
Bitter melon does not only show potential as a<br />
hypoglycemic agent, but also as a potential cancer<br />
and osteoarthritis therapy (Raina, 2016; Soo May,<br />
<strong>2018</strong>). Guided by bitter melon’s targeted effect against<br />
Type 2 Diabetes, Raina et al. attempted to provide a<br />
comprehensive view of the bioactivity of bitter melon’s<br />
different components and determine if they are applicable<br />
to cancer treatment (Raina et al., 2016). Specifically, they<br />
focus on how bitter melon interacts with other drugs.<br />
This is an important aspect to consider concerning<br />
24 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> BIOLOGY
diabetes as well, to see how bitter melon interacts with<br />
the mechanisms of insulin (Raina et al., 2016). Soo May<br />
et al. focus on bitter melon’s anti-inflammatory effects and<br />
how they can potentially reduce knee pain in osteoarthritis<br />
patients (Soo May et al., <strong>2018</strong>). They concluded that with<br />
3 months of supplementation, bitter melon can reduce<br />
the need for analgesia consumption, while also showing<br />
reductions in body weight, body mass index, and fasting<br />
blood glucose (Soo May et al., <strong>2018</strong>). Overall, bitter melon<br />
has a variety of beneficial effects that are not well studied,<br />
so it is important to understand how it affects the body to<br />
better recommend this natural remedy.<br />
Once this has been completed, health care providers<br />
can use this information, especially in Asian countries, to<br />
inform their patients about additional foods to include into<br />
their diet, with the appropriate intake. As there are many<br />
different vegetables and roots that are said to exhibit this<br />
hypoglycemic effect, they can be tested similarly to how it<br />
has been done in this paper and examine the effects before<br />
formally recommending inclusion into the diet.<br />
6. Acknowledgments<br />
I would like to thank Dr. Kimberly Monahan for being<br />
an encouraging mentor and guiding me through the<br />
research process. Thank you to the Research in Biology<br />
class of <strong>2019</strong> for providing support throughout this project.<br />
Thank you to Kevin Zhang and Tyler Edwards for being<br />
my lab assistants over the summer. Finally, I would like to<br />
thank Dr. Sheck, the North Carolina School of Science and<br />
Mathematics, and the Glaxo Endowment for allowing me<br />
the opportunity to experience research.<br />
7. References<br />
Fuangchan, A. (2011) Hypoglycemic effect of bitter melon<br />
compared with metformin in newly diagnosed type 2<br />
diabetes patients. Journal of Ethnopharmacology, 134(2),<br />
422–428. https://doi.org/10.1016/j.jep.2010.12.045<br />
Insulin, Medicines, & Other Diabetes Treatments. (2016,<br />
November 01). Retrieved January 27, <strong>2018</strong>, from https://<br />
www.niddk.nih.gov/health-information/diabetes/<br />
overview/insulin-medicines-treatments<br />
Jin Yang, S., Mook Choi, Jung. (2015). Preventive effects<br />
of bitter melon (Momordica charantia) against insulin<br />
resistance and diabetes are associated with the inhibition<br />
of NF-κB and JNK pathways in high-fat-fed OLETF rats.<br />
The Journal of Nutritional Biochemistry, 26(3), 234–240.<br />
https://doi.org/10.1016/j.jnutbio.2014.10.010<br />
Matsumoto, Y., Sumiya, E., Sugita, T., & Sekimizu, K.<br />
(2011). An Invertebrate Hyperglycemic Model for the<br />
Identification of Anti-Diabetic Drugs. PLOS ONE, 6(3),<br />
e18292. https://doi.org/10.1371/journal.pone.0018292<br />
Raina, K., Kumar, D., & Agarwal, R. (2016). Promise<br />
of bitter melon (Momordica charantia) bioactives in<br />
cancer prevention and therapy. Seminars in Cancer<br />
Biology, 40–41, 116–129. https://doi.org/10.1016/j.<br />
semcancer.2016.07.002<br />
Soo May, L., Sanip, Z., Ahmed Shokri, A., Abdul Kadir,<br />
A., & Md Lazin, M. R. (<strong>2018</strong>). The effects of Momordica<br />
charantia (bitter melon) supplementation in patients with<br />
primary knee osteoarthritis: A single-blinded, randomized<br />
controlled trial. Complementary Therapies in Clinical Practice,<br />
32, 181–186. https://doi.org/10.1016/j.ctcp.<strong>2018</strong>.06.012<br />
Axe, J. (2015, August 29). 7 Benefits of Ayurvedic Med<br />
icine: Lower Stress, Blood Pressure & More. (n.d.).<br />
Retrieved September 22, <strong>2018</strong>, from https://draxe.com/<br />
ayurvedic-medicine/<br />
CDC Press Releases. (2016, January 1). Retrieved<br />
January 27, <strong>2018</strong>, from https://www.cdc.gov/media/<br />
releases/2017/p0718-diabetes-report.html<br />
Chen, K., & Lu, Z. (<strong>2018</strong>). Immune responses to bacterial<br />
and fungal infections in the silkworm, Bombyx mori.<br />
Developmental & Comparative Immunology, 83, 3–11. https://<br />
doi.org/10.1016/j.dci.2017.12.024<br />
Drive, A. D. A. 2451 C., Arlington, S. 900, & Va 22202<br />
1-800-Diabetes. (n.d.). Diabetes Symptoms. Retrieved<br />
October 26, <strong>2018</strong>, from http://www.diabetes.org/<br />
diabetes-basics/symptoms/<br />
BIOLOGY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 25
TETRAETHYL ORTHOSILICATE-POLYACRYLONITRILE<br />
HYBRID MEMBRANES AND THEIR APPLICATION IN<br />
REDOX FLOW BATTERIES<br />
Ethan Frey<br />
Abstract<br />
Redox flow batteries (RFBs) are a reliable solution to long term energy storage, but lack an inexpensive and effective<br />
proton exchange membrane. Polyacrylonitrile (PAN) nanoporous membranes have a high chemical stability but<br />
low hydrophilicity when compared to Nafion, the standard membrane. The addition of tetraethyl orthosilicate (TEOS)<br />
increases the mechanical and thermal properties of membranes and may also increase their hydrophilicity due to the<br />
presence of hydrophilic silicon hydroxide bonds. Therefore, doping a nanoporous hydrophobic PAN membrane with<br />
TEOS is hypothesized to increase the hydrophilicity of the membrane, while still maintaining a high chemical stability<br />
and low vanadium crossover. Membranes of Nafion 212, nanoporous PAN, and a nanoporous hybrid TEOS/PAN were<br />
prepared through a phase inversion method and tested for chemical stability, proton and vanadium crossover in a model<br />
RFB, and water contact angle. The TEOS/PAN hybrid membrane had a higher hydrophilicity than both PAN and Nafion.<br />
The addition of TEOS had no impact on chemical stability. However, the TEOS/PAN hybrid membrane did have a higher<br />
vanadium crossover and lower proton/vanadium selectivity. It was concluded that TEOS can increase hydrophilicity,<br />
but more research needs to be done to improve proton/vanadium selectivity, potentially by optimizing pore size. Since<br />
TEOS was proven as an effective additive to membranes, progress was made towards the development of an ideal proton<br />
exchange membrane and a solution to long-term energy storage.<br />
1. Introduction<br />
Recently, polyacrylonitrile (PAN) nanofiltration<br />
membranes have proven to be a promising alternative to<br />
Nafion membranes due to their high chemical stability,<br />
and inexpensive cost. However, PAN membranes have<br />
been found to lack the proton conductivity of Nafion, a<br />
property that could be increased through the addition of<br />
additives to a PAN membrane. The addition of tetraethyl<br />
orthosilicate to a PAN membrane may not only increase<br />
the hydrophilicity and proton conductivity of PAN, but<br />
also improve the membrane’s mechanical strength and<br />
thermal properties, while still maintaining the chemical<br />
stability of PAN. The creation of a hybrid membrane of<br />
PAN and TEOS could make progress towards the creation<br />
of a cheaper membrane with properties comparable to that<br />
of Nafion.<br />
As renewable energy has become increasingly popular,<br />
demand for long term energy storage increased as well. As<br />
a result, a lot of attention has been given to redox flow<br />
batteries (RFBs) due to their ability to store energy for<br />
an indefinite period of time. What differentiates a RFB<br />
from other battery types is that it behaves essentially as a<br />
reversible fuel cell.<br />
Figure 1. Redox Flow Battery Design: Two different<br />
oxidation states of vanadium are stored in the tanks<br />
on either side of the battery and pumped into two<br />
adjacent half cells where the vanadium is reduced or<br />
oxidized and a flow of electrons is created. However,<br />
a proton exchange membrane is essential to allow<br />
this reaction to occur.<br />
The fuel is stored in tanks separate from where the<br />
oxidation and reduction occurs (Fig. 1). This fuel is<br />
pumped into two adjacent half cells separated by a proton<br />
exchange membrane. The vanadium on one side of the<br />
half cell is reduced and the other side is oxidized before<br />
being pumped back into the fuel tank. The battery is<br />
finished charging or discharging when all of the fuel has<br />
been reduced or oxidized. Commonly used batteries, such<br />
as lithium-ion batteries, can slowly discharge while not in<br />
use, resulting in a loss of charge over time. This is due to<br />
the fuel being stored where the reduction and oxidation<br />
occurs, allowing spontaneous reactions to take place even<br />
26 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
when the battery is not in use. Since the fuel in RFBs is<br />
stored externally, the battery cannot slowly discharge over<br />
time while not in use, and energy can be stored for an<br />
indefinite period of time. Vanadium is most typically used<br />
in RFBs due to its multiple oxidation states and large ion<br />
size (Alotto et al., 2014).<br />
The expensive proton exchange membrane prevents<br />
the use of RFBs on a commercial scale. The most widely<br />
used membrane is Nafion. This membrane is expensive and<br />
its properties could still be improved upon. It still exhibits<br />
vanadium ion crossover, and a higher hydrophilicity and<br />
proton conductivity could increase its efficiency. However,<br />
it is challenging to manipulate these properties while still<br />
maintaining a high level of chemical stability. Vanadium<br />
ion crossover is difficult to decrease while still maintaining<br />
proton conductivity. Similarly, proton conductivity is<br />
difficult to increase without decreasing chemical stability<br />
or increasing vanadium crossover. A membrane must<br />
be both hydrophobic to maintain chemical stability and<br />
hydrophilic to conduct protons. An alternative is to create<br />
a membrane that is simply very hydrophobic and has<br />
nanopores to allow protons to pass through. However, an<br />
extremely hydrophobic membrane struggles to keep the<br />
nanopores big enough to allow protons to pass through<br />
but small enough to prevent vanadium crossover. Nafion is<br />
designed (Fig. 2) such that it has a fluorinated carbon chain<br />
that allows for high chemical stability and hydrophobicity.<br />
Yet, Nafion still has a S-OH bond that allows for some<br />
hydrophilicity.<br />
exceeding 170°, indicating its high chemical stability (Feng<br />
et al., 2002). However, high hydrophobicity can become<br />
problematic when it prevents the membrane from<br />
conducting protons. Polyacrylonitrile membranes with<br />
nanopores from a phase inversion method (Zhang et al.,<br />
2011) and conditioning in an alkali solution (Karpushkin<br />
et al., 2017) have been investigated. These investigations<br />
found that, as expected, polyacrylonitrile lacks the proton<br />
conductivity of Nafion. Therefore, doping PAN to<br />
increase its hydrophilicity could create a membrane that<br />
is comparable to Nafion.<br />
Doping Nafion with metal oxides to improve its<br />
hydrophilicity has been explored repeatedly and proven<br />
successful (Noto et al., 2007). Specifically, silicon dioxide<br />
has been proven effective due to its ability to significantly<br />
increase the thermal properties and hydrophilicity<br />
of Nafion (Yu et al., 2007). Doping PAN with metal<br />
oxides should also increase its hydrophilicity. Tetraethyl<br />
orthosilicate (TEOS) has been explored as an additive to<br />
a polyvinylidene fluoride membrane (Liu et al., 2008).<br />
However, only the enhanced mechanical properties and<br />
effect on pore size were explored and the doped PVDF<br />
membrane was not tested in application for RFBs. TEOS<br />
has also been tested and used for the creation of a super<br />
hydrophilic surface in photovoltaic cells, proving its use as<br />
a hydrophilic material (Yan et al., 2015). Its hydrophilicity<br />
is due to the presence of hydroxide bonds after being<br />
polymerized and hydrolyzed, like the ones in Nafion (Fig.<br />
3).<br />
Figure 2. Nafion Structure: Nafion contains<br />
a hydrophobic fluorinated backbone with a<br />
hydrophilic sulfur hydroxide bond.<br />
As a result of its structure, Nafion maintains a high<br />
chemical stability while still allowing protons to cross over<br />
the membrane. The cost of Nafion is mainly due to the<br />
manufacturing cost of making fluorinated membranes.<br />
Therefore, non-fluorinated membranes have been<br />
investigated as a cheaper alternative. However, many<br />
lack the chemical stability of fluorinated membranes,<br />
which poses a challenge for their application in vanadium<br />
RFBs. Polyacrylonitrile has recently been recognized<br />
as a promising option for non-fluorinated membranes<br />
due to its high chemical stability despite the absence of<br />
fluorine. In fact, PAN has been explored in applications as<br />
a superhydrophobic polymer with a water contact angle<br />
CHEMISTRY<br />
Figure 3. Hydrolyzed and Polymerized TEOS: Just<br />
like Nafion’s hydrophilic sulfur hydroxide bonds,<br />
TEOS contains hydrophilic silicon hydroxide bonds.<br />
Doping PAN with TEOS should have the same effect<br />
that adding a metal oxide, like silicon oxide, would have.<br />
TEOS has also demonstrated an increase in the thermal<br />
stability of membranes (Liu et al., 2008). Therefore,<br />
doping PAN with TEOS should combine the superhydrophobicity<br />
and chemical stability of PAN with the<br />
super-hydrophilicity, thermal stability, and high tensile<br />
strength of TEOS without compromising the high<br />
chemical stability or proton/vanadium selectivity of PAN.<br />
Engineering a more efficient and less expensive proton<br />
exchange membrane will allow energy to be generated<br />
and stored for an indefinite period of time, making<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 27
enewable energy much more reliable and allowing entire<br />
cities to depend on it. The future of RFB membranes lies<br />
in the development of a non-fluorinated membrane, due<br />
to the cheaper manufacturing cost. Through testing how<br />
to improve the properties of a promising non-fluorinated<br />
membrane, progress is made towards engineering an ideal<br />
proton-exchange membrane.<br />
2. Methods<br />
2.1 – Preparing the Membranes<br />
The following procedure was adapted from Liu (2008).<br />
The nanoporous PAN membrane was prepared by casting<br />
a 15 wt% PAN (M w<br />
= 150,000), 3 wt% LiCl, and 4 wt%<br />
polyvinylpyrrolidone (PVP) solution in DMSO on a glass<br />
plate and leveling it with an RDS40 wire round rod (RD<br />
Specialties, USA). A normal phase inversion was then<br />
conducted using a water bath at room temperature. The<br />
membrane was left in the water bath for 1-2 days to<br />
remove any remaining solvents. The hybrid membrane<br />
was prepared by lowering the weight percentages to 12.5<br />
wt% PAN, 2.5 wt% LiCl, 3.25 wt% PVP, and adding 6.7<br />
wt% TEOS. A normal phase inversion was then conducted<br />
in an acid bath (pH = 1), to allow the polymerization of<br />
TEOS, and then transferred to a water bath for 1-2 days.<br />
2.2 – Model Redox Flow Battery<br />
The model redox flow battery was designed using<br />
two mini-variable flow peristaltic pumps (Fisher) and<br />
a modified fuel cell (Heliocentris). The fuel cell was<br />
composed of graphite, two carbon felt electrodes, and a<br />
membrane. The design is shown below (Fig. 4).<br />
60-90°C. After the dissolution of V 2<br />
O 5<br />
, temperature was<br />
maintained while .7mL of glycerol was stirred in. Once a<br />
uniform blue color was obtained, indicating the formation<br />
of V 4+ , the reaction was completed.<br />
2.4 – Proton/Vanadium Selectivity Test<br />
Proton vanadium selectivity was tested by filling one<br />
side of the fuel cell with water and the other side with<br />
2M VO 2<br />
+<br />
and 7M HCl. The pumps were run for 45<br />
minutes with samples being taken every 5 minutes. The<br />
concentration of vanadium in these samples was measured<br />
using a spectrophotometer. The absorption at 765nm<br />
was measured and Beer’s law was used to calculate the<br />
concentration using a molar absorptivity of 13.40 (Choi<br />
et al., 2013). The concentration of protons was measured<br />
through pH measurement using a pH meter (Vernier). All<br />
data were recorded in LoggerPro (Vernier).<br />
2.5 – Chemical Stability<br />
The chemical stability of the membranes was estimated<br />
by placing the membranes in a solution that consisted of<br />
1M V 2<br />
O 5<br />
and 5M HCl at 50°C for 30 days. The presence of<br />
VO 2<br />
+<br />
indicated that the membrane had been oxidized. The<br />
stability of the membrane was determined by calculating<br />
the percentage of VO 2<br />
+<br />
in the sample in comparison to a<br />
control solution without a membrane.<br />
3. Results<br />
3.1 – Water Contact Angle<br />
Figure 4. Model Redox Flow Battery Design: The<br />
design consists of two peristaltic pumps, two fuel<br />
tanks, and two adjacent half cells with electrodes<br />
separated by a membrane.<br />
2.3 – Vanadium 4+ Preparation<br />
All fuel was prepared through the reduction of V 2<br />
O 5<br />
to VO 2<br />
+<br />
using glycerol in the presence of HCl (Small et<br />
al., 2017). 38.9 mL of deionized water, 50.0 mL of 12.1M<br />
HCl, and 5.0g of V 2<br />
O 5<br />
was added to a beaker and stirred at<br />
Figure 5. Water contact angle of a drop of water on<br />
the membranes. The photo was taken on an IPhone<br />
6s and the angles were measured using Logger Pro.<br />
(A) Nafion (B) PAN (C) Hybrid<br />
28 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
Hydrophilicity of Nafion, PAN, and the hybrid<br />
membrane was demonstrated by measuring the contact<br />
angle of a water droplet (Fig. 5). It was found that Nafion<br />
had the largest water contact angle of 87.09° followed by<br />
PAN at 55.01° and the hybrid membrane at 42.27°.<br />
3.2 – Proton and Vanadium Crossover<br />
membrane the highest (Fig. 6A). The TEOS/PAN hybrid<br />
membrane also had the highest proton conductivity (Fig.<br />
6B). The overall proton/vanadium selectivity was similar<br />
for the Nafion and PAN membranes. However, the hybrid<br />
membrane had a much lower selectivity (Fig. 6C).<br />
3.3 – Chemical Stability<br />
Table 1. The percent of V 5+ reduced to V 4+ .This<br />
indicated ongoing oxidation of the membrane<br />
and therefore can be used to analyze the chemical<br />
stability of the membrane. Concentrations were<br />
measured using a spectrophotometer.<br />
Percent<br />
Vanadium<br />
Reduced (%)<br />
Percent<br />
Reduced<br />
Compared<br />
to reference<br />
Nafion PAN Hybrid Reference<br />
4.72% 1.20% 1.67% 2.41%<br />
96.06% -49.94% -30.54% ________<br />
The chemical stability of the prepared membranes was<br />
measured as the percent of the original vanadium that was<br />
reduced (Table 1), indicating ongoing oxidation of the<br />
membrane. The percent of vanadium reduced was highest<br />
for Nafion with 4.72% followed by the hybrid membrane<br />
with 1.67% and the PAN membrane with 1.20%. However,<br />
2.41% of the vanadium in the reference sample (the sample<br />
without a membrane) was reduced. When comparing the<br />
measured percentages to the reference percentages, it is<br />
found that the percent reduced in the Nafion was 96%<br />
higher than that of the reference sample, PAN was 49.9%<br />
smaller, and the hybrid was 30.5% smaller.<br />
4. Discussion and Conclusion<br />
Figure 6. (A) Vanadium Crossover (B) Proton<br />
Crossover (C) Proton/Vanadium Selectivity<br />
The vanadium and proton crossovers were measured<br />
over a 45 minute period. A model redox flow battery<br />
was used. An acidic V 4+ solution was placed on one side<br />
of the battery and deionized water on the other side.<br />
The concentration of vanadium was measured over time<br />
using a spectrophotometer and the proton concentration<br />
was measured using a Vernier pH probe. The proton/<br />
vanadium selectivity was determined as the ratio of the<br />
proton crossover to the vanadium crossover.<br />
Nafion was found to have the lowest vanadium<br />
crossover, PAN the second highest, and the hybrid<br />
The goal of this project was to demonstrate that<br />
hydrolyzed and polymerized TEOS could effectively<br />
increase hydrophilicity and provide a suitable substitute<br />
for Nafion in a vanadium redox flow battery. The results<br />
of the experiments are summarized in Table 2. Introducing<br />
TEOS to the PAN membrane improved its hydrophilicity<br />
as demonstrated by the water contact angle test (Fig. 5).<br />
The smaller the water contact angle, the more hydrophilic<br />
the material, because the water is not as repelled to the<br />
polymer. The hybrid membrane had a smaller water<br />
contact angle than both PAN and Nafion, indicating a<br />
high hydrophilicity. Its high hydrophilicity was further<br />
demonstrated when tested in a model redox flow battery.<br />
The hybrid membrane showed a higher proton crossover<br />
than both Nafion and PAN (Fig. 6). These tests show that<br />
the proton conductivity of the PAN membrane was most<br />
likely successfully increased. The chemical stability of<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 29
Table 2. Data Summary: The water contact angle, vanadium crossover, proton crossover, proton/vanadium<br />
selectivity, and chemical stability of Nafion, PAN, and the hybrid TEOS/PAN membrane.<br />
Water<br />
Contact<br />
Angle (°)<br />
Permeability to<br />
V 4+ (cm 2 min -1 )<br />
Permeability to<br />
H + (cm 2 min -1 )<br />
Proton/<br />
Vanadium<br />
Selectivity<br />
Chemical Stability (% Reduced<br />
Compared to Reference)<br />
Nafion 87.09 7.54x10 -5 3.77x10 -4 18.19 96.06%<br />
PAN 55.01 1.57x10 -4 7.83x10 -4 15.74 -49.94%<br />
Hybrid 42.27 2.37x10 -4 1.18x10 -5 8.75 -30.54%<br />
the membrane was also maintained with the addition of<br />
TEOS, as none of the membranes had a significant amount<br />
of oxidation occur in the presence of a strong oxidizer.<br />
However, the TEOS-PAN hybrid did show increased<br />
vanadium crossover. Prevention of vanadium crossover<br />
is an essential function of a proton-exchange membrane.<br />
The different oxidation states of vanadium on either side of<br />
the membrane need to remain unmixed while still allowing<br />
protons to cross over. Therefore, proton/vanadium<br />
selectivity is measured as the ability of the membrane to<br />
allow protons to cross over but prevent vanadium ions<br />
from crossing over. A higher proton/vanadium selectivity<br />
is ideal. However, the hybrid membrane displayed a<br />
lower proton/vanadium selectivity than both Nafion<br />
and the PAN membrane. Further optimization will need<br />
to be performed in order to improve proton/vanadium<br />
selectivity.<br />
There are several areas that can be explored to improve<br />
upon this research. The PAN and TEOS-PAN membranes<br />
were cast through a phase inversion method and developed<br />
nanopores, which allow these ions to cross over. The size<br />
of the nanopores has a significant effect on the selectivity<br />
of the membrane. To aid in the casting process, a lower<br />
polymer concentration was used for the hybrid membrane.<br />
However, this may have resulted in an increased pore size,<br />
causing the decreased selectivity and increased vanadium<br />
crossover. This could be examined with scanning electron<br />
microscopy to verify the pore sizes. It could be inferred<br />
that if the increased vanadium crossover is only due to an<br />
increased pore size, then the increased proton crossover<br />
is also only due to an increased pore size. However, it was<br />
demonstrated that the membrane was more hydrophilic<br />
in the water contact angle test. Therefore, the hybrid<br />
membrane should easily allow protons to cross over even<br />
with a reduced pore size.<br />
This study successfully demonstrated that the addition<br />
of hydrolyzed and polymerized TEOS to a PAN membrane<br />
was effective in increasing membrane hydrophilicity.<br />
Further research needs to be done to investigate if the<br />
addition of TEOS results in an increased vanadium crossover<br />
or if this could be overcome through the optimization of<br />
the pore size of the hybrid membrane. The membranes’<br />
properties should also be tested in a functional redox flow<br />
battery to test the effects of the increased properties on the<br />
efficiency of the battery. TEOS was proven as an effective<br />
additive to membranes in increasing their mechanical and<br />
thermal properties and hydrophilicity. A hybrid TEOS-<br />
PAN membrane with an optimized pore size may create<br />
a membrane with properties comparable to that of Nafion<br />
at a cheaper cost.<br />
5. Acknowledgments<br />
I would like to thank Dr. Michael Bruno for his help<br />
and mentorship throughout the research project as<br />
well as the help and support of my fellow Research in<br />
Chemistry peers. Finally, I would like to thank the NCSSM<br />
Foundation for funding my research project as it has been<br />
an invaluable experience.<br />
6. References<br />
Alotto, P., Guarnieri, M., Moro, F. (2014). Redox flow<br />
batteries for the storage of renewable energy: A review.<br />
Renewable and Sustainable Energy Reviews, 29, 325-335.<br />
doi:10.1016/j.rser.2013.08.001<br />
Choi, N. H., Kwon, S., Kim, H. (2013). Analysis of the<br />
Oxidation of the V(II) by Dissolved Oxygen Using UV-<br />
Visible Spectrophotometry in a Vanadium Redox Flow<br />
Battery. Journal of The Electrochemical Society, 160(6).<br />
doi:10.1149/2.145306jes<br />
Feng, L., Li, S., Li, H., Zhai, J., Song, Y., Jiang, L., Zhu,<br />
D. (2002). Super-Hydrophobic Surface of Aligned<br />
Polyacrylonitrile Nanofibers. Angewandte Chemie<br />
International Edition, 41(7), 1221-1223. doi:10.1002/1521-<br />
3773(20020402)41:73.0.co;2-g<br />
Karpushkin, E. A., Gvozdik, N. A., Stevenson, K. J.,<br />
Sergeyev, V. G. (2017). Membranes based on carboxylcontaining<br />
polyacrylonitrile for applications in vanadium<br />
redox-flow batteries. Mendeleev Communications,<br />
27(4), 390-391. doi:10.1016/j.mencom.2017.07.024<br />
Liu, X., Peng, Y., Ji, S. (2008). A new method to prepare<br />
organic–inorganic hybrid membranes. Desalination,<br />
221(1-3), 376-382. doi:10.1016/j.desal.2007.02.056<br />
30 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
Noto, V. D., Gliubizzi, R., Negro, E., Vittadello,<br />
M., Pace, G. (2007). Hybrid inorganic–organic proton<br />
conducting membranes based on Nafion and 5wt.% of<br />
M x<br />
O y<br />
(M=Ti, Zr, Hf, Ta and W). Electrochimica Acta,<br />
53(4), 1618-1627. doi:10.1016/j.electacta.2007.05.00<br />
Small, L. J., Pratt, H., Staiger, C., Martin, R. I., Anderson, T.<br />
M., Chalamala, B., Subarmanian, V. R. (2017). Vanadium<br />
Flow Battery Electrolyte Synthesis via Chemical Reduction<br />
of V 2<br />
O 5<br />
in Aqueous HCl and H 2<br />
SO 4<br />
. doi:10.2172/1342368<br />
Yan, H., Yuanhao, W., Hongxing, Y. (2015). TEOS/Silane-<br />
Coupling Agent Composed Double Layers Structure: A<br />
Novel Super-hydrophilic Surface. Energy Procedia, 75,<br />
349-354. doi:10.1016/j.egypro.2015.07.384<br />
Yu, J., Pan, M., Yuan, R. (2007). Nafion/Silicon oxide<br />
composite membrane for high temperature proton<br />
exchange membrane fuel cell. Journal of Wuhan<br />
University of Technology- Mater. Sci. Ed., 22(3), 478-481.<br />
doi:10.1007/s11595-006-3478-3<br />
Zhang, H., Zhang, H., Li, X., Mai, Z., Zhang, J. (2011).<br />
Nanofiltration (NF) membranes: The next generation<br />
separators for all vanadium redox flow batteries (VRBs)?<br />
Energy Environmental Science, 4(5), 1676. doi:10.1039/<br />
c1ee.<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 31
NOVEL SYNERGISTIC ANTIOXIDATIVE<br />
INTERACTIONS BETWEEN SOY LECITHIN AND<br />
CYCLODEXTRIN-ENCAPSULATED QUERCETIN IN A<br />
LIPID MATRIX<br />
Anirudh Hari<br />
Abstract<br />
Food oils stale via multiple mechanisms, the most damaging being oxidation by free radicals through reaction with oxygen<br />
in the air. Antioxidants are used to combat this oxidation, but many that are commonly used have carcinogenic properties.<br />
Quercetin is a safer polyphenolic phytochemical known to possess antioxidative properties in lipid matrices. Soy lecithin,<br />
a common food emulsifier primarily composed of phospholipids, also possesses antioxidative properties in lipid matrices,<br />
one of its primary mechanisms being the dispersion of less lipid-soluble antioxidants in the matrix. Phosphatidylcholine,<br />
the primary component of soy lecithin, is capable of forming a hydrogen bond from its polar head to a hydroxyl group of<br />
quercetin to create a complex known as a phenolipid. This phenolipid has a greater antioxidative effect than soy lecithin<br />
or quercetin do alone. However, one issue that remains prevalent is rapid degradation of quercetin in the lipid matrix.<br />
Beta-cyclodextrin is a ring-shaped molecule which can encapsulate quercetin, but it has not been tested for its ability to<br />
protect quercetin against degradation in oils.<br />
A novel phenolipid was formulated between a quercetin-cyclodextrin complex and soy lecithin, thus doubly encapsulating<br />
quercetin in order to potentially increase antioxidative lifetime by protecting the molecule while still maintaining the<br />
dispersive effect of lecithin. An accelerated oxidation test was conducted and time points were analyzed for radical<br />
scavenging activity. Results revealed that the novel phenolipid scavenged radicals more effectively than quercetin or<br />
lecithin by themselves, and also had a greater antioxidative lifetime, showing much higher radical scavenging activity than<br />
the quercetin-lecithin phenolipid and quercetin or lecithin alone after 12 days of the oxidation test. This implies novel<br />
applications for beta-cyclodextrins in the protection of polyphenolic antioxidants in lipid matrices.<br />
1. Introduction<br />
The oxidation of lipids is a major concern in the food<br />
industry, especially with unsaturated and polyunsaturated<br />
fats which are very sensitive to degradation (Ramadan et<br />
al., 2012). When excited by light, molecular oxygen in the<br />
air forms the superoxide anion, a free radical that oxidizes<br />
molecules in the oil, causing a radical chain reaction that<br />
results in cleavage of the double bonds in unsaturated fatty<br />
acids, resulting in their degradation into aldehydes and<br />
ketones. This process, known as oxidative rancidification,<br />
is responsible for the characteristic stale smell of old oils.<br />
While there are a number of ways to prevent oxidative<br />
rancidification, including wrapping containers with foil<br />
to prevent reactions catalyzed by sunlight and vacuumsealing<br />
containers to prevent interactions with oxygen,<br />
the most effective way to protect oils is the addition of<br />
antioxidants to scavenge free radicals (Judde et al., 2003).<br />
Antioxidants are used in the food industry to protect<br />
oils by reducing reactive free radicals. The addition of<br />
antioxidants to oils inhibits oxidative rancidification,<br />
slowing the rate of decline in oil quality (Ramadan et<br />
al., 2012). However, many of the most commonly used<br />
synthetic antioxidants, including butylated hydroxyanisole,<br />
butylated hydroxytoluene, propyl gallate, and tert-butyl<br />
hydroquinone, have been shown to promote carcinogenesis<br />
(National Toxicology Program).<br />
Flavonoids are a group of natural polyphenolic<br />
phytochemicals consisting of more than 4000<br />
molecules that vary in structure and properties.<br />
3,5,7,3′,4′-pentahydroxyflavone, also known as quercetin,<br />
is a yellow-colored flavonoid that possesses antioxidative<br />
properties in lipid matrices and is considered safe at much<br />
higher doses than most common synthetic antioxidants.<br />
Quercetin is used in the food industry as an alternative to<br />
synthetic antioxidants, but the degradation of quercetin<br />
via glycosylation at its hydroxyl groups is a major limit<br />
to its application in foods. The structural feature of<br />
quercetin most involved in its antioxidative mechanism is<br />
the hydroxyl group on the 4′ carbon, which can donate a<br />
hydrogen atom to a free radical to reduce it (Ozgen et al.,<br />
2016) (Fig. 1).<br />
Figure 1. Quercetin reduces a free radical, labelled R,<br />
by donating a hydrogen atom from its 4′ hydroxyl<br />
group to form a stable radical 4′-quercetin.<br />
32 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
Phospholipids are another class of antioxidants that<br />
have different antioxidative mechanisms than flavonoids.<br />
Soy lecithin is a mixture of amphipathic phospholipids,<br />
primarily phosphatidylcholine, and is a common<br />
emulsifying agent that also possesses antioxidative<br />
properties in lipid matrices. Soy lecithin has multiple<br />
antioxidative mechanisms by itself. The choline group on<br />
the phospholipid head of phosphatidylcholine is capable<br />
of accepting a free electron from radical molecules.<br />
Phospholipids also form an oxygen barrier at the<br />
atmospheric interface of oils to prevent oxidation (Judde<br />
et al., 2003).<br />
It has been shown that soy lecithin also helps to<br />
disperse other antioxidants present within the oil to allow<br />
them to scavenge radicals more efficiently. Quercetin<br />
and soy lecithin exhibit higher radical scavenging activity<br />
when mixed together than when tested individually. The<br />
catechol group on the 3′ and 4′ carbons of quercetin allows<br />
for intramolecular and intermolecular hydrogen bonding<br />
with phosphatidylcholine to create a “phenolipid” between<br />
quercetin and the phospholipid (Fig. 2). This phenolipid is<br />
fat soluble, increasing the accessibility of quercetin in oil.<br />
However, although soy lecithin fully surrounds quercetin<br />
in the phenolipid formation, it does not inhibit the<br />
breakdown of quercetin, which remains a limiting factor<br />
in its application (Ramadan et al., 2012).<br />
of beta-cyclodextrin may be able to bond with soy lecithin<br />
through hydrogen bonds to form a novel phenolipid,<br />
increasing the solubility of a complex of quercetin and<br />
beta-cyclodextrin in oil. This would form a double<br />
encapsulation of quercetin (Fig. 3). Beta-cyclodextrin<br />
could prevent the degradation of quercetin with its<br />
encapsulation of the molecule, while soy lecithin facilitates<br />
its dispersion in oil, allowing this novel phenolipid to have<br />
a higher antioxidative effect compared to quercetin or<br />
lecithin by themselves as well as an increased antioxidative<br />
lifetime. This could be important in controlling the rate of<br />
oxidation of quercetin both in food protection and medical<br />
applications.<br />
Figure 3. Potential double encapsulation of quercetin<br />
by beta-cyclodextrin and phosphatidylcholine.<br />
Figure 2. Quercetin-phosphatidylcholine phenolipid<br />
complex.<br />
Cyclodextrins are ring-shaped molecules that can<br />
encapsulate certain molecules through hydrogen bonding.<br />
Such encapsulation has been performed with quercetin<br />
and has shown an increase in solubility (Zheng et al.,<br />
2005). However, an important potential application of<br />
the cyclodextrin-quercetin complex that has not been<br />
previously investigated is the possible protection of<br />
quercetin from degradation.<br />
Since beta-cyclodextrin increases the water solubility<br />
of quercetin, it would decrease the lipid solubility, making<br />
the cyclodextrin-quercetin complex unsuitable for use in a<br />
lipid matrix by itself. However, the outer hydroxyl groups<br />
It was hypothesized that a phenolipid between lecithin<br />
and the quercetin-cyclodextrin complex would increase<br />
availability of quercetin in sunflower oil, and that this<br />
phenolipid would have a greater antioxidative lifetime<br />
than the phenolipid made of quercetin and lecithin.<br />
Alternatively, the double encapsulation may prevent<br />
degradation of quercetin without increasing the availability<br />
of quercetin in sunflower oil.<br />
In the present study, a molecular docking model of the<br />
encapsulation of quercetin by beta-cyclodextrin indicated<br />
that in the most stable conformation, the 4′ hydroxyl group<br />
important to the antioxidative mechanism of quercetin is<br />
not encompassed by the cyclodextrin, while the rest of<br />
the quercetin molecule is, suggesting that quercetin could<br />
retain its antioxidative ability in the beta-cyclodextrin<br />
complex.<br />
The novel phenolipid was prepared along with<br />
the quercetin-lecithin phenolipid and the quercetincyclodextrin<br />
complex. Each antioxidant sample was mixed<br />
in sunflower oil and incubated in an oven to accelerate<br />
oxidation. Radical scavenging activity assay was conducted<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 33
periodically to measure reduction in antioxidant<br />
effectiveness over time. Results from radical scavenging<br />
activity assay indicate that the doubly encapsulated<br />
quercetin does have a greater antioxidative lifetime than<br />
the quercetin-lecithin phenolipid, and it also has a greater<br />
initial ability to scavenge radicals than quercetin or soy<br />
lecithin alone. This suggests a new potential application of<br />
beta-cyclodextrins to allow antioxidants to last longer in<br />
oils, which would also increase the lifetime of the oils due<br />
to longer term protection from oxidative rancidification.<br />
2. Materials and Methods<br />
2.1 – Molecular Modeling<br />
Before experimentation, the encapsulation of quercetin<br />
by beta-cyclodextrin was computationally modelled<br />
by molecular docking using PatchDock. PDB files for<br />
quercetin and beta-cyclodextrin were obtained and fed<br />
into the server, which found the most stable conformation<br />
of quercetin inside beta-cyclodextrin using shape<br />
complementarity and electrostatic interactions.<br />
2.2 – Encapsulation of Quercetin in Beta-Cyclodextrin<br />
Quercetin dihydrate, 97% (Alfa Aesar) was encapsulated<br />
in beta-cyclodextrin hydrate, 99% (Acros Organics) using<br />
physical kneading. An equimolar ratio of quercetin and<br />
beta-cyclodextrin powder was mixed in a mortar using a<br />
pestle for 10 minutes. Then, a small amount of ethanol<br />
(Fisher <strong>Scientific</strong>) was added and the mixture was kneaded<br />
for 40 more minutes. After kneading, the mixture was<br />
dried in a vacuum desiccator for 24 hours.<br />
50 mg of the dried mixture was dissolved in 50 mL<br />
of acetonitrile (Fisher <strong>Scientific</strong>), causing the betacyclodextrin<br />
and quercetin+beta-cyclodextrin complexes<br />
to precipitate, while the free quercetin that did not<br />
get complexed remained in solution. The absorbance<br />
spectrum of quercetin was taken using a Vernier UV-<br />
Vis spectrophotometer in a Hellma QS 282 1.000 quartz<br />
cuvette, showing 2 UV peaks: one at 260 nm and one at<br />
370 nm. A standard curve was made with absorption at<br />
370 nm as a function of quercetin concentration (Santos et<br />
al., 2015) (Fig. 4).<br />
Figure 4. Standard curve of quercetin in acetonitrile<br />
at 370 nm.<br />
The solution of complex in acetonitrile was allowed<br />
to settle for 3 days. The concentration of free quercetin<br />
in solution was determined using the standard curve.<br />
This concentration was compared to the total quercetin<br />
concentration in the solution, and the entrapment<br />
efficiency (EE) was determined using the following<br />
equation:<br />
free quercetin concentration<br />
EE = 1 -<br />
total quercetin concentration<br />
2.3 – Formation of Phenolipid Complexes<br />
The complex was removed from acetonitrile solution<br />
by vacuum filtration and mixed with soy lecithin (Alfa<br />
Aesar) at a 3:97 ratio complex to lecithin by mass. The<br />
complex was then dissolved in 10 mL ethyl acetate (Fisher<br />
<strong>Scientific</strong>). Several control groups were also dissolved<br />
in ethyl acetate: quercetin, quercetin with lecithin 3:97<br />
(phenolipid), and quercetin encapsulated in cyclodextrin<br />
without lecithin.<br />
Each sample was incubated at 40°C for 24 hours to<br />
facilitate dissolution. The samples were then dried by<br />
creating a vacuum within a chamber using a Chemglass<br />
<strong>Scientific</strong> Apparatus Vacuum for 2 hours.<br />
2.4 – Accelerated Oxidation<br />
Each sample was added to 100% sunflower oil (Loriva,<br />
cold pressed) at a concentration of 500 parts per million.<br />
The Schaal oven accelerated oxidation test was run on the<br />
4 samples as well as a sample with only sunflower oil as<br />
a negative control. Each mixture was placed in a 20 mL<br />
clear glass bottle. Each bottle was completely sealed and<br />
incubated in an oven at 60°C (Ramadan et al., 2012).<br />
Samples were withdrawn at 0, 3, 9, and 12 days and<br />
analyzed by Radical Scavenging Activity (RSA) assay.<br />
1,1-Diphenyl-2-picrylhydrazyl (DPPH) radical (Alfa Aesar)<br />
was dissolved in reagent grade toluene (Fisher <strong>Scientific</strong>)<br />
at a concentration of 10-4 M. 10 mg of each experimental<br />
sample was dissolved in 100 µL of toluene. This solution<br />
was mixed with 390 µL of the DPPH solution, and the<br />
mixture was vortexed at maximum speed for 20 seconds<br />
at ambient temperature. The decrease in absorbance at<br />
515 nm between the time of making the mixture and 1<br />
hour later was measured in a quartz cuvette using a UV-<br />
Vis spectrophotometer. As a control, radical scavenging<br />
activity towards the toluenic DPPH solution was measured<br />
without addition of sample. Percent inhibition was<br />
calculated by comparing the absorbance after 1 hour of the<br />
control to each of the test samples:<br />
% inhibition =<br />
abs of control - abs of test sample<br />
abs of control<br />
RSA was measured as the difference in 515 nm<br />
absorption between the beginning and end of the assay.<br />
RSA was compared between each time-point taken for<br />
each sample (Ramadan et al., 2012).<br />
34 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
3. Results and Discussion<br />
In order to determine if quercetin would likely maintain<br />
its antioxidative properties while encapsulated by betacyclodextrin,<br />
molecular docking was computationally<br />
modelled. In the lowest energy conformation, the 4′<br />
hydroxyl group of quercetin, shown in light blue, extends<br />
out of the cyclodextrin ring, while the rest of the molecule<br />
sits inside the cyclodextrin (Fig. 5). This suggests that<br />
beta-cyclodextrin can protect quercetin from degradation<br />
without compromising its effectiveness as an antioxidant.<br />
radical scavenging activity with the least decrease in<br />
activity over time, while the samples that did not include<br />
cyclodextrin scavenged radicals less effectively after 12<br />
days, degrading more quickly. The quercetin encapsulated<br />
with cyclodextrin without lecithin also showed increased<br />
antioxidative lifetime compared to quercetin alone (Fig. 7,<br />
8).<br />
Figure 6. Radical scavenging activity assay was<br />
conducted immediately after mixing the antioxidant<br />
formulations in sunflower oil.<br />
Figure 5. Molecular docking model of quercetin<br />
in beta-cyclodextrin. Beta-cyclodextrin is shown<br />
in pink, quercetin is shown in yellow, and the<br />
4′ hydroxyl group of quercetin is shown in light<br />
blue. The 4′ hydroxyl group is key to quercetin’s<br />
antioxidative effect.<br />
According to the hypothesis, the doubly encapsulated<br />
quercetin formulation would scavenge radicals more<br />
effectively than quercetin or lecithin alone before the<br />
acceleration test, and have a smaller decrease in radical<br />
scavenging activity over time than the quercetin-lecithin<br />
phenolipid formulation. The entrapment of quercetin<br />
in beta-cyclodextrin was successful, and the entrapment<br />
efficiency was determined by UV absorbance to be 45%.<br />
Radical scavenging activity assay conducted on the<br />
day the complexes were mixed in sunflower oil revealed<br />
that the quercetin-lecithin phenolipid formulation had<br />
the highest radical scavenging activity, followed by the<br />
novel double encapsulation formulation. Quercetin and<br />
soy lecithin alone had similar radical scavenging activity<br />
results (Fig. 6).<br />
Samples in the Schaal oven accelerated oxidation test<br />
were withdrawn at 3, 6, and 12 days and assayed for<br />
radical scavenging activity. After incubation for 12 days,<br />
the doubly encapsulated quercetin sample had the highest<br />
Figure 7. Radical scavenging activity was assayed<br />
at 3, 6, and 12 days after initiating the accelerated<br />
oxidation test.<br />
Figure 8. Decrease of RSA after 12 days of oxidation<br />
test compared to RSA before oxidation test was<br />
begun.<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 35
The higher initial RSA of the doubly-encapsulated<br />
quercetin compared to quercetin and lecithin alone<br />
suggests that the polar head of phosphatidylcholine did<br />
hydrogen bond to the exterior hydroxyl groups of the<br />
quercetin-cyclodextrin complex, dispersing the quercetin<br />
in the sunflower oil as hypothesized (Fig. 6). The initial<br />
RSA of the quercetin-cyclodextrin complex without<br />
lecithin was lowest of the groups tested, which was<br />
expected since cyclodextrin increases the water-solubility<br />
of quercetin, decreasing its availability in sunflower oil.<br />
The higher antioxidative lifetimes of both the<br />
doubly encapsulated quercetin and the single quercetincyclodextrin<br />
complex suggest that beta-cyclodextrin<br />
provides protection to quercetin from degradation in<br />
sunflower oil, consistent with the hypothesis (Fig. 6, 7).<br />
4. Conclusion<br />
The use of cyclodextrins in the protection of flavonoidclass<br />
antioxidants from degradation in lipid matrices has<br />
been unexplored, as have phenolipid bonds between<br />
cyclodextrins and phospholipids. The novel phenolipid<br />
formulation constructed in this experiment, consisting of<br />
quercetin doubly encapsulated in beta-cyclodextrin and<br />
soy lecithin, had a higher radical scavenging activity than<br />
quercetin or soy lecithin alone and a higher antioxidative<br />
lifetime than known phenolipid formulations of quercetin<br />
and lecithin. These results indicate that cyclodextrins<br />
can increase the antioxidative lifetime of flavonoids<br />
without compromising antioxidative ability if paired<br />
with a phospholipid to disperse the complex in the<br />
lipid matrix, opening up new avenues of lipid oxidation<br />
research with applications in food oils. Future work<br />
would include repeating the accelerated oxidation test<br />
and radical scavenging activity assays for improved<br />
statistical significance, as well as testing different types of<br />
polyphenols, phospholipids, and oils to determine whether<br />
the same effects are observed.<br />
5. Acknowledgments<br />
I would like to thank Dr. Michael Bruno for selecting<br />
me for the Research in Chemistry program and providing<br />
guidance throughout the development and execution of my<br />
project. I would also like to thank the NCSSM Foundation<br />
for providing funding for the purchase of materials and<br />
equipment used in my experimentation.<br />
6. References<br />
Di Donato, C., et al. (2016). Alpha- and Beta-Cyclodextrin<br />
Inclusion Complexes with 5-Fluorouracil: Characterization<br />
and Cytotoxic Activity Evaluation. Molecules, 21(12),<br />
1644. doi:10.3390/molecules21121644<br />
Judde, A., Villeneuve, P., Rossignol-Castera, A., & Guillou,<br />
A. L. (2003). Antioxidant effect of soy lecithins on vegetable<br />
oil stability and their synergism with tocopherols. Journal<br />
of the American Oil Chemists Society, 80(12), 1209-1215.<br />
doi:10.1007/s11746-003-0844-4<br />
Kahveci, D., Laguerre, M., & Villeneuve, P. (2015).<br />
Phenolipids as New Antioxidants: Production, Activity,<br />
and Potential Applications. Polar Lipids, 185-214.<br />
doi:10.1016/b978-1-63067-044-3.50011-x<br />
National Toxicology Program (2001). Carcinogens<br />
Nominated for 11th Report on Carcinogens. JNCI Journal<br />
of the National Cancer Institute, 93(18), 1372-1372.<br />
doi:10.1093/jnci/93.18.1372-a<br />
Ozgen, S., Kilinc, O. K., & Selamoğlu, Z. (2016). Antioxidant<br />
Activity of Quercetin: A Mechanistic Review. Turkish<br />
Journal of Agriculture - Food Science and Technology,<br />
4(12), 1134. doi:10.24925/turjaf.v4i12.1134-1138.1069<br />
Panya, A., Laguerre, M., Bayrasy, C., Lecomte, J.,<br />
Villeneuve, P., Mcclements, D. J., & Decker, E. A. (2012).<br />
An Investigation of the Versatile Antioxidant Mechanisms<br />
of Action of Rosmarinate Alkyl Esters in Oil-in-Water<br />
Emulsions. Journal of Agricultural and Food Chemistry,<br />
60(10), 2692-2700. doi:10.1021/jf204848b<br />
Ramadan, M. F. (2012). Antioxidant characteristics<br />
of phenolipids (quercetin-enriched lecithin) in lipid<br />
matrices. Industrial Crops and Products, 36(1), 363-369.<br />
doi:10.1016/j.indcrop.2011.10.008<br />
Santos, E. H., Kamimura, J. A., Hill, L. E., & Gomes, C.<br />
L. (2015). Characterization of carvacrol beta-cyclodextrin<br />
inclusion complexes as delivery systems for antibacterial<br />
and antioxidant applications. LWT - Food Science and<br />
Technology, 60(1), 583-592. doi:10.1016/j.lwt.2014.08.046<br />
Tanhuanpää, K., Cheng, K. H., Anttonen, K., Virtanen,<br />
J. A., & Somerharju, P. (2001). Characteristics of Pyrene<br />
Phospholipid/ γ -Cyclodextrin Complex. Biophysical<br />
Journal, 81(3), 1501-1510. doi:10.1016/s0006-<br />
3495(01)75804-3<br />
Zheng, Y., Haworth, I. S., Zuo, Z., Chow, M. S., &<br />
Chow, A. H. (2005). Physicochemical and Structural<br />
Characterization of Quercetin-β-Cyclodextrin Complexes.<br />
Journal of Pharmaceutical Sciences, 94(5), 1079-1089.<br />
doi:10.1002/jps.20325<br />
36 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
UTILIZATION OF ATOMIC LAYER DEPOSITION TO<br />
CREATE NOVEL METAL OXIDE PHOTOANODES FOR<br />
SOLAR-DRIVEN WATER SPLITTING<br />
Annie Wang<br />
Abstract<br />
A major obstacle of dye-sensitized photoelectrosynthesis cells is the recombination of 60% of the injected electrons from<br />
the dye into the photoanode. Creating core/shell structures is one technique of slowing down electron recombination.<br />
There has been no work done on TiO 2<br />
/SnO 2<br />
structures or on TiO 2<br />
/TiO 2<br />
structures using atomic layer deposition, so the<br />
aim of the project was to successfully deposit these materials, optimize the deposition, and compare the behavior of the<br />
structures to the standard SnO 2<br />
/TiO 2<br />
core/shell. Novel deposition of TiO 2<br />
and SnO 2<br />
onto mesoporous TiO 2<br />
thin films<br />
was achieved using atomic layer deposition with the TDMAT and TDMASn precursors. Subsequently, the dye loading<br />
capabilities of the core/shell structures were measured after being loaded with the RuP chromophore. The samples were<br />
characterized through XPS after varying deposition parameters to optimize deposition conditions in order to create TiO 2<br />
and SnO 2<br />
shells of comparable thicknesses. Dye loading onto TiO 2<br />
/TiO 2<br />
was found to be affected by parameters other<br />
than pore size, including type of TiO 2<br />
used and processing conditions. Deposition of SnO 2<br />
initially resulted in SnO, but<br />
TiO 2<br />
/SnO 2<br />
structures were able to be synthesized by using dyesol TiO 2<br />
instead of mixed-phase TiO 2<br />
. The successfully<br />
created TiO 2<br />
/SnO 2<br />
and TiO 2<br />
/TiO 2<br />
core/shells can be studied to differentiate competing electron recombination theories.<br />
1. Introduction<br />
As the world is becoming increasingly dependent on<br />
our dwindling supply of nonrenewable sources of energy,<br />
clean energy is the only viable long-term option. A<br />
promising method for solar energy conversion is the use<br />
of dye-sensitized photoelectrosynthesis cells (DSPECs)<br />
(Brennaman et al., 2016). The DSPEC shares similar design<br />
features and applies similar principles as the dye-sensitized<br />
solar cell (DSSC), and although less developed, holds<br />
much promise for the future of solar energy conversion.<br />
Photoelectrosynthesis cells convert light to chemical<br />
energy in the form of stored hydrogen fuel. Rather than<br />
producing electrical energy as in solar cells, DSPECs use<br />
photons from sunlight to split water into hydrogen and<br />
oxygen gases (Fujishima & Honda, 1972). The oxidation<br />
of water occurs at the anode and the reduction of hydrogen<br />
occurs at the cathode. The key advantage of this model<br />
is that hydrogen is able to be stored as chemical fuel for<br />
future use. Photoanodes used in these cells are often made<br />
of metal oxide semiconductors due to their ability to form<br />
high surface area films, ability to accept photoinjected<br />
electrons from dye molecules, and transparency in the<br />
visible spectrum because of their optimally high band gap<br />
energies. (Ashford et al., 2015).<br />
In addition, a crucial component of the DSPEC is the<br />
electron injection from chromophores (dye molecules)<br />
attached to the surface of the mesoporous (containing<br />
pores with diameters between 2 and 50 nm) film into<br />
the semiconductor. It is therefore essential to minimize<br />
undesired back electron transfer (BET) in these devices.<br />
Back electron transfer/electron recombination occurs<br />
CHEMISTRY<br />
when electrons injected into the semiconductor conduction<br />
band recombine with the oxidized dye, which ultimately<br />
results in lower DSPEC performance because the electrons<br />
are not able to travel to the cathode to reduce hydrogen.<br />
One technique used to slow BET rates in DSPECs is<br />
the application of SnO 2<br />
/TiO 2<br />
core/shell photoanode<br />
structures (Bakke et al., 2011). Core/shell structures<br />
allow for electron injection without interference, while<br />
maintaining a barrier against electron recombination. It<br />
has been proven by many past studies that these structures<br />
greatly reduce back electron transfer and enhance DSPEC<br />
efficiencies (Gish et al., 2016). There is still much debate<br />
over the underlying theory of how electron recombination<br />
is reduced in core/shell structures. Two competing<br />
theories shown in Figure 1a and 1b include the band edge<br />
offset model (proposing an energy barrier created by the<br />
difference in band edge between the core and shell) and a<br />
model proposing the existence of a unique electronic state<br />
at the core/shell interface (James et al., <strong>2018</strong>).<br />
To study this more closely, it is therefore necessary to<br />
create samples with different band edges for the core and<br />
shell as well as structures with the core and shell made<br />
of the same material in order to compare their electron<br />
kinetics. In addition, the different samples must have<br />
comparable shell thicknesses.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 37
Figure 1a. In the band edge model, an energy barrier<br />
created by conduction band (CB) edge differences<br />
prevents electrons from traveling back and<br />
recombining from the fluorine doped tin oxide<br />
(FTO).<br />
Figure 1b. In this model, a unique electronic state<br />
between the core and shell (left) exhibits special<br />
properties that cause a change in electron transfer<br />
behavior, in contrast to electronic states within the<br />
core and shell (right).<br />
Atomic layer deposition (ALD) is one method of<br />
creating core/shell structures (George, 2010). The<br />
technique involves depositing the shell layer onto<br />
nanoparticles through successive self-limiting reactions on<br />
the surface of the material. ALD consists of multiple cycles<br />
of precursor pulsing and purging to obtain extremely<br />
precise monolayers on the Angstrom scale. Due to its selflimiting<br />
nature, ALD produces very smooth, conformal<br />
films because all parts of the surface react completely with<br />
the precursor to grow the film (Wang et al., 2017). ALD<br />
has been used widely to create metal oxide films such as<br />
Al 2<br />
O 3<br />
, TiO 2<br />
, ZnO, ZrO 2<br />
, SiO 2<br />
, and VO 2<br />
(George, 2010).<br />
This study will mainly focus on deposition of SnO 2<br />
and<br />
TiO 2<br />
on TiO 2<br />
. TiO 2<br />
has thus far produced the highest light<br />
conversion efficiencies out of all the metal oxides, and is<br />
widely used in DSSCs as a photoanode (Jafari et al., 2016).<br />
SnO 2<br />
also has favorable characteristics for its anodic<br />
abilities, such as its stability, high reversible capacity, nontoxicity<br />
and low cost (Knauf et al., 2015).<br />
The goal of this study was to successfully synthesize<br />
and characterize TiO 2<br />
/SnO 2<br />
and TiO 2<br />
/TiO 2<br />
core/<br />
shell nanostructures using tetrakis(dimethylamido)<br />
titanium (TDMAT) and tetrakis(dimethylamido)tin(IV)<br />
(TDMASn) precursors. Since TiO 2<br />
/SnO 2<br />
has not been<br />
created before, the hypothesis was that TiO 2<br />
/SnO 2<br />
would<br />
behave similarly to the more common SnO 2<br />
/TiO 2<br />
core<br />
shells and would help differentiate the mechanism actually<br />
in use by core/shell structures to inhibit recombination.<br />
Previously, it had been common practice to deposit TiO 2<br />
onto TiO 2<br />
by treating the TiO 2<br />
thin film with a TiCl 4<br />
chemical bath deposition, which was demonstrated to<br />
reduce back electron transfer (Lee et al., 2012). However,<br />
this method is very unreliable and difficult to control.<br />
According to the hypothesis, it would be possible to<br />
create TiO 2<br />
/TiO 2<br />
core/shell structures using ALD for<br />
the first time which would allow a much more controlled<br />
deposition while still reducing electron recombination.<br />
The second aim of this project was therefore to deposit<br />
TiO 2<br />
on TiO 2<br />
using solely atomic layer deposition, a much<br />
more controllable and reproducible method.<br />
The TiO 2<br />
-TiO 2<br />
deposition was in fact found to be<br />
successful without necessitating the TiCl 4<br />
treatment which<br />
was previously utilized to create TiO 2<br />
/TiO 2<br />
structures.<br />
It was also found that using the TDMASn precursor to<br />
deposit tin resulted in stannous oxide (SnO) rather than<br />
the expected SnO 2<br />
. After thorough studies, the stannous<br />
oxide was successfully removed by using pure anatase<br />
dyesol TiO 2<br />
paste, a commercial paste, for the thin films<br />
instead of mixed-phase TiO 2<br />
. In addition, dye loading was<br />
measured for each of the samples. It was found that dye<br />
loading in TiO 2<br />
/TiO 2<br />
slides does not decrease consistently<br />
as in SnO 2<br />
/TiO 2<br />
slides, so there are other factors besides<br />
pore size that have an effect on dye loading.<br />
2. Materials and Methods<br />
2.1 – Thin Film Preparation<br />
FTO (fluorine doped tin oxide) glass plates were<br />
washed in an ultrasonic bath immersed in ethanol, then<br />
acetone, for 20 minutes each. Previously prepared TiO 2<br />
paste was coated on the slides through doctor blading<br />
and tape-casting. The thin films were stored in a 125°C<br />
oven to prevent water adsorption on the TiO 2<br />
. They were<br />
then sintered at 450°C for 60 minutes with a 120 minute<br />
ramp-up time. Selected films were annealed at 450°C for<br />
30 minutes with a 120 minute ramp up time.<br />
38 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
2.2 – Atomic Layer Deposition<br />
Atomic layer deposition was conducted using<br />
an Ultratech/Cambridge Nanotech Savannah S200.<br />
TDMASn and TDMAT precursor reactant gases were<br />
transported to the reactor chamber through heated gas<br />
lines using nitrogen carrier flow. Nitrogen gas was used<br />
to purge the reactant chamber after each precursor step.<br />
Deposition was performed at 150°C while TDMAT and<br />
TDMASn were held at 75 °C and 60°C, respectively. Gas<br />
flow and purge times were controlled electronically by a<br />
LabVIEW sequencer.<br />
minimized with additional shell coatings. After examining<br />
the results, it can be concluded that the change in pore size<br />
is not the only factor affecting dye loading levels. Rather,<br />
it is hypothesized that there is an uneven preferential<br />
deposition of TiO 2<br />
onto the TiO 2<br />
causing increased dye<br />
loading which does not occur on the SnO 2<br />
thin films.<br />
The decrease in dye loading from 25 to 45 cycles can be<br />
attributed to the decreased pore size, the effect of which<br />
eventually overbears that of the preferential deposition<br />
and leads to an overall decrease in dye loading.<br />
2.3 – Dye Loading<br />
The RuP chromophore was loaded to the films by<br />
soaking the slides in anhydrous methanol solutions<br />
containing 0.0003 M RuP for several days. The slides were<br />
removed and subsequently soaked in methanol to remove<br />
unadsorbed dye. UV-vis absorbances of the dye-loaded<br />
thin films were taken in 0.1 M HClO 4<br />
using a Cary 60 UV−<br />
vis absorbance spectrophotometer.<br />
2.4 – Characterization<br />
Profilometry measurements were done with a Bruker<br />
Optics DektakXT® stylus profiler. All films were between<br />
4-6 μm thick. Characterization of the deposited thin films<br />
was done through infrared spectroscopy using a Bruker<br />
Optics Alpha FTIR Spectrometer, transmission electron<br />
microscopy using a TEM JEOL 2010F-FasTEM, X-ray<br />
photoelectron spectroscopy (XPS) using a Kratos Axis<br />
Ultra DLD X-ray Photoelectron Spectrometer, and Raman<br />
spectroscopy using a Renishaw inVia Raman microscope.<br />
Ellipsometry to measure ALD-deposited shell thickness<br />
was also conducted using a JA Woollam ellipsometer. All<br />
data were analyzed using Igor Pro (WaveMetrics Inc.).<br />
Figure 2a. Infrared spectrum of TiO 2<br />
/TiO 2<br />
core/<br />
shells of various numbers of ALD cycles confirming<br />
successful deposition.<br />
3. Results<br />
The goal of this project was to deposit both SnO 2<br />
and<br />
TiO 2<br />
onto mesoporous TiO 2<br />
thin films using atomic layer<br />
deposition (ALD).<br />
3.1 – TiO 2<br />
/TiO 2<br />
Deposition<br />
The TiO 2<br />
/TiO 2<br />
deposition was successfully achieved<br />
using ALD with the TDMAT and water precursors.<br />
The slides were characterized using FTIR and TEM and<br />
confirmed to have shells made of the correct material (Fig.<br />
2). Following this, the dye loading of the samples was<br />
collected (Fig. 3). The results reveal different trends from<br />
those of the more commonly studied SnO 2<br />
/TiO 2<br />
core/<br />
shell structures. While the data show a clear continuous<br />
decrease in dye loading of SnO 2<br />
/TiO 2<br />
with increasing<br />
ALD cycles, the dye loading of TiO 2<br />
/TiO 2<br />
increases from<br />
0 to 25 cycles and then decreases. This is inconsistent<br />
with previous theories that suggested dye loading always<br />
decreases with increasing ALD cycles because pore sizes are<br />
CHEMISTRY<br />
Figure 2b. TEM image of an anatase TiO 2<br />
nanoparticle<br />
with an amorphous TiO 2<br />
shell created using 20 ALD<br />
cycles.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 39
Figure 3. Dye loading of mixed phase TiO 2<br />
/TiO 2<br />
samples and SnO 2<br />
/TiO 2<br />
samples.<br />
This phenomenon of an initial increase then decrease in<br />
dye loading was further studied with dyesol (pure-phase)<br />
TiO 2<br />
/TiO 2<br />
samples, annealed and unannealed, as well as<br />
annealed mixed phase TiO 2<br />
/TiO 2<br />
samples (Fig. 4). Dyesol<br />
TiO 2<br />
contains larger pores and anneals more easily than<br />
mixed-phase TiO 2<br />
. In addition, mixed-phase TiO 2<br />
creates<br />
a less well-connected film. All of the samples exhibited<br />
higher dye loading than the unannealed mixed phase TiO 2<br />
/<br />
TiO 2<br />
samples. The data clearly display an overall trend for<br />
each sample type. The unannealed dyesol slides increase<br />
initially in dye loading but decrease starting at 35 cycles,<br />
while the annealed dyesol slides demonstrate the same<br />
behavior but do not decrease in dye loading until 40 cycles.<br />
These results are consistent with the trends observed<br />
for the unannealed mixed phase TiO 2<br />
/TiO 2<br />
structures.<br />
The annealed mixed phase TiO 2<br />
/TiO 2<br />
samples, however,<br />
continuously decrease in dye loading from 0 all the way<br />
to 50 cycles, suggesting that annealing the samples affects<br />
the dye loading behavior of mixed phase TiO 2<br />
. Based on<br />
the results, it can be concluded that dye loading is not<br />
solely determined based on pore size and can be affected<br />
by different processing conditions as well as the type of<br />
TiO 2<br />
used.<br />
Figure 4a. Dye loading on dyesol TiO 2<br />
/TiO 2<br />
slides<br />
created with dyesol TiO 2<br />
paste, unannealed.<br />
Figure 4b. Dye loading on dyesol TiO 2<br />
/TiO 2<br />
slides<br />
created with dyesol TiO 2<br />
paste, annealed.<br />
Figure 4c. Dye loading on TiO 2<br />
/TiO 2<br />
slides created<br />
with mixed phase TiO 2<br />
paste, annealed.<br />
3.2 – TiO 2<br />
/SnO 2<br />
Deposition<br />
The TDMASn precursor deposition was initially<br />
performed with the standard recipe used for the mixed<br />
phase TiO 2<br />
/TiO 2<br />
structures and resulted in a brown layer<br />
on the slides, which is not the normal appearance of SnO 2<br />
shells. Upon further characterization, the layer was found<br />
to be SnO. The SnO formation can be attributed to the<br />
poor oxidative properties of water. Furthermore, the ALD<br />
growth rate of SnO 2<br />
is naturally higher than that of TiO 2<br />
.<br />
When the same recipe is used for depositing both SnO 2<br />
40 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
and TiO 2<br />
, there is a larger growth per cycle for SnO 2<br />
. Thus,<br />
it was necessary to vary the ALD conditions in order to<br />
gain a better understanding of how each parameter affects<br />
growth so the SnO 2<br />
growth per cycle could be equalized to<br />
the TiO 2<br />
growth per cycle.<br />
Numerous attempts were made to first convert the<br />
SnO to SnO 2<br />
by changing the deposition parameters.<br />
The TDMASn deposition was initially attempted using<br />
an ozone precursor and using a combination of water<br />
and ozone precursors, but both approaches still resulted<br />
in SnO shells instead of SnO 2<br />
. In addition, an increase<br />
of the ALD reactor chamber temperature from 150°C to<br />
250°C resulted in an uncontrolled island-like growth of<br />
the SnO, as calculated from ellipsometry measurements of<br />
the shell thicknesses. The exponential growth of the SnO<br />
thicknesses for the 250°C samples (Fig. 5) indicates the<br />
uncontrolled nature of the deposition, which often results<br />
in island-like growth rather than a smooth conformal<br />
coating as desired.<br />
After testing the effects of increasing the reactor<br />
temperature and using the ozone precursor, a postdeposition<br />
heat treatment was administered at 210°C in an<br />
attempt to remove the SnO, but this was unsuccessful and<br />
resulted in increased SnO peaks in the Raman spectra of<br />
the sample (Fig. 6). Next, the films were annealed at 450°C<br />
in an effort to convert the existing SnO to SnO 2<br />
. Although<br />
the SnO was successfully converted, the annealing process<br />
led to delamination of the TiO 2<br />
. This occurred because<br />
the extreme heat induced expansion of the TiO 2<br />
, but<br />
the rigidity of the crystal structure forced the TiO 2<br />
to<br />
eventually crack and delaminate from the slide due to<br />
internal pressure.<br />
Figure 5. Comparison of growth rates of SnO on<br />
planar silicon at 150°C and 250°C based on ellipsometry<br />
of shell thickness.<br />
Figure 6. Raman spectra characterizing TiO 2<br />
/SnO 2<br />
samples before and after heat treatments.<br />
The SnO 2<br />
deposition was then attempted using a<br />
H 2<br />
O 2<br />
precursor instead of the water precursor with other<br />
varying parameters. H 2<br />
O 2<br />
is a stronger oxidant than water<br />
and does not degrade as easily as ozone, so it offered a<br />
possible option to convert the SnO to SnO 2<br />
during the<br />
ALD process. In order to study the effects of varying each<br />
parameter, samples with varying precursor pulse, hold,<br />
and purge times were created and analyzed through XPS<br />
and ellipsometry on planar silicon. X-ray photoelectron<br />
spectroscopy (XPS) is a spectroscopic technique that is<br />
used to analyze the elemental composition of the surface<br />
of a material by measuring the kinetic energy of escaped<br />
electrons after focusing a beam of X-rays into the material,<br />
while ellipsometry is an optical technique used to measure<br />
thin film thickness.<br />
Table 1. XPS atomic concentrations obtained for<br />
each sample created with different ALD deposition<br />
parameters using the H 2<br />
O 2<br />
precursor.<br />
TD-<br />
MASn<br />
Pulse<br />
H 2<br />
O 2<br />
Pulse<br />
Ti<br />
Atomic<br />
Concentration<br />
(%)<br />
Sn<br />
Atomic<br />
Concentration<br />
(%)<br />
Sn/Ti<br />
Atomic<br />
Ratio<br />
0.5 sec 0.02 sec 10.84 18.89 1.74<br />
0.5 sec 0.1 sec 13.6 16.98 1.25<br />
0.5 sec 1.0 sec 6.1 24.54 4.02<br />
0.1 sec 1.0 sec 9.69 21.01 2.17<br />
0.5 sec 0.02 sec,<br />
40 sec<br />
hold<br />
0.5 sec 0.5 sec,<br />
60 sec<br />
hold<br />
15.75 15.28 0.97<br />
15.66 15.13 0.97<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 41
Table 1 shows the atomic concentrations and Sn/<br />
Ti ratio obtained through the XPS analysis. The atomic<br />
concentrations for Sn using the H 2<br />
O 2<br />
precursor were<br />
much higher than the typical values encountered during<br />
ALD deposition, indicating that the precursor is extremely<br />
reactive and causing highly uncontrolled growth onto the<br />
films. In this case, increasing the TDMASn pulse increased<br />
growth. Increasing the H 2<br />
O 2<br />
precursor pulse from 0.02<br />
seconds to 0.1 seconds did not have much effect on growth,<br />
but increasing its pulse time from 0.1 seconds to 1.0 seconds<br />
increased SnO 2<br />
deposition significantly. Based on the data,<br />
it is clear that H 2<br />
O 2<br />
results in heavy growth of the shells,<br />
but the growth is most likely uneven. Another weakness<br />
of H 2<br />
O 2<br />
is its inconsistency as a precursor because of its<br />
tendency to disproportionate in the precursor cylinder to<br />
water and O 2<br />
. H 2<br />
O 2<br />
is overall not an optimal precursor to<br />
use in SnO 2<br />
deposition on TiO 2<br />
for the purposes of this<br />
study, but holds promise for future research.<br />
The deposition was further optimized by utilizing<br />
dyesol TiO 2<br />
paste to doctor blade the slides instead of<br />
mixed-phase TiO 2<br />
, because of the characteristics of dyesol<br />
TiO 2<br />
as a pure phase substance. This left a slight amount<br />
of SnO on the films immediately after deposition, but the<br />
SnO was completely removed after heating slightly at<br />
200°C. Unlike the mixed phase TiO 2<br />
/SnO 2<br />
structures, the<br />
dyesol TiO 2<br />
/SnO 2<br />
did not require annealing to convert the<br />
SnO to SnO 2<br />
, which would have been impractical for realworld<br />
purposes.<br />
The dyesol TiO 2<br />
/SnO 2<br />
was then characterized using<br />
XPS to confirm that the correct form of the material was<br />
deposited. The correct peak for Sn 4+ was observed at 486.3<br />
eV (Fig. 7), which was extremely close to the recorded<br />
value of 486.6 eV (Stranick & Moskwa, 1993).<br />
confirming that the correct form of Sn 4+ was formed and<br />
not Sn 2+ .<br />
Table 2. Atomic concentrations of Ti, Sn, O obtained<br />
through XPS of dyesol TiO 2<br />
/SnO 2<br />
samples.<br />
ALD<br />
Cycles<br />
Ti<br />
Atomic<br />
Conc.(%)<br />
Sn<br />
Atomic<br />
Conc.(%)<br />
O Atomic<br />
Conc.<br />
(%)<br />
(Ti% +<br />
Sn%) /<br />
O%<br />
30 8.17 24.27 61.96 0.52<br />
40 3.48 30.78 59.63 0.57<br />
50 3.91 30.47 60.18 0.57<br />
Samples created using varied parameters were analyzed<br />
again using XPS and ellipsometry to determine the effect<br />
of changing each condition on deposition of SnO 2<br />
using<br />
dyesol TiO 2<br />
. Figure 8 shows the effect of changing each<br />
ALD parameter other than temperature on both the<br />
thickness of SnO 2<br />
deposited on planar silicon obtained<br />
through ellipsometry as well as the ratio of Sn to Ti atomic<br />
concentrations from TiO 2<br />
/SnO 2<br />
samples determined by<br />
XPS. A lower growth rate is desired for this deposition<br />
because the SnO 2<br />
shell naturally is thicker than the TiO 2<br />
shell, but they should be similar thicknesses in order to<br />
compare their electron transfer kinetics. The optimal hold<br />
time is around 60 seconds for decreasing SnO 2<br />
thickness.<br />
The decreased growth caused by both increased hold<br />
and purge time is likely due to removal of moisture and<br />
impurities introduced into the chamber during the pulse<br />
and hold times. The lowest growth rate occurred on the<br />
sample with 0.1 second TDMASn pulse, 0.02 second H 2<br />
O<br />
pulse, 20 second hold time, 60 second purge time. This<br />
recipe resulted in a growth rate of 0.07 nm per cycle, which<br />
decreased from the 0.09 nm per cycle growth rate achieved<br />
with the standard recipe used for TiO 2<br />
deposition.<br />
Figure 7. XPS spectra of Sn 3d region, displaying peak<br />
at 486.3 eV.<br />
In addition, the atomic concentrations were collected<br />
of Ti, Sn, and O (Table 2). If the deposited material was<br />
all SnO 2<br />
, the ratio of (Ti % + Sn %):O% should be 1:2. The<br />
ratios calculated for the samples were all very close to 0.5,<br />
Figure 8. SnO 2<br />
shell thickness determined by<br />
ellipsometry (left axis) and atomic ratio of Sn to Ti<br />
determined by XPS (right axis) with varying ALD<br />
conditions, at 15 cycles.<br />
42 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> CHEMISTRY
4. Conclusion<br />
Atomic layer deposition was conducted to create<br />
novel TiO 2<br />
/TiO 2<br />
and TiO 2<br />
/SnO 2<br />
core/shell structures.<br />
Dye loading studies conducted on TiO 2<br />
/TiO 2<br />
with the<br />
RuP chromophore revealed that dye loading in TiO 2<br />
/<br />
TiO 2<br />
increases to a certain point, and then decreases,<br />
contradicting the trends of SnO 2<br />
/TiO 2<br />
which show<br />
continuously decreasing dye loading which was attributed<br />
to decreasing pore size. This inconsistency suggests the<br />
importance of multiple other factors such as processing<br />
conditions and the type of TiO 2<br />
used to synthesize the<br />
core. Moreover, initial attempts to create TiO 2<br />
/SnO 2<br />
resulted in the formation of SnO, but this was removed<br />
by using dyesol TiO 2<br />
to create the thin films rather than<br />
mixed phase TiO 2<br />
. The effects of each ALD parameter<br />
were studied to create films of similar thicknesses for both<br />
TiO 2<br />
/SnO 2<br />
and TiO 2<br />
/TiO 2<br />
, and the growth rate of the<br />
SnO 2<br />
was able to be decreased from the standard recipe.<br />
Future directions will include the conduction of transient<br />
absorption spectroscopy in order to understand the<br />
differences in the dynamics of interfacial electron kinetics<br />
between the TiO 2<br />
/TiO 2<br />
and TiO 2<br />
/SnO 2<br />
structures. In<br />
addition, the electron kinetics should be studied in core/<br />
shells of various other oxide materials.<br />
5. Acknowledgments<br />
We would like to thank Dr. Jillian Dempsey as well as<br />
Michael Mortelliti for their incredible mentorship over<br />
this project. This work was performed in part at the<br />
Chapel Hill Analytical and Nanofabrication Laboratory,<br />
CHANL, a member of the North Carolina Research<br />
Triangle Nanotechnology Network, RTNN, which is<br />
supported by the National Science Foundation, Grant<br />
ECCS-1542015, as part of the National Nanotechnology<br />
Coordinated Infrastructure, NNCI. In addition, the<br />
project was funded by a grant from the RTNN Kickstarter<br />
Program for fabrication & analytical costs.<br />
6. References<br />
Fujishima, A., & Honda, K. (1972). Electrochemical<br />
Photolysis of Water at a Semiconductor Electrode. Nature,<br />
238(5358), 37-38. https://doi.org/10.1038/238037a0<br />
George, S. M. (2010). Atomic Layer Deposition: An<br />
Overview. Chemical Reviews, 110(1), 111–131. https://<br />
doi.org/10.1021/cr900056b<br />
Gish, M. K., Lapides, A. M., Brennaman, M. K., Templeton,<br />
J. L., Meyer, T. J., & Papanikolas, J. M. (2016). Ultrafast<br />
Recombination Dynamics in Dye-Sensitized SnO 2<br />
/TiO 2<br />
Core/Shell Films. The Journal of Physical Chemistry<br />
Letters, 7(24), 5297–5301. https://doi.org/10.1021/acs.<br />
jpclett.6b02388<br />
Jafari, T., Moharreri, E., Amin, A. S., Miao, R., Song, W.,<br />
& Suib, S. L. (2016). Photocatalytic water splitting - The<br />
untamed dream: A review of recent advances. Molecules,<br />
21(7). https://doi.org/10.3390/molecules21070900<br />
James, E. M., Barr, T. J., & Meyer, G. J. (<strong>2018</strong>). Evidence<br />
for an Electronic State at the Interface between the SnO 2<br />
Core and the TiO 2<br />
Shell in Mesoporous SnO 2<br />
/TiO 2<br />
Thin<br />
Films. ACS Applied Energy Materials, acsaem.7b00274.<br />
https://doi.org/10.1021/acsaem.7b00274<br />
Knauf, R. R., Kalanyan, B., Parsons, G. N., & Dempsey, J.<br />
L. (2015). Charge Recombination Dynamics in Sensitized<br />
SnO 2<br />
/TiO 2<br />
Core/Shell Photoanodes. Journal of Physical<br />
Chemistry C, 119(51), 28353–28360. https://doi.<br />
org/10.1021/acs.jpcc.5b10574<br />
Stranick, M. A., & Moskwa, A. (1993). SnO 2<br />
by XPS.<br />
Surface Science Spectra, 2(1), 50–54. https://doi.<br />
org/10.1116/1.1247724<br />
Wang, D., et al. (2017). Layer-by-Layer Molecular<br />
Assemblies for Dye-Sensitized Photoelectrosynthesis<br />
Cells Prepared by Atomic Layer Deposition. Journal of<br />
the American Chemical Society, 139(41), 14518–14525.<br />
https://doi.org/10.1021/jacs.7b07216<br />
Ashford, D. L., Gish, M. K., Vannucci, A. K., Brennaman,<br />
M. K., Templeton, J. L., Papanikolas, J. M., & Meyer, T.<br />
J. (2015). Molecular Chromophore-Catalyst Assemblies<br />
for Solar Fuel Applications. Chemical Reviews,<br />
115(23), 13006–13049. https://doi.org/10.1021/acs.<br />
chemrev.5b00229<br />
Brennaman, M. K., et al. (2016). Finding the Way to<br />
Solar Fuels with Dye-Sensitized Photoelectrosynthesis<br />
Cells. Journal of the American Chemical Society, 138(40),<br />
13085–13102. https://doi.org/10.1021/jacs.6b06466<br />
CHEMISTRY<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 43
USING A HYBRID MACHINE LEARNING APPROACH<br />
FOR TEST COST OPTIMIZATION IN SCAN CHAIN<br />
TESTING<br />
Luke Duan<br />
Abstract<br />
Continual technological advances have led to more complex microchip designs, which in turn, have led to the need for<br />
more complex fault testing. As a result, higher testing costs (increased test time and data volume) have emerged as well.<br />
This work examines one application of hybrid machine learning (ML) to optimize the costs of scan chain testing. We<br />
used fifty-one benchmark circuits to train the models and analyze their performances. We generated training data by<br />
performing scan chain test simulations on each of these circuits using MentorGraphics tools DFTAdvisor and FastScan<br />
and compiled them into files readable by the ML framework Weka. We then trained three individual ML models and<br />
evaluated their accuracies by comparing them against a test set. Finally, we created a hybrid model by combining these<br />
individual models, with different weights allotted to each model based on their individual accuracy. Findings showed that<br />
there was a slight increase in performance by using a hybrid approach. We concluded that this method can be improved<br />
by using larger training sets and better heuristic algorithms when assigning weights. This research could be useful for the<br />
microchip industry by reducing time-to-market.<br />
1. Introduction<br />
Technological advances in the field of engineering have<br />
allowed integrated circuit/microchip design companies<br />
to figure out how to continuously add more and more<br />
transistors (along with gates) onto smaller and smaller<br />
devices. In order to completely test for all the possible faults<br />
in a microchip, more complex and costly testing is needed<br />
on these denser designs (Bushnell & Agrawal, 2005).<br />
One procedure for fault testing occurs during the design<br />
phase of chips - in the form of scan chain testing. In this<br />
type of testing, a certain number of scan chains are chosen<br />
for insertion into a circuit, with varying numbers of scan<br />
chains having different test costs. It can become extremely<br />
tedious to test all possible scan chain numbers, and<br />
manually pick out the most cost-efficient number to use.<br />
In order to make that decision, machine learning models<br />
can be trained with circuit data, along with the number<br />
of scan chains inserted. Then, when provided with a new<br />
circuit, they would be able to predict the best number of<br />
scan chains to use. (Zipeng & Chakrabarty, 2016) proposed<br />
a method to optimize test cost by choosing parameters,<br />
such as scan chain length, using a support vector regression<br />
(SVR) machine learning model. In this work, we will<br />
examine the parameter optimization of the number of<br />
scan chains. The primary focus is to explore how well a<br />
hybrid machine learning model performs in predicting the<br />
optimal number of scan chains to use in scan chain testing.<br />
1.1 – Design for Testability (DFT)<br />
Design for Testability, or DFT, can be described as the<br />
set of methods that make testing for faults in microchips<br />
easier. In the next section, we break down DFT and explain<br />
the connections between digital logic, data flip-flops, shift<br />
registers, and scan chain testing.<br />
1.1.1 – Context<br />
There exist two types of digital logic: combinational and<br />
sequential, with the latter involving a memory component<br />
as well as a clock signal for regulation. The physical<br />
manifestation of digital logic can be found in digital circuits.<br />
A flip-flop (FF) is a prime example of a component in a<br />
sequential digital circuit. It is not uncommon for instances<br />
of sequential logic/circuits to incorporate combinational<br />
logic.<br />
The Data FF (Fig. 1), or DFF, is the simplest type of<br />
FF, and consists of an input (D), a clock signal (CLK) and<br />
an output (Q). The “scan-enabled” DFF comes with an<br />
additional scan-in and scan-out port (scan-out port not<br />
pictured).<br />
Figure 1. A typical scan-enabled flip-flop (Gupta, 2014)<br />
It is a basic storage element in sequential logic, able to<br />
hold a stable state of either 0 or 1. The DFF may receive<br />
an input, but unless the clock signal is turned “on,” the<br />
output will not change. This reduces the occurrence of<br />
any unnecessary output changes, thus saving power. A<br />
44 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
shift register is essentially a linear chain of these DFF’s,<br />
all connected to and regulated by the same clock signal.<br />
The output of one DFF is directly connected to the input<br />
of the next. The input can be controlled, and the output of<br />
the register can be observed. For our purposes, we do not<br />
worry about what happens within the register.<br />
1.1.2 – Scan Chain Testing<br />
Scan chain testing (Fig. 2) is a common method for testing<br />
for faults in silicon when manufacturing circuits. Either<br />
one or multiple scan enabled shift registers are formed,<br />
with each DFF being replaced with their “scan-enabled”<br />
versions, which simply means they come equipped with<br />
scan-in and scan-out ports. The total number of flip-flops<br />
are divided as equally as possible over the number of scan<br />
chains in a circuit. A clock signal is established, and testing<br />
begins. An input test pattern generated by pseudo-random<br />
methods is scanned in by each register, and the scannedout<br />
output will be compared to the expected output.<br />
The expected output is the output that would have been<br />
reached if all gates in the combinatorial logic had been<br />
working correctly. If the two outputs do not match, then<br />
a fault is detected. Scan chain testing can be characterized<br />
by its test application time (time for the test to occur),<br />
and test data volume (number of test patterns inserted to<br />
test for all faults) (Gupta, 2014). These costs can change<br />
depending on the number of scan chains used.<br />
connected to every single neuron in the next layer. The<br />
input values, each multiplied by a unique weight, are<br />
summed up and passed through an activation function.<br />
If above a certain value, the neuron “fires” (information<br />
is passed on to the next layer). A neural network uses<br />
feedback (comparison to actual value) to learn and slowly<br />
correct itself to become the best predictor it can be<br />
(Mitchell, 1997).<br />
Figure 3. A visual representation of an artificial<br />
neural network; two hidden layers.<br />
Random forests (Fig. 4) essentially take a collection<br />
of decision trees, and output either the mode or mean<br />
predictions of the individual trees. Decision trees work by<br />
breaking a dataset into smaller pieces and formulating a set<br />
of rules for decision-making based on previous data. They<br />
have the ability to decide which features are important and<br />
which features can be dropped (as they contribute little to<br />
the prediction process) (Donges, <strong>2018</strong>).<br />
Figure 2. A typical scan chain (Gupta, 2014)<br />
1.2 – Machine Learning (ML)<br />
1.2.1 – Basic Principles<br />
Machine learning (ML) is a subset of artificial<br />
intelligence, which is built around the idea of self-learning<br />
and self-improvement. To begin, a ML model is trained<br />
with a set of training data. In supervised learning, both the<br />
input and expected output are fed into the model. After<br />
becoming sufficiently trained, the model can be tested<br />
against a test set. Accuracies for the model can be found<br />
by comparing the predicted outputs from the model to the<br />
actual outputs of the test set (Mitchell, 1997).<br />
1.2.2 – Machine Learning Model Descriptions<br />
The artificial neural network (NN) (Fig. 3) consists of<br />
an input layer, one or several hidden layers, and an output<br />
layer. Each layer consists of several neurons, which are<br />
Figure 4. A visual representation of a random forest;<br />
two separate decision trees - red nodes represent<br />
the individual output of each tree, which are then<br />
combined in some way to form the output of the<br />
random forest (Donges, <strong>2018</strong>).<br />
Support vector regression (SVR) (Fig. 5) works by<br />
optimizing a line between two sets or classes of data. In<br />
other words, while learning, it attempts to minimize<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 45
error by adjusting a hyperplane. The accuracy is generally<br />
dependent on setting good parameters (Cortes and<br />
Vapnik, 1995).<br />
A formula for a Test Cost (TC) may be obtained from Test<br />
Application Time (TT) and Test Data Volume (TV):<br />
TT max<br />
and TV max<br />
represented the maximum test<br />
application time and maximum test data volume for a<br />
circuit, respectively. This was to normalize a value for TC<br />
(Zipeng and Chakrabarty, 2016).<br />
Figure 5. A visual representation of SVR applied to<br />
two classes of data (black circles and blue squares);<br />
hyperplane represented by the green line.<br />
1.2.3 – Weka Software<br />
Weka is a software tool that provides a collection of<br />
many developed ML models, including neural networks,<br />
random forests, and support vector regression. This<br />
application contains a user interface, which simplifies the<br />
experience when working with and applying ML to data<br />
(Weka Machine Learning Group, n.d.).<br />
2. Methodology<br />
This project was divided into four phases: Training Data<br />
Generation, Individual ML Model Training, Allotment of<br />
Weights, and Hybrid ML Model Performance.<br />
2.1 – Training Data Generation<br />
In this phase, the tools DFTAdvisor (MentorGraphics,<br />
n.d.) (to insert scan chains) and FastScan (MentorGraphics,<br />
n.d) (to generate and compare test patterns) were applied<br />
on a collection of 51 pre-constructed benchmark digital<br />
circuits from the ISCAS89 library. With each circuit, we<br />
recorded several features: the number of primary inputs,<br />
the number of primary outputs, the number of gates,<br />
the number of flip-flops, and the number of scan chains<br />
inserted. Five variations of each circuit were tested, from<br />
one scan chain inserted to five scan chains inserted.<br />
For context, the features had the following ranges<br />
(Table 1):<br />
Table 1. Range of values for features.<br />
# of Primary Inputs 6 - 80<br />
# of Primary Outputs 1 - 320<br />
# of Gates 26 - 26115<br />
# of Flip-Flops 3 - 1728<br />
# of Scan Chains Inserted 1 - 5 for each circuit<br />
We also took note of the test application time and test<br />
data volume in performing each scan chain test.<br />
2.2 – Individual ML Model Training<br />
In this phase, Weka was used to individually train three<br />
types of regression ML models: artificial neural networks,<br />
random forest, and SVR. Out of the 51 total circuits that<br />
were given, 42 circuits were used for training the models,<br />
while the remaining 9 circuits were used for testing. A true<br />
random number generator was used to select the circuits<br />
in each set. The ML model was trained and run against the<br />
testing set. The outputted TC was compared to a manually<br />
calculated TC from the actual FastScan data.<br />
2.3 – Allotment of Weights<br />
In this phase, the weights that each individual ML<br />
model will have in the hybrid model were empirically<br />
selected. This was performed on the following basis: the<br />
higher the accuracy, the more weight it had. There were<br />
many different methods for weight selection, which left<br />
this phase open to a lot of trial and error.<br />
2.4 – Hybrid ML Model Performance<br />
In this phase, the hybrid model was fed a different<br />
set of training data and tested against a different testing<br />
set (though still chosen out of the same collection of<br />
benchmark circuits). The minimum TC was chosen,<br />
and the scan chain number correlated with that TC was<br />
compared to the actual FastScan output. The accuracy of<br />
the hybrid model was evaluated.<br />
3. Data Analysis<br />
We performed tests on the 9 circuits not used for<br />
training.<br />
3.1 – Weighting (Individual Models)<br />
For each individual model, results were labeled with the<br />
following:<br />
• Off if the scan chain number correlating with the<br />
lowest ML test cost prediction didn’t match the scan<br />
chain number correlating with the lowest actual<br />
FastScan output (Table 2)<br />
• Success if the scan chain number correlating with<br />
the lowest ML test cost prediction did match the<br />
scan chain number correlating with the lowest actual<br />
FastScan output (Table 3)<br />
46 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
Example of Off for a test circuit:<br />
Table 2. Comparison between actual and predicted<br />
test cost example 1 - from NN.<br />
The scan chain number correlating with the lowest<br />
ML test cost prediction (0.9474) is 1, while the scan chain<br />
number correlating with the lowest actual FastScan output<br />
(0.8733) is 4.<br />
Example of Success for a test circuit:<br />
random forest had a weight of 0.8000, and the support<br />
vector regression had a weight 0.1000 (a 1:8:1 ratio).<br />
3.2 – Hybrid Model<br />
Both hybrid models had 3 Successes, so further<br />
evaluation had to be completed. Specifically, the total<br />
difference between the actual test cost corresponding<br />
with the predicted SC number and the actual lowest test<br />
cost was found for the 9 testing circuits. A lower total<br />
difference is indicative of a more accurate model.<br />
The differences with the first weighting combination<br />
(2:3:1) are shown in Table 5.<br />
Table 5. Hybrid model (Weights 2:3:1) total differences.<br />
Table 3. Comparison between actual and predicted<br />
test cost example 2 - from NN.<br />
The scan chain number correlating with the lowest ML<br />
test cost prediction (0.9951) is 5, matching the scan chain<br />
number correlating with the lowest actual FastScan output<br />
(0.9504).<br />
The lowest values for Test Cost are highlighted in<br />
boldface. If the lowest Predicted Test Cost does not<br />
match the lowest Actual Test Cost, then Off. If the lowest<br />
Predicted TC matched the lowest Actual TC, then Success.<br />
We only focus the lowest values of cost, because this is the<br />
main objective of our optimization.<br />
Weights for the hybrid model were assigned based on<br />
the number of Successes (Table 4).<br />
Note: A difference of 0.0000 means Success.<br />
The total difference for the hybrid model with weights<br />
2:3:1 is 0.3502.<br />
The differences with the second weighting combination<br />
(1:8:1) are shown in Table 6.<br />
Table 6. Hybrid model (Weights 1:8:1) total differences.<br />
Table 4. Number of Successes for each model.<br />
The total difference for the hybrid model with weights<br />
1:8:1 is 0.3560.<br />
Thus, our initial weighting of the hybrid model was in<br />
a 2:3:1 ratio. The artificial neural network had a weight of<br />
0.3333, the random forest had a weight of 0.5000, and the<br />
support vector regression had a weight 0.1667 in the hybrid<br />
model. We also decided to investigate heavily weighting<br />
the best-performing individual model as compared to the<br />
other two models. In this weighting of the hybrid model,<br />
the artificial neural network had a weight of 0.1000, the<br />
We next compare these differences to those of the<br />
individual models. The Artificial Neural Network isn’t<br />
considered in these comparisons, due to predicting invalid<br />
test costs.<br />
The differences for the random forest model are shown<br />
in Table 7.<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 47
Table 7. Random forest model total differences.<br />
optimized weights. There could very well exist a weight<br />
set for the hybrid model that provides an even better<br />
performance. Moreover, we hope that with promising<br />
results, this methodology may be applied to industriallevel<br />
circuits for real-world use.<br />
5. Acknowledgments<br />
The total difference for the random forest model is 0.3560.<br />
The differences for the support vector regression model<br />
are shown in Table 8.<br />
Table 8. Support vector regression model total<br />
differences.<br />
I would like to express my sincerest thanks towards<br />
Dr. Jonathan Bennett for his constant encouragement<br />
and accepting me into the Research in Physics program<br />
at NCSSM. I would also like to acknowledge Dr. Sarah<br />
Shoemaker for organizing and directing the Summer<br />
Research Internship Program.<br />
I am very thankful to Dr. Krishnendu Chakrabarty for<br />
granting me permission to work with his research group<br />
at Duke University.<br />
I would like to thank Zhanwei Zhong, Shi Jin, Thomas<br />
Napoles, and the Duke Office of Information Technology<br />
for their assistance with local issues.<br />
Last but not least, I would like to express my gratitude<br />
towards my mentor, Arjun Chaudhuri, for his patience<br />
and dedication in guiding and challenging me.<br />
6. References<br />
The total difference for the support vector regression<br />
model is 0.4894.<br />
The hybrid model with weights 2:3:1 had lower total<br />
differences compared to the total differences of the<br />
individual models, as well as a hybrid model with<br />
nonoptimal weighting, showing evidence of a slightly<br />
better performance. This provides basic evidence that<br />
there is, in fact, an improvement in accuracy by using a<br />
hybrid ML method.<br />
4. Conclusion and Future Work<br />
4.1 – Conclusion<br />
This work offered the possibility of using a hybrid<br />
machine learning model to predict the best number of scan<br />
chains to use for cost optimization. Though individual<br />
ML models, such as the artificial neural network, random<br />
forest, and support vector regression work well on their<br />
own, a hybrid model with correct weighting appears to<br />
offer a slightly better performance. With this in mind,<br />
microchip testers could potentially use this new method<br />
to further decrease test costs and improve time-to-market.<br />
4.2 – Future Work<br />
Running a program or algorithm may offer further<br />
Bushnell, M., Agrawal, V. (2005). Essentials of Electronic<br />
Testing, Springer.<br />
Zipeng, L., Chakrabarty, K. (2016). Test Cost Optimization<br />
in a Scan-Compression Architecture using Support-<br />
Vector Regression. Proc. IEEE Test Symposium (VTS).<br />
Gupta, N. (2014). Overview and Dynamics of Scan<br />
Chain Testing, Retrieved from https://anysilicon.com/<br />
overview-and-dynamics-of-scan-testing/<br />
Mitchell, T.M. (1997). Machine Learning, McGraw-Hill.<br />
Donges, S. (<strong>2018</strong>). The Random Forest Algorithm.<br />
Retrieved from https://towardsdatascience.com/therandom-forest-algorithm-d457d499ffcd<br />
Cortes C., Vapnik V. (1995). In Support-vector networks,<br />
Machine Learning (vol. 20, pp. 273-297).<br />
Machine Learning Group. (n.d). Weka 3: Data Mining<br />
Software in Java, University of Waikato.<br />
DFTAdvisor Reference Manual. (n.d). MentorGraphics.<br />
FastScan and FlexTest Reference Manual. (n.d).<br />
MentorGraphics.<br />
48 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
NOVEL WATER DESALINATION FILTER UTILIZING<br />
GRANULAR ACTIVATED CARBON<br />
Geoffrey Fylak<br />
Abstract<br />
As the human population continues increasing, so does the demand for freshwater resources. The scarcity of freshwater<br />
will likely impact one-third of the world’s population within the next decade. While there are many proven methods of<br />
water desalination, most are cost- and energy-intensive. Our research seeks to improve upon capacitive deionization:<br />
an emerging, yet proven, scalable method of desalination that removes charged species from water using low levels of<br />
electricity. The filter utilizes granular activated carbon (GAC), an affordable, naturally abundant material commonly used<br />
in industrial Brita® water filters to remove uncharged contaminants. We anticipate that GAC’s electrically conductive<br />
properties will enable the material to adsorb sodium chloride. Our goal is to determine and enhance the performance<br />
capabilities of GAC by altering operational parameters and system design. Initial tests demonstrated low performance<br />
due to inadequate operational parameters and design flaws. Through systematic improvements, researchers have greatly<br />
increased system performance. The filter’s charge efficiency has increased from 13% to 63% while the adsorption capacity<br />
has increased from 10.3 µg/g to 452.0 µg/g. Based upon success in removing sodium chloride, our filter’s application could<br />
be extended to remove more harmful, charged water contaminants in the future.<br />
1. Introduction<br />
1.1 – Significance<br />
As the human population continues increasing, so<br />
does the demand for freshwater resources. The scarcity<br />
of freshwater will likely impact one-third of the world’s<br />
population within the next decade. While there are many<br />
proven methods of water desalination, most are cost<br />
and energy-intensive. Our research seeks to improve<br />
upon a novel desalination technique, which would<br />
expand available drinking water sources on a global<br />
scale. The technology investigated is based on capacitive<br />
deionization (CDI), an emerging, yet proven, scalable<br />
method of desalination that removes charged species from<br />
water using low levels of electricity. The filter will utilize<br />
granular activated carbon (GAC), an affordable, naturally<br />
abundant material commonly used in industrial Brita®<br />
water filters to remove uncharged contaminants. We<br />
anticipate that GAC’s electrically conductive properties<br />
will enable the material to adsorb sodium chloride.<br />
Our goal is to determine and enhance the performance<br />
capabilities of GAC by altering operational parameters<br />
and system design. Emerging contaminants widely exist<br />
in raw and treated drinking water and present an ongoing<br />
threat to human health and the planet. Certain substances,<br />
such as PFAS, are suspected carcinogens and pose a risk to<br />
humans even at trace levels (ng/L to µg/L). Thus, there<br />
exists a need to develop viable methods and technologies<br />
to remove charged contaminants from water resources.<br />
Ultimately, our filter’s application can be extended to<br />
remove more harmful charged contaminants in the future.<br />
1.2 – Background Literature Review<br />
Water treatment is a broad field consisting of many<br />
different methods and focuses. Water desalination is a<br />
sub-field which focuses on removing salt from water.<br />
Many industrial scale water desalination techniques<br />
exist, such as reverse osmosis and thermal distillation;<br />
however, these techniques are highly energy intensive.<br />
CDI technology improves upon these other techniques<br />
through its low energy requirement.<br />
CDI cells operate based off of the electrochemical<br />
principles of charge. Essentially, saltwater is a solution<br />
containing two sets of molecules: salt compounds and<br />
water molecules. Salt compounds are composed of two<br />
types of ions: positively charged sodium ions and negatively<br />
charged chloride ions. Moreover, when opposite electrical<br />
charges are given to two parallel plates, an electric field is<br />
created. This electric field will immobilize sodium chloride<br />
ions and separate them based off of their respective<br />
electrical charge, directing the positively charged ions to<br />
attach to the negatively charged plate and vice versa for<br />
the negatively charged ions. However, the most crucial<br />
component of a CDI system is the electrode, the part that<br />
captures the charged salt ions, thus removing them from<br />
the water, resulting in pure water (Suss et al., 2015).<br />
Previous research has proven CDI technology to<br />
successfully remove salt on the lab scale (Porada et al.,<br />
2013) and industrial scale (Welgemoed & Schutte, 2005).<br />
These experiments describe the salt removal process, as<br />
well as detail the various essentials of a successful CDI<br />
system. The most important physical component is the<br />
electrode material, as the resistivity and specific surface<br />
area of the material determine the amount of salt that can<br />
be adsorbed. Materials with high specific surface areas and<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 49
porosity are most efficient at removing salt.<br />
As researchers attempt to expand the applicability of<br />
CDI technology, they are experimenting with a variety<br />
of electrode materials. One particular electrode material,<br />
granular activated carbon, is contained within Brita®<br />
water filters, removing uncharged contaminants with its<br />
desirable properties. Researchers determined granular<br />
activated carbon (GAC) to have a promising surface<br />
conductivity and adsorption capacity (Jia & Zhang, 2016).<br />
Another set of researchers packed an electrode chamber<br />
with granular activated carbon and discovered up to two<br />
and a half times more salt removal (Bian et al., 2015).<br />
However, their research did not assess the potential of<br />
GAC as a primary electrode material. Our study seeks<br />
to determine performance metrics, as well as compare<br />
our findings with pre-existing data. In doing so, we will<br />
be able to gain a holistic view of the efficiency of GAC<br />
as an electrode material. Since many industrial water<br />
filters, such as Brita®’s, utilize GAC, the transition to an<br />
industrial-scale desalination system will be feasible if GAC<br />
is proven to be efficient.<br />
However, to accurately assess the efficiency of GAC<br />
as an electrode material, we must first ensure that the<br />
CDI system’s design is sufficient. Charge efficiency<br />
is an important, quantifiable indication of a system’s<br />
effectiveness. A system’s charge efficiency is a measurement<br />
in the form of a percentage, which demonstrates the moles<br />
of salt removed per moles of electrical charge emitted<br />
to electrodes. A system with a charge efficiency of 100%<br />
removes one mole of salt per mole of electrical charge.<br />
One set of researchers discovered that CDI cells must be<br />
charged at a positive voltage to achieve the highest charge<br />
efficiency (Avraham et al., 2009). Therefore, our project<br />
will utilize critical findings to ensure that the electrode<br />
parameters are under enable the maximum performance<br />
of GAC.<br />
Though there is substantial research surrounding<br />
the CDI process, there is no significant information<br />
concerning the efficiency of GAC as an electrode material.<br />
By conducting this research, GAC could potentially prove<br />
to be a useful electrode material, consequently sparking<br />
feasible industrial filter production. Conversely, GAC<br />
could prove to be inefficient, allowing researchers to<br />
focus on other potential modifications. The purpose of<br />
this study is to determine the efficiency of the electrode<br />
material granular activated carbon in comparison with<br />
pre-existing materials.<br />
2. Materials<br />
2.1 – Novel CDI System Design<br />
The novel filter was designed, modeled, and assembled<br />
using materials funded by the Call Lab at NC State<br />
University. As a novel design, each material and component<br />
must be considered to achieve optimal functionality.<br />
The assembled and disassembled GAC filter design is<br />
illustrated below (Fig. 1, Fig. 2). Certain materials such<br />
as the hex nuts, screws, and barbed tube fittings did not<br />
require modification; however, the polycarbonate plate,<br />
graphite plates, rubber gaskets, and glass fibre prefilters<br />
needed to be cut. Each part plays an instrumental role<br />
in adapting GAC to carry electrical charge and remove<br />
sodium chloride.<br />
Figure 1. A rendered model of the assembled filter.<br />
Figure 2. A disassembled model of the GAC filter.<br />
Numbers coincide with different parts and materials:<br />
1. Rubber gaskets and glass fibre prefilter; 2. Nylon<br />
screws; 3. Barbed Tube Fittings; 4. Polycarbonate<br />
plates; 5. Graphite plates 1/8” thick; 6. Graphite plates<br />
1” thick.<br />
Water will enter through the top barbed tube fitting<br />
and exit through the bottom, passing through the<br />
cylindrical chambers that contain the electrode material.<br />
An electrical charge must be given to the system through<br />
an anode and a cathode. Hence, the 1/8” thick graphite<br />
plates have an extended area designated for anode and<br />
cathode attachment. Graphite was chosen as the material<br />
to house the granular activated carbon (GAC) because it<br />
is electrically conductive. However, since two oppositely<br />
50 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
charged chambers are created, they must be separated<br />
to ensure the system does not short-circuit. A series of<br />
glass fibre prefilters (spacers) accomplishes this goal. The<br />
middle spacer separates the anode and cathode chambers,<br />
ensuring the GACs in either chamber do not touch and<br />
cause system failure.<br />
Gaskets are used in combination with the spacers to<br />
prevent leakage from occurring. Each of these components<br />
is held together by two nylon screws. It is essential to use<br />
nylon, plastic, or any other non-conductive material so<br />
that the system does not short-circuit when an object is in<br />
contact with both the anode and cathode chambers at the<br />
same time. The nylon hex nuts allow researchers to tighten<br />
the system, preventing leakages and pressure build-ups.<br />
With computer-aided design, the 3D model was<br />
converted into 2D sketches and each individual part<br />
was able to be cut in NC State’s Machine Shop. Lastly, a<br />
3D-printer in the NCSSM Fabrication Lab was used to<br />
create a stand to hold the filter upright and prevent the<br />
filter from lying horizontally (Fig. 3).<br />
Figure 4. A top view of the resistance experienced<br />
from the anode/cathode connection sites to various<br />
locations within the electrode chamber.<br />
Figure 5 shows a few photos of the GAC filter<br />
completely assembled.<br />
Figure 5. The GAC filter completely assembled, from<br />
a variety of angles.<br />
3. Specific Aims and Research Design<br />
We seek to address the following research questions:<br />
Figure 3. A red, 3D-printed stand supports the GAC<br />
filter and enhances the system’s vertical flow path.<br />
Aside from flow path, system resistance was a challenge<br />
that the design needed to overcome. Thus, researchers<br />
filled the chambers with GAC and measured the resistance<br />
from the anode or cathode connection points to various<br />
locations within the chamber (Fig. 4). These data<br />
demonstrate that graphite sufficiently emits charge to all<br />
of the electrode material. Although resistance increases in<br />
areas furthest away from the graphite, electrical charge can<br />
still travel to those areas and facilitate salt removal (Fig. 4).<br />
3.1 – Specific Aim 1<br />
Determine the relationship between flow rate and CDI<br />
system performance by running tests with different flow<br />
rates and comparing the respective performances.<br />
3.2 – Rationale and Hypothesis<br />
The flow rate of water through a CDI system impacts<br />
the volume of salt entering the system. Exposure to higher<br />
salt concentrations should enable electrodes to capture<br />
more salt. However, increased flow rates facilitate pressure<br />
build-ups and leakage issues that may negatively impact<br />
system performance. By analyzing the impact of flow<br />
rate on system efficiency, researchers can discover the<br />
operational parameters necessary to yield maximum salt<br />
removal.<br />
Typical lab-scale, flow-by CDI cells utilize 0.200 g of<br />
electrode material; however, this novel design incorporates<br />
20.0409 g of electrode materials. Due to the much higher<br />
system volume, researchers expect higher flow rates to<br />
increase CDI cell performance. Moreover, incremental<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 51
changes in flow rate will likely impact performance less<br />
because of the large volume. Thus, researchers may need<br />
to greatly increase flow rate to produce significant changes<br />
in performance.<br />
3.3 – Supporting Preliminary Data<br />
We previously analyzed the relationship between flow<br />
rate and cell performance using flow-by CDI cells. We<br />
concluded that increasing flow rate negatively impacted<br />
CDI performance across all performance metrics (Table<br />
1). Our current experiment utilizes flow-through CDI<br />
cells; thus, our design differs from the one featured in this<br />
study.<br />
Nevertheless, it is important to observe the implications<br />
of these findings on the electrochemical level, as this study<br />
indicates that higher flow rates induce pressure build-ups<br />
and consequently, Faradaic Reactions. Faradaic Reactions<br />
contribute to pH fluctuations and electronic charge storage<br />
without salt ion adsorption (Na + or Cl - ).<br />
Table 1. Comprehensive visualization of the<br />
impact of increasing flow rate on flow-by CDI<br />
cell performance. Noticeably, each performance<br />
parameter decreases as flow rate increases.<br />
Adsorption<br />
Capacity<br />
Charge<br />
Efficiency<br />
4 mL/min 6 mL/min 8 mL/min<br />
2.507mg/g 1.212mg/g 1.075mg/g<br />
18.87% 9.44 % 10.07 %<br />
3.4. – Methods<br />
After calibrating the pump, tubing was attached from<br />
the pump through the CDI system, then directed into a<br />
properly labeled waste container. Next, distilled water<br />
was pumped through the system to ensure that no leakage<br />
occurred.<br />
Finally, flow cells were attached outside of the system<br />
to allow researchers to measure the conductivity of water<br />
exiting the system (Fig. 6).<br />
Figure 6. A visualization of the research setup,<br />
including the pump, salt solution, CDI cell,<br />
conductivity flow cell, pH flow cell, waste bucket,<br />
and tubing.<br />
With the system assembled, we created one liter of 100<br />
mM salt solution. The 100 mM solution is then diluted<br />
into a 10 mM salt solution and pumped through the CDI<br />
cell. This step saves time creating solutions in the future,<br />
as it is much easier to dilute a solution than create one.<br />
For this specific project, we chose to test flow rates of<br />
5 mL/min and 10 mL/min. Using the calibration which<br />
we previously conducted, we programmed the pump to<br />
each of these flow rates in different tests. All of our other<br />
system parameters were kept constant during this test:<br />
voltage during charge was 1.2 V, charge cycle time was 5<br />
minutes, the alligator clips were positioned from anode to<br />
cathode, and the system ran for three cycles.<br />
We first measure the conductivity and pH of the water<br />
before it enters the system. The flow cells containing<br />
conductivity and pH probes are used to measure the<br />
conductivity and pH of the water exiting the system.<br />
Conductivity is directly related to salt concentration, so the<br />
combination of these measurements enables researchers to<br />
analyze salt removal over time. Each probe captures data<br />
points one minute apart, allowing researchers to observe<br />
the behavior of the cell over time, minute by minute.<br />
3.5 – Data Analysis<br />
A charge cycle occurs under an applied voltage while the<br />
system is removing salt. However, the electrodes will reach<br />
an adsorption capacity and cannot remove salt forever. A<br />
52 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
discharge cycle occurs when the voltage is removed or<br />
reversed, allowing electrodes to flush captured salt ions<br />
into a brine stream. During each cycle there are various<br />
performance metrics that researchers observe to assess<br />
system efficiency. These metrics are adsorption capacity,<br />
adsorption rate, and charge efficiency. Adsorption capacity<br />
refers to the mass of salt collected per mass of electrode<br />
material. Adsorption rate is an indication of the rate of salt<br />
adsorption as per mass of electrode. Charge efficiency is a<br />
measurement in the form of a percentage; a ratio of moles<br />
of salt removed per mole of electric charge.<br />
Since conductivity is directly proportional to salt<br />
concentration, we were able to derive each performance<br />
metric by finding the area under the effluent conductivity<br />
curve (Fig. 7).<br />
Figure 7. Conductivity versus time graph that<br />
graphically illustrates the importance of the<br />
integral of effluent conductivity in determining salt<br />
removed.<br />
The following demonstrates the mathematical analysis<br />
performed to derive each performance metric.<br />
Adsorption Capacity:<br />
Ultimately, these mathematical formulas are the key<br />
to transform raw data into meaningful analysis. These<br />
performance metrics are accepted throughout the larger<br />
CDI community.<br />
3.6 – Specific Aim 2<br />
Determine the relationship between charge and<br />
discharge cycle length and CDI cell performance by<br />
increasing the time during which voltage is applied to the<br />
system.<br />
3.7 – Rationale and Hypothesis<br />
The charge and discharge cycle length determine the<br />
time during which salt removal will occur. However,<br />
considering the adsorption capacity of electrodes, we<br />
expect for salt removal rates to vary as the pores become<br />
more filled with salt. Accordingly, proper cycle lengths<br />
are essential for an accurate measurement of electrode<br />
material performance. By analyzing the impact of cycle<br />
length on system efficiency, researchers can maximize<br />
the effectiveness of the electrode and determine the true<br />
potential of the material.<br />
Researchers expect longer cycle times to coincide<br />
with increased system performance. The large volume of<br />
electrode material should theoretically require more time<br />
to reach maximum adsorption. However, exceedingly long<br />
cycle times will decrease charge efficiency, as charge enters<br />
into electrodes that are unable to hold more salt ions. Thus,<br />
it is imperative that researchers systematically determine<br />
the proper charge cycle to enhance system performance.<br />
Ultimately, researchers expect cycle time to be<br />
significantly longer than the five-minute period that is<br />
adequate for smaller cells.<br />
3.8 – Supporting Preliminary Data<br />
Figure 8 illustrates the salt concentration over time<br />
for the first test run on the CDI cell. The test run below<br />
consisted of three, five-minute charging and discharging<br />
cycles.<br />
Charge Efficiency:<br />
Figure 8. Conductivity versus time graph for a flow<br />
rate of 5 mL/min, at 1200 mV, for 3 complete charge<br />
and discharge cycles each 5-minutes long.<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 53
This length was not adequate since the system was still<br />
removing salt at the end of the charging cycle. At the end<br />
of a charging period, effluent conductivity should return to<br />
influent conductivity so that the system reaches maximum<br />
adsorption and returns to a state of equilibrium. These<br />
findings demonstrate that the CDI system needs a longer<br />
charging cycle period, likely because of the large relative<br />
volume of the cell. This result is limited because it does<br />
not indicate what an adequate length would be, it merely<br />
demonstrates that it needs to be longer than 5 minutes.<br />
Thus, the researchers will be conducting systematic testing<br />
to determine the appropriate charge cycle time.<br />
3.9 – Methods<br />
For this specific test, researchers knew that the charge<br />
and discharge cycle needed to be longer than 5-minutes<br />
however, they did not know how long it needed to<br />
be. First, researchers decided to increase cycle time<br />
gradually in order to analyze the system behavior. This<br />
enabled researchers to analyze the system’s consistency<br />
as well as GAC performance under different operational<br />
parameters. Consequently, this heuristic continually<br />
yielded an inadequate cycle time. Hence, researchers<br />
decided to systematically determine the charge cycle by<br />
conducting a ‘single-cycle test’. In this test, researchers set<br />
the charge length to 300 minutes, and observed the data<br />
to determine the time at which the electrodes had reached<br />
their maximum adsorption and returned to equilibrium.<br />
3.10 – Data Analysis<br />
Researchers used the same mathematical and graphical<br />
approach to derive the performance metrics for the CDI<br />
cell as in Specific Aim 1.<br />
In addition to this quantitative data analysis, this data<br />
required graphical analysis based off of graph qualities.<br />
Researchers focused on observing the effluent versus<br />
influent conductivity at the end of each cycle time to<br />
observe whether the system was at equilibrium at the end<br />
of the cycle.<br />
3.11 – Specific Aim 3<br />
Determine the impact of design modifications on CDI<br />
cell performance by decreasing the total volume of the<br />
system.<br />
3.12 – Rationale and Hypothesis<br />
Although the filter was experiencing great increases<br />
in adsorption capacity, the charge efficiency was still very<br />
low. Charge efficiency is a measure of the percentage of<br />
electrical charge allotted to salt removal. A low charge<br />
efficiency indicates that much of the GAC is not removing<br />
salt and not receiving electrical charge. Researchers<br />
hypothesized that the large volume of the system was<br />
contributing to a poor distribution of electrical charge.<br />
Thus, system performance is expected to increase as the<br />
filter’s volume decreases.<br />
3.13 – Methods<br />
For this specific test, researchers decided to decrease<br />
the system’s volume by half. Researchers hypothesized that<br />
the large volume of GAC in the filter was contributing to<br />
the low charge efficiency, thus researchers anticipated that<br />
this modification would improve adsorption capacity and<br />
charge efficiency. The following image demonstrates the<br />
design modification that occurred.<br />
Figure 9. Graphic illustration of the design<br />
modification that decreased the filter volume from<br />
45.23 mL to 25.24 mL.<br />
Researchers decided to test the filter using the 5 mL/<br />
min flow rate because higher flow rates caused too many<br />
leakage issues. Moreover, the applied voltage of 1.2 V<br />
remained constant. A charge cycle time of 20 minutes was<br />
deemed appropriate after qualitative graph analysis.<br />
3.14 – Data Analysis<br />
Researchers used the same mathematical and graphical<br />
approach to derive the performance metrics for the CDI<br />
cell as in the previous specific aims.<br />
3.15 – Specific Aim 4<br />
Determine the impact of design modifications on<br />
CDI cell performance by rearranging anode and cathode<br />
attachment locations.<br />
3.16 – Rationale and Hypothesis:<br />
Although the filter once again experienced an increase<br />
in adsorption capacity, the charge efficiency decreased.<br />
Researchers hypothesized that the arrangement of the<br />
anode and cathode connection was not facilitating the ideal<br />
electron flow. Thus, researchers decided that increasing<br />
the distance between the applied voltages was necessary<br />
for the electrical field to encompass all of the GAC within<br />
the filter. Researchers hypothesize that this change may<br />
increase charge efficiency and overall system performance.<br />
3.17 – Methods<br />
For this specific test, researchers decided to change<br />
the location of the anode and cathode connection points.<br />
Researchers hypothesized that this modification would<br />
increase the reach of the electrical field, enable more GAC<br />
54 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
to be charged, and increase the filter’s charge efficiency.<br />
Figure 10 demonstrates the design modification that<br />
occurred.<br />
Figure 10. Graphic illustration of the design<br />
modification that increased the reach of the electric<br />
field by moving the anode and cathode connection<br />
plates further away from one another.<br />
Researchers decided to keep the operational parameters<br />
constant to ensure that the design modification was<br />
the only factor that could contribute to differences in<br />
performance. Thus, the flow rate remained 5 mL/min,<br />
the voltage applied remained 1.2 V, and the cycle time<br />
remained 20 minutes long throughout testing.<br />
3.18 – Data Analysis<br />
Researchers used the same mathematical and graphical<br />
approach to derive the performance metrics for the CDI<br />
cell as in the previous specific aims.<br />
4. Results<br />
to use a flow rate of 5 mL/min in future testing to avoid<br />
these issues. Nevertheless, these tests were successful in<br />
establishing baseline performance capabilities of GAC.<br />
4.2 – Impact of Cycle Time on Filter Performance<br />
The following table displays the performance of the<br />
CDI system as the cycle time increases. As expected,<br />
system performance increased as cycle time increased,<br />
since the cell spent more time at peak adsorption (Table 3).<br />
Additionally, the cell spent more time expelling salt during<br />
discharge cycles so the GAC was able to adsorb even more<br />
salt for a longer period of time.<br />
Table 3. Performance metrics comparison between<br />
elongated cycle periods demonstrates that the longer<br />
cycle time increased performance efficiency.<br />
Adsorption<br />
Capacity<br />
Charge<br />
Efficiency<br />
5 min 10 min 20 min 50 min<br />
20.2<br />
µg/g<br />
31.9<br />
µg/g<br />
96.0<br />
µg/g<br />
155.4<br />
µg/g<br />
22.32% 30.65 % 35.04% 35.64%<br />
Figure 11 displays the salt concentration over time for<br />
the lowest charge time tested (five minutes).<br />
4.1 – Impact of Flow Rate on Filter Performance<br />
After testing, researchers observed that a higher flow<br />
rate yielded more efficient filter performance. Flow rate<br />
directly impacts the performance metrics of the flowthrough<br />
CDI cell (Table 2). The lower flow rate was<br />
significantly less efficient than the higher flow rate.<br />
Table 2. The performance metrics of the flowthrough<br />
CDI cell at two different flow rates: 5 mL/<br />
min and 10 mL/min. Each performance metric rises<br />
with flow rate, demonstrating that higher flow<br />
rates increase performance.<br />
Adsorption<br />
Capacity<br />
Charge<br />
Efficiency<br />
5 mL/min 10 mL/min<br />
10.7 µg/g 20.2 µg/g<br />
13.125 % 22.32 %<br />
The novel system has a relatively large electrode<br />
volume, causing alterations in operation parameters<br />
to impact system performance less than expected.<br />
Accordingly, additional testing with a higher range of flow<br />
rate values may be necessary to cause greater variations in<br />
performance. Moreover, the larger flow rate introduced<br />
many leakage issues and pressure build-ups which<br />
increased internal system resistance. Researchers chose<br />
Figure 11. Salt concentration over time for a cycle<br />
time of 5 minutes. Operational parameters: applied<br />
voltage of 1.2 V, cycle time of 5 minutes, and flow rate<br />
of 5 mL/min.<br />
During this test, the filter was not at equilibrium at the<br />
end of the charge and discharge cycle periods. Here, very<br />
brief, ineffective discharge periods inhibited the amount<br />
of salt that the electrodes were able to adsorb. From this<br />
qualitative analysis, it was evident that cycle time must be<br />
increased. Figure 12 displays the salt concentration over<br />
time for an increased charge and discharge cycle length of<br />
10 minutes.<br />
ENGINEERING<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 55
indication regarding the potential of GAC to adsorb salt<br />
when operating under ideal conditions.<br />
Figure 12. Salt concentration over time for a cycle<br />
time of 10 minutes. Operational parameters: applied<br />
voltage of 1.2 V, cycle time of 5 minutes, and flow rate<br />
of 5 mL/min.<br />
Noticeably, the discharge cycles were more effective as<br />
the area under the curve during discharge cycles appears<br />
much larger, which was confirmed through quantitative<br />
analysis. However, the effluent and influent conductivities<br />
were still not equal at the end of the respective cycle time<br />
lengths (Fig 12). After increasing the cycle time again to<br />
20 minutes, graphical analysis once again demonstrated<br />
a need for increased cycle time. However, these results<br />
were limited because they did not indicate the ideal cycle<br />
time. Researchers conducted a ‘single-charge test’ to<br />
finally determine the optimal cycle time. In doing so, 50<br />
minutes was found to be ideal. The system performance<br />
was considerably higher under the 50-minute charge and<br />
discharge cycle time (Table 3). Figure 13 illustrates the salt<br />
removal over time under this elongated cycle time.<br />
Figure 13. Salt concentration over time for a cycle<br />
time of 50-minutes. Operational parameters: applied<br />
voltage of 1.2 V, cycle time of 5 minutes, and flow rate<br />
of 5 mL/min.<br />
4.3 – Impact of System Volume on Filter Performance<br />
Due to the exceptional volume of electrode material<br />
contained within the original design, researchers<br />
decided to decrease system size and analyze the impact<br />
on performance. The system design was maintained,<br />
researchers merely decreased the volume of each large<br />
graphite chamber to half of its original size. This change<br />
decreased the amount of electrode material from 20.04 g<br />
to 8.44 g. Figure 14 illustrates the salt removal over time<br />
using the smaller system.<br />
Figure 14. Salt concentration over time for the<br />
system after the design modification. Operational<br />
parameters: applied voltage of 1.2 V, flow rate of 5<br />
mL/min, and a cycle time of 20 minutes.<br />
Qualitative analysis demonstrates that the conductivity<br />
was nearing equilibrium at the end of the charge and<br />
discharge cycles, so a cycle time of 20 minutes was adequate<br />
for the smaller system. The performance metrics of the<br />
system were considerably higher than the larger systems,<br />
indicating an improved performance with the design<br />
modifications (Table 4).<br />
Table 4. Performance metrics and size comparisons<br />
between the two filters of different sizes illustrate<br />
that a decrease in filter size coincides with an<br />
increase in adsorption capacity but a decrease in<br />
charge efficiency. Researchers attribute the decrease<br />
in charge efficiency to an inadvertent decrease in<br />
GAC density.<br />
Large Filter<br />
Small Filter<br />
Volume 45.23 cm 3 25.24 cm 3<br />
Mass of GAC 20.04 g 8.433 g<br />
Density of GAC 0.433 g/cm 3 0.334 g/cm 3<br />
In this test, GAC demonstrated the impressive ability<br />
to remove salt at maximum adsorption for an extended<br />
period of time (~35 min.), which is a positive indication<br />
of GAC capability and system performance (Fig. 13). In<br />
conclusion, the results of this experiment were a positive<br />
Adsorption<br />
Capacity<br />
Charge<br />
Efficiency<br />
155.4 µg/g 287.7 µg/g<br />
35.6 % 27.3 %<br />
56 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
The adsorption capacity increased, indicating that the<br />
GAC in the filter adsorbed more salt than in previous tests.<br />
However, the charge efficiency decreased which meant that<br />
less charge was directed towards salt removal. Researchers<br />
hypothesize that the difference in GAC densities between<br />
the chambers caused this decrease in charge efficiency. As<br />
the chamber becomes less dense, it is more difficult for<br />
charge to be administered across the GAC, thus a lower<br />
charge efficiency should coincide with a lower electrode<br />
density. Nevertheless, the broader goal of this research<br />
project was to study the adsorption capabilities of GAC,<br />
using our filter as the avenue to do so. Thus, this increase<br />
in adsorption capacity was another promising sign.<br />
4.4 – Impact of Anode/Cathode Arrangement on Filter<br />
Performance<br />
In this design modification, researchers changed the<br />
location of the anode and cathode attachments to expand<br />
the amount of GAC impacted by the applied voltage (Fig.<br />
10). The results from this design modification are shown<br />
in Table 5.<br />
Table 5. Performance metrics before and after<br />
increasing the distance between anode and cathode<br />
attachment plates demonstrate that a wider<br />
electrical field significantly increases the charge<br />
efficiency and adsorption capacity of the system.<br />
Adsorption<br />
Capacity<br />
Charge<br />
Efficiency<br />
ENGINEERING<br />
Previous<br />
Design<br />
New Design<br />
287.7 µg/g 452.0 µg/g<br />
7.3 % 63.1 %<br />
This modification caused the most significant increase<br />
in charge efficiency experienced by the filter. Additionally,<br />
there was a large increase in adsorption capacity which was<br />
likely due to the amount of charge contributing towards<br />
salt removal. The distance between the applied voltages<br />
was much larger than before, likely causing the increase<br />
in charge efficiency. Moreover, charge efficiency reflects<br />
the performance of the filter, while adsorption capacity<br />
reflects the performance of the GAC. Thus, the correlation<br />
between increases in filter performance and increases in<br />
GAC performance indicate that GAC has even more<br />
potential to serve as an electrode material as the system<br />
design continues to improve.<br />
5. Discussion and Conclusions<br />
The aforementioned study established the efficiency<br />
of a novel electrode material, granular activated carbon,<br />
commonly used in portable water filters. Many industrial<br />
water filtration companies leverage GAC’s adsorptive<br />
capabilities to remove uncharged contaminants. Without a<br />
preexisting design, researchers leveraged their innovation<br />
and created a system that dispersed electrical charge<br />
across a chamber of GAC. Throughout experimentation,<br />
researchers have improved GAC’s adsorption capacity<br />
from ~10 µg/g to ~450 µg/g. The filter was initially<br />
invented as a lab-scale device aimed to determine the<br />
potential of GAC. This profound increase in adsorption<br />
capacity proves GAC to be a promising potential electrode<br />
material for use on the industrial scale. Moreover, the<br />
charge efficiency of the system was increased from ~13% to<br />
~63% over the course of various design modifications. It is<br />
important to consider that operational conditions are not<br />
yet ideal, and are evidently limiting system performance.<br />
Nevertheless, researchers proved that the electrochemical<br />
technique capacitive deionization is compatible for use<br />
with granular activated carbon. Thus, researchers have<br />
created a portable device that feasibly adapts GAC for salt<br />
removal. The low cost and energy requirements of this<br />
desalination technique will become a valuable resource<br />
to those impacted by the growing demand for freshwater<br />
resources. Furthermore, though currently untested,<br />
the device may have the potential to adapt GAC for the<br />
removal of other, more harmful, charged contaminants.<br />
6. Acknowledgement<br />
My work for this project was completed at North<br />
Carolina State University’s Environmental Engineering<br />
Lab from June <strong>2018</strong>- January <strong>2019</strong> under the mentorship<br />
of Dr. Douglas Call and Dr. Shan Zhu. Both of my mentors<br />
played fundamental roles in building my competency with<br />
capacitive deionization technology. Moreover, these<br />
mentors initially proposed the idea to design a filter that<br />
could utilize granular activated carbon (GAC) based upon<br />
their knowledge of the advantages of GAC. While the filter<br />
was designed, modeled, and assembled entirely by myself,<br />
I have sought their guidance throughout the development<br />
of my specific research aims to ensure continual system<br />
enhancement.<br />
7. References<br />
Jia, B., Zheng, W. (2016). Preparation and Application of<br />
Electrodes in Capacitive Deionization (CD): a State-of-Art<br />
Review. Nanoscale Research Letters, 11.<br />
Schutteb, C. F., Welgemoeda, T. J. (2005). Capacitive<br />
Deionization Technology TM : An Alternative Desalination<br />
Solution. Desalination, 183, 327-340.<br />
Avraham, E., Noked, M., Bouhadana, Y., Soffer, A.,<br />
Aurbach, D. (2009). Limitations of Charge Efficiency in<br />
Capacitive Deionization. Journal of the Electrochemical<br />
Society, 156, 157-162.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 57
Suss, M. E., Porada, S., Sun, X., Biesheuvel, P. M., Yoon,<br />
J., Presser, V. (2015). Water desalination via capacitive<br />
deionization: what is it and what can we expect from it?<br />
Energy Environmental Science, 8, 2296-2319.<br />
Porada, S., Zhao, R., Van der Wal, A., Presser, V.,<br />
Biesheuvel, P. M. (2013). Review on the science and<br />
technology of water desalination by capacitive deionization.<br />
Progress in Materials Science, 58, 1388-1442.<br />
Bian, Y., Huang, X., Jiang, Y., Liang, P., Yang, X., Zhang, C.<br />
(2015). Enhanced desalination performance of membrane<br />
capacitive deionization cells by packing the flow chamber<br />
with granular activated carbon. Water Research, 85, 371-<br />
376.<br />
58 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> ENGINEERING
LONG PRIME JUGGLING PATTERNS<br />
Daniel Carter and Zach Hunter<br />
Abstract<br />
There are a large variety of ways to juggle balls. Different juggling patterns can be modeled by a sequence of states that<br />
describe the positions of the balls in regular time intervals. A pattern is said to be prime if it does not repeat states more<br />
than once per cycle. We investigate the problem of finding the longest prime pattern for a given number of balls and<br />
maximum throw height. Solutions up to a maximum throw height of 9 were found by computer search. We completely<br />
solve the 2-ball case and provide a very strong upper bound for all other cases. This upper bound differs by no more than<br />
1 from every computed case.<br />
1. Introduction<br />
Juggling and mathematics are intricately connected.<br />
The math YouTube channels Mathologer (Polster &<br />
Geracitano, 2015) and Numberphile (Wright & Haran,<br />
2017) have both released videos on juggling. This introduction<br />
reiterates the information in those videos and introduces<br />
the main problem of this paper.<br />
There are many ways to juggle balls. For example, two<br />
basic 3-ball patterns are cascade, where the balls travel in a<br />
figure eight, and shower, where they travel in a circle. We<br />
can represent these patterns by following which hand the<br />
balls are in or traveling to over time in a ladder diagram,<br />
such as the ones shown in Figure 1.1.<br />
The left and right columns of dots represent the left<br />
and right hands, and the lines represent the paths of the<br />
balls. For the cascade, every ball is thrown so that it lands<br />
in the opposite hand 3 steps later. In other words, the ball<br />
is thrown to height 3. However, for the shower, the right<br />
hand throws balls to height 5 and the left hand throws balls<br />
to height 1.<br />
Jugglers assign siteswap notation to these patterns. This<br />
notation lists the sequence of throw heights in a pattern.<br />
For example, cascade has a siteswap of “3” and shower has<br />
a siteswap of “51.” It is worth noting that siteswap notation<br />
does not distinguish the left hand from the right. In fact,<br />
these patterns could be juggled using just one hand. Also<br />
worth noting is that there may be multiple siteswaps that<br />
refer to one pattern; for example, 51 and 15 represent the<br />
same pattern. Finally, a 0 in siteswap means all balls are in<br />
the air and there is no ball ready to be thrown.<br />
We can also describe the states reached by a pattern.<br />
Each state is a sequence of 1’s and 0’s representing the<br />
positions of the balls in the air. A 1 in the kth position<br />
indicates a ball in the air will land k steps later. At most<br />
one ball may be in each position, because two balls in the<br />
same position will fall into the same hand at the same time,<br />
which isn’t allowed in basic juggling. In the cascade, the<br />
only state is (111): one ball is always just about to land,<br />
one will land in two steps, and one will land in three steps.<br />
Jugglers call this state ground state, as it is the state with all<br />
balls in the lowest position. For the shower, the two states<br />
are (11010) before a throw of height 5 and (10101) before<br />
a throw of height 1. Two throws of a shower are shown<br />
diagrammatically in Figure 1.2.<br />
Figure 1.1. Ladder diagrams of the 3-ball cascade and<br />
3-ball shower.<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 59
e decomposed into the prime patterns 42 and 3. Prime<br />
patterns correspond to cycles on the graph, which are<br />
closed walks that do not repeat vertices.<br />
The number of (not necessarily prime) patterns is wellestablished<br />
(Takahashi, 2015). The more difficult question<br />
of the number of prime patterns has a partial answer<br />
(Banaian et al., 2015). We attempt to find the longest<br />
prime pattern for each combination of balls and maximum<br />
throw height.<br />
2. Empirical Results and Symmetry<br />
Figure 1.2. Converting location of balls to states.<br />
Bolded arrows indicate throws and are labeled<br />
with the throw height. Dotted arrows show balls<br />
dropping due to gravity.<br />
Reading from the bottom to the top, marking a 1 for<br />
every ball and a 0 for every gap gives the states (11010)<br />
and (10101). In this diagram, the balls are colored<br />
differently for clarity. However, we will consider each ball<br />
indistinguishable for our analysis.<br />
As seen, throws can change the state of the balls. In<br />
general, every throw each “1” moves left one place (i.e.<br />
the corresponding ball falls slightly) except for a ball in<br />
the leftmost position, which is thrown to some currently<br />
empty spot. We can make a directed graph describing<br />
every possible state and throw. A closed walk in this graph<br />
is a repeating pattern of throws — a juggling pattern. The<br />
graph for 3 balls with a maximum throw of 5 is shown in<br />
Figure 1.3.<br />
There are finitely many prime patterns, because for any<br />
finite graph, there are finitely many cycles. Therefore, a<br />
computer can search and find the longest prime pattern.<br />
Call L(n, b) the length of the longest prime pattern for b<br />
balls and maximum height n. The values of L(n, b) for 0 ≤ n<br />
≤ 9 are given in Table 2.1.<br />
Table 2.1. Lengths of longest prime patterns for max<br />
height 9 or less.<br />
For example, the value at b = 3, n = 5 is 8 because the<br />
longest prime pattern has siteswap 55150530, which is<br />
length 8. In Figure 1.3, this corresponds to the 8 states in<br />
the center of the diagram that are arranged in an octagon.<br />
Interestingly, the table appears symmetrical, with L(n, b) =<br />
L(n, n−b). We will now prove this. In fact, we will prove a<br />
somewhat stronger result.<br />
Figure 1.3. Juggling graph from 3 balls and max height<br />
5. Vertices represent states and edges represent<br />
throws. Vertices are labeled with the state they<br />
represent, and edges are labeled by throw height.<br />
We will denote the graph for b balls with max throw<br />
height n as J(n, b). The diagram above represents J(5,3).<br />
If a pattern visits each state no more than once, jugglers<br />
call it a prime pattern. This is because if a state is visited<br />
multiple times, the pattern can be decomposed into two<br />
or more prime patterns. For example, the pattern with<br />
a siteswap of 423 visits the state (11100) twice and can<br />
Theorem 2.1. There exists a bijection between patterns with b<br />
balls and n − b balls.<br />
Proof. Consider a valid juggling pattern for b balls with<br />
maximum height n. List the states of this pattern in order.<br />
Now, switch 0’s and 1’s, mirror each state left-to-right,<br />
and reverse the order of the list. This new list is a valid<br />
pattern for n − b balls. For example, there is a 3-ball pattern<br />
of height 5 with siteswap 5511. The states reached are, in<br />
order:<br />
(11100)<br />
(11001)<br />
(10011)<br />
(10110)<br />
60 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
The new list in this case is<br />
(10010)<br />
(00110)<br />
(01100)<br />
(11000)<br />
Which are, in fact, the states reached by the 2-ball<br />
pattern with siteswap 4004.<br />
To see why this bijection works, consider two seemingly<br />
unrelated questions: “What happens to the 0’s in the state<br />
after each throw?” and “What states could have led into<br />
some particular state?”<br />
For the first problem, there are three cases. The first<br />
is the case where there is a 0 in the leftmost position, so<br />
a throw of height 0 is the only option. In this case, all 0’s<br />
except the leftmost move left one position (i.e. fall) and<br />
a 0 appears in the rightmost position. Next, if a throw<br />
of maximum height is made, all 0’s simply move left one<br />
position. Finally, for any other throw, all 0’s move left one<br />
position, a 0 appears in the rightmost position, and one<br />
of the 0’s disappears because it was filled by the ball just<br />
thrown.<br />
For the second problem, there are also three cases.<br />
The first is if there is a 1 in the rightmost position, so the<br />
previous throw must have been maximum height. In this<br />
case, the previous state had the 1’s (except the rightmost 1)<br />
moved right one step, and there was a 1 was in the leftmost<br />
position. Next, there is the case where the previous throw<br />
was height 0, and the previous state had all 1’s simply<br />
moved right one position. Finally, for any other throw, all<br />
1’s were moved right one position, a 1 was in the leftmost<br />
slot, and one of the 1’s disappears because it has not been<br />
thrown yet.<br />
Clearly, these problems are equivalent! Simply swap 0<br />
and 1 and left and right. This accounts for the swapping of<br />
0’s and 1’s and the left-to-right mirroring in the bijection.<br />
The reversal of the order of states is a reversal of time,<br />
which comes from the statement of the second question.<br />
Due to this bijection, any pattern for b balls with max<br />
height n corresponds to a pattern for b gaps — that is, n − b<br />
balls.<br />
Corollary 2.2. L(n, b) = L(n, n − b).<br />
Corollary 2.3. To construct J(n, n − b) given J(n, b), reverse<br />
all arrows and relabel each vertex by switching 0 and 1 and<br />
mirroring left-to-right.<br />
Borrowing terminology from graphical linear algebra,<br />
we call the state formed after doing the bijection the bizarro<br />
of the initial state, denoted S * . We introduce the functions<br />
next and prev of a state S which return the set of possible<br />
states that could follow or precede S, respectively. From<br />
this theorem, if S 2<br />
∈ prev(S 1<br />
), then S 2<br />
*<br />
∈ next(S 1*<br />
).<br />
We will derive some basic upper bounds on the lengths<br />
of the longest prime patterns.<br />
3. Basic Upper Bounds<br />
Obviously, we cannot have a prime pattern with more<br />
states than the number of possible states.<br />
Lemma 3.1. The number of possible states is<br />
J(n, b) has vertices.<br />
. Equivalently,<br />
Proof. Each state is a permutation of b copies of 1 and<br />
n − b copies of 0. Therefore, the number of distinct states<br />
is .<br />
Corollary 3.2. L(n, b) ≤ .<br />
In fact, if b > 1 and n−b > 1, this inequality is strict because<br />
it is impossible to reach all states without repetition. This<br />
is proven below.<br />
Lemma 3.3. If b > 1 and n − b > 1, L(n, b) < .<br />
Proof. Consider the state with all balls in the highest<br />
possible position,<br />
Because this state ends in 1, the previous throw must<br />
have been max height and the previous state was<br />
However, this new state also ends in 1, so the previous<br />
throw must have been max height, and so on until all b<br />
copies of 1 are exhausted and the state is ground state,<br />
In other words, the only way to get to the original<br />
state S is to do b max height throws from ground state.<br />
However, because S begins with n − b copies of 0, the next<br />
n − b throws must all be height 0. After those throws, we<br />
return to ground state, closing the walk.<br />
This means that S is only reached by a single prime<br />
pattern. This pattern has length n, and for b > 1 and n −<br />
b > 1, n < . In other words, the longest prime pattern<br />
will either reach S and be length n or not reach S at all.<br />
Therefore, if b > 1 and n − b > 1, L(n, b) < .<br />
Now, consider the simple case where b = 2. Using a more<br />
complex argument, stronger bounds can be constructed.<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 61
4. The Case b = 2<br />
The argument hinges on simplifying the problem by<br />
considering the distance between the two balls, rather than<br />
their exact positions in the states. The distance between<br />
two balls in a state is the difference in the position of their<br />
corresponding 1’s. For example, the distance between the<br />
balls in the state (01001) is 5 − 2=3, because the first 1 is in<br />
position 2 and the second 1 is in position 5.<br />
With only two balls, the distance between the balls and<br />
the position of the first ball completely describe a state.<br />
However, notice that if the first ball is not in the leftmost<br />
position, the only possible throw is 0 until that ball falls<br />
into the leftmost position. Therefore, any pattern that<br />
reaches a state with some distance d necessarily reaches<br />
the state with distance d and a 1 in the leftmost position.<br />
This implies that each throw where height ≠ 0 in a prime<br />
pattern must lead to a unique distance.<br />
By considering only the states with a 1 in the leftmost<br />
position, we can construct a weighted directed graph with<br />
each vertex representing a unique distance and the weights<br />
on the edges indicating the maximum number of throws<br />
from one distance to another, using only one throw<br />
of height > 0. For example, the graph for 2 balls with<br />
maximum height 5 (or maximum distance 4), is shown in<br />
Figure 4.1.<br />
Then, an edge exists between states d and d′ if d′ < d or<br />
d + d′ ≤ n. Its weight is<br />
Rather than drawing the graph, it is simpler to consider<br />
a modified adjacency matrix where the entry in row x and<br />
column y is the W(x, y), if that edge exists. The example<br />
n = 5 is below.<br />
A cycle on the weighted graph does not repeat states,<br />
so it is also a prime pattern. Its length is the sum of the<br />
weights of the edges that it traverses. In the n = 5 case, the<br />
longest prime pattern created using this strategy is length<br />
8. The edges it traverses are circled in the matrix below.<br />
In fact, there is a general pattern that gives very long<br />
prime patterns and a lower bound on L(n,2).<br />
Lemma 4.1.<br />
Figure 4.1. Condensed 2-ball juggling graph with<br />
max height 5. Vertices represent states with a 1 in<br />
the leftmost position and edges represent throws.<br />
Vertices are labeled with distance, and edges are<br />
labeled with the number of states reached in the<br />
transition from one state to another.<br />
The edge from distance 3 to distance 2 has weight 3<br />
because the longest path from (10010) to (10100) is length<br />
3, given by the throws 5, 0, 0.<br />
The edge weights can be calculated easily. Every time a<br />
ball is thrown, it can either be thrown to a higher position<br />
than the other ball or to a lower position. If it is thrown<br />
higher, the next ball to land will be the second ball, which<br />
happens in d steps. If the first ball is thrown lower, say<br />
height h, it will be the next to land, h steps later. Let d′ be<br />
the target distance. Then h = d - d′ . Finally, this transition<br />
is only possible if d′ < d (so we can throw lower) or d + d′ ≤<br />
n (so we don’t throw above max height).<br />
Proof. Construct a sequence of distances as follows:<br />
• Begin with distance n − 1.<br />
• Go to .<br />
• Alternate across n/2, each time going to the distance<br />
closest to n/2 not yet reached. For odd n, begin by<br />
increasing the distance, and for even n, begin by<br />
decreasing the distance.<br />
• When distance 1 is reached, go to distance n − 1.<br />
For example, take n = 10. The sequence of distances<br />
formed by this procedure is 9, 5, 4, 6, 3, 7, 2, 8, 1. Looking<br />
at the matrix representation makes it much more obvious<br />
what this process does. For n = 10, the edges traversed are<br />
The sum of the weights of the edges traversed (the<br />
circled numbers) is the length of the pattern. This sum is<br />
62 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
upper bound on L in terms of C.<br />
For upper integer bound n, this is equal to .<br />
Writing it in this way shows a difference of just from<br />
the upper bound .<br />
In fact, we will later see that the notion is<br />
exactly L(n,2). This is due to a stronger upper bound for L<br />
that is derived by extending the notion of distance to cases<br />
with b > 2.<br />
5. Extension to b > 2<br />
For some state S with a ball in the lowest position, write<br />
the sequence of distances between each ball and the nexthighest<br />
ball, starting with the lowest ball. Call the sum<br />
of this sequence m and append n − m to the sequence to<br />
construct the distance notation of a state. We write distance<br />
notation in brackets and without commas or space<br />
between entries. For example, for the state (100101), the<br />
distance notation is [321].<br />
Distance notation is useful because after a max height<br />
throw and all subsequent height 0 throws, the distance<br />
notation rotates one place. Again taking the state (100110),<br />
after the siteswap 600, the state is (1010100) and the<br />
distance notation is [213]. The states corresponding to all<br />
unique rotations of a distance notation and all “in-between”<br />
states that have a 0 in the leftmost position form a subcycle:<br />
a set of states formed when doing only max height and<br />
height 0 throws. Each state is part of exactly one subcycle.<br />
Subcycles are very useful for finding long prime<br />
patterns, because a particular subcycle contains many<br />
states that cannot be reached by and cannot reach any state<br />
outside the subcycle. All states that end in 1 must have had<br />
the previous throw be max height, so the previous state<br />
was in the subcycle. Furthermore, all states that begin in 0<br />
must have the next throw be height 0, so the next state will<br />
be in the subcycle.<br />
Not all subcycles have the same number of states. For<br />
example, (1010) and (0101) form a subcycle with b = 2<br />
and n = 4, but (1100), (1001), (0011), and (0110) are also a<br />
subcycle with b = 2 and n = 4. If the number of states in a<br />
subcycle is m, the ratio n/m is the multiplicity of the subcycle,<br />
denoted with the letter x. Multiplicity can also be seen as a<br />
property of a state and is the number of times a string of 1’s<br />
and 0’s is repeated to form that state. For example, (1010)<br />
has multiplicity 2 because it is (10) repeated 2 times. States<br />
have the same multiplicity of the subcycle of which they<br />
are part. Clearly, x must be a divisor of n. x must also be a<br />
divisor of b, because each repetition must include the same<br />
number of balls. Therefore, x must be a divisor of gcd(n, b).<br />
Let C x<br />
(n, b) be the number of subcycles of multiplicity<br />
x with max throw n and b balls. We obtain the following<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
Theorem 5.1. If b > 1 and n − b > 1, L(n, b) ≤<br />
or equivalently L(n, b)≤<br />
means α is a divisor of β.<br />
. The notation α|β<br />
Proof. This bound essentially states that in each subcycle,<br />
we can hit at most one fewer state than the number of<br />
states in that subcycle.<br />
To see why this is true, consider all states with a 1 in<br />
the leftmost position. For brevity, we call these states<br />
grounded. These are the only states that can reach any state<br />
outside the subcycle. Consider a particular grounded state<br />
S. Then there is the next state in the subcycle S′ formed<br />
after doing a max height throw from S. The only state that<br />
can reach S′ is S.<br />
Now consider a prime pattern that includes all<br />
grounded states in a subcycle S 1<br />
,S 2<br />
,...,S n/x<br />
. Unless the prime<br />
pattern has no states outside this subcycle, at some point a<br />
throw lower than max height must be made from one of<br />
the grounded states S i<br />
. However, the state after S i<br />
in this<br />
subcycle could not be reached without repeating S i<br />
.<br />
Therefore, if a prime pattern includes states from<br />
multiple subcycles, it can hit at most one fewer than the<br />
number of states in each subcycle. The number of states<br />
in a subcycle of multiplicity x is n/x, so multiplying n/x<br />
− 1 by the number of subcycles with multiplicity x, then<br />
summing across all possible multiplicities gives the upper<br />
bound<br />
Equivalently, we can start with the total number of<br />
states and subtract 1 for each cycle to get<br />
The exceptions are when the longest prime pattern is<br />
actually just one subcycle, and the length of that subcycle<br />
is greater than the bound above. This only occurs when<br />
there is only one subcycle, which happens when b = 1 or<br />
b = 0.<br />
We will define<br />
for<br />
simplicity.<br />
How many subcycles of a particular multiplicity are<br />
there? We can construct several recurrence relations that<br />
uniquely define C x<br />
.<br />
Lemma 5.2.<br />
Proof. Recall that x counts the number of repetitions of a<br />
string needed to form a state with multiplicity x. Each of<br />
these strings is also a state with b/x balls and max height<br />
n/x. For example, (101010) is a state with 3 balls and max<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 63
height 6, and the repeating unit (10) is a state with 1 ball<br />
and max height 2.<br />
Every subcycle of multiplicity 1 with b/x balls and<br />
max height n/x uniquely determines a subcycle of<br />
multiplicity x with b balls and max height n. Therefore,<br />
.<br />
Table 5.2. Upper bound on length of longest prime<br />
pattern given by Theorem 5.1.<br />
Let s x<br />
(n, b) be the number of states with multiplicity x,<br />
max height n, and b balls. Clearly, C x<br />
(n, b) = , because<br />
each subcycle of multiplicity x has n/x states by definition.<br />
From the previous lemma, we have<br />
. We<br />
also have the following relation that involves s.<br />
Lemma 5.3.<br />
Proof. Each state has a unique multiplicity, so summing<br />
across all possible multiplicities yields all states.<br />
Table 5.3. Difference between the upper bound and<br />
actual value of longest prime pattern.<br />
This is enough information to calculate any value of C<br />
and s, and therefore the upper bound on L. As an example,<br />
we will find L≤(6,3):<br />
In fact, L(6,3) = 15.<br />
Below are tables for values of C 1<br />
, L≤, and L≤ − L. C x<br />
, and<br />
therefore L≤, is not defined for b = 0 or n − b = 0, so those<br />
entries are omitted.<br />
Table 5.1. Number of subcycles with multiplicity 1.<br />
Table 5.3 shows that in many cases, L≤ = L. However, for<br />
cases where n = 2b and b > 2 (the central values in every<br />
other row), L(2b, b) < L≤(2b, b). Before this is proven, we<br />
will establish some necessary conditions to lose only 1<br />
state in each subcycle, instead of more.<br />
We define the first grounded state in a subcycle. As the<br />
name implies, this is the grounded state in a subcycle that<br />
first appears in a prime pattern. Not every grounded state<br />
can be a first grounded state.<br />
Lemma 5.4. If a grounded state has 1 as its last distance, it<br />
cannot be a first grounded state.<br />
Proof. If a state has 1 as its last distance, it must have a 1 in<br />
the rightmost position. However, this means the previous<br />
throw must have been a max height throw, so the previous<br />
state was a grounded state in the same subcycle. Thus, a<br />
state with 1 as its last distance cannot be the first grounded<br />
state reached in a subcycle.<br />
Now consider, for example, the state (1101000), or in<br />
distance notation, [124]. If this is the first grounded state<br />
and the prime pattern only misses 1 state from its subcycle,<br />
then [241] and [412] will also be reached. The states<br />
missed are the non-grounded states “in-between” [412]<br />
and [124]. These are S 1<br />
= (0001101), S 2<br />
= (0011010), and<br />
64 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
S 3<br />
= (0110100). Out of these, only S 2<br />
and S 3<br />
can be reached<br />
from a state outside this subcycle, because S 1<br />
has a 1 in the<br />
rightmost position. If S 3<br />
was reached first, then both S 1<br />
and<br />
S 2<br />
will be missed. Then to miss exactly one state, assuming<br />
[124] is the first grounded state reached in the subcycle, S 2<br />
must be the first state reached in the subcycle.<br />
In general, the first state reached in a subcycle must be<br />
two states after some grounded state S G<br />
in a subcycle, if<br />
only 1 state is to be missed in that subcycle. We call this<br />
state an entry state for a subcycle. The singular state missed<br />
in that case is the state from throwing maximum height<br />
from S G<br />
.<br />
Which states can reach this particular entry state? From<br />
Theorem 2.1, this question is equivalent to asking for the<br />
bizarro of the states that can be reached by the bizarro of<br />
the entry state. In our example, the entry state is (0011010),<br />
which has bizarro (1010011). There are 4 states that can<br />
immediately follow this state: (1100110), (0110110),<br />
(0101110), and (0100111), which have bizarros (1001100),<br />
(1001001), (1000101), and (0001101). Obviously, we<br />
discard the last of these, because it is in the same subcycle<br />
as the entry state.<br />
The distance notation for the three states that work<br />
are [313], [331], and [421]. These would be the last states<br />
reached in their subcycle, or the leaving states, and the<br />
corresponding first ground states reached would be [133],<br />
[313], and [214]. Recall our original grounded state [124].<br />
Notice the state [133] is just [124] with the second-to-last<br />
distance incremented and the last distance decremented.<br />
Notice as well that the other two states both have 1 as their<br />
second-to-last distance. These are in fact the only two<br />
possibilities, a fact that we will prove.<br />
Before the proof, we introduce the function entry of a<br />
grounded state S G<br />
, which returns the unique entry state if<br />
the first grounded state reached in S G<br />
’s subcycle is S G<br />
. From<br />
the above example, entry((1101000)) = (0011010). We also<br />
introduce the function fg of a grounded state S H<br />
, which<br />
returns the unique first grounded state of S H<br />
’s subcycle if<br />
S H<br />
is the leaving state. fg(S H<br />
) is also the next grounded state<br />
after S H<br />
in S H<br />
’s subcycle. If fg(S H<br />
) = S G<br />
, then entry(S G<br />
) is the<br />
state formed after a max height and height 0 throw from<br />
S H<br />
.<br />
Lemma 5.5. Let S G<br />
with distance notation [d 1<br />
d 2<br />
...d b−1<br />
d b<br />
] be the<br />
first grounded state of a subcycle. Then let {S p1<br />
,S p2<br />
,...} be all states<br />
in prev(entry(S G<br />
)) but not in the same subcycle as S G<br />
. For each<br />
S pi<br />
, let S qi<br />
= fg(S pi<br />
). Then the distance notation of each S qi<br />
is either<br />
[d 1<br />
...d b−2<br />
(d b−1<br />
+ 1)(d b<br />
− 1)], or S qi<br />
has 1 as the second-to-last distance<br />
and either d b<br />
− 1 or d b<br />
− 1 + d 1<br />
as the last distance.<br />
Let fg(S H<br />
) = S G<br />
. Then<br />
We know the entry state is the state after a max height<br />
throw and one throw of height 0 from S H<br />
, so<br />
The bizzaro is<br />
We know<br />
. In fact, only<br />
entry(S G<br />
)* is omitted in<br />
There are three cases<br />
for possible throws from entry(S G<br />
)*: the aforementioned<br />
max height throw, a throw of height 1, and every other<br />
throw. We must consider only the latter two.<br />
Case 1: After a throw of height 1, the state is<br />
which has bizarro<br />
Then<br />
, which is<br />
S qi<br />
has distance notation [d 1<br />
...d b−2<br />
(d b−1<br />
+ 1)(d b<br />
− 1)]. This<br />
is the first possibility described by the lemma.<br />
Case 2: After a throw of height < n, the state is either<br />
Case 2a, where the ball is thrown somewhere in the middle:<br />
or Case 2b, where the ball is thrown close to the end:<br />
with the circled 1 representing the thrown ball. These<br />
two subcases are essentially the same and correspond to<br />
the two possibilities for the last distance, d b<br />
− 1 and d b<br />
− 1<br />
+ d 1<br />
. We will only show the rest of Case 2a, but Case 2b<br />
follows similarly.<br />
After absorbing the circle 1 into the adjacent groups,<br />
we have<br />
Proof. We will begin by constructing entry(S G<br />
). We have<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 65
which has bizarro<br />
not work because its last distance is 1. Therefore, pfg(S 1<br />
)<br />
consists of exactly one state S′ 1<br />
, which has distance notation<br />
We have S qi<br />
= fg(S pi<br />
), so<br />
Then S qi<br />
has a 1 as its second-to-last distance and d b<br />
− 1<br />
as its last distance, which is the second possibility described<br />
in the lemma. As mentioned before, Case 2b corresponds<br />
to the final possibility described in the lemma, with 1 as the<br />
second-to-last distance and d b<br />
− 1 + d 1<br />
as the last distance.<br />
As these are the only two possibilities, the proof is<br />
complete.<br />
We will denote pfg as the set of these previous first<br />
grounded states. That is, pfg(S G<br />
) is the set of fg(S pi<br />
) for each<br />
S pi<br />
in prev(entry(S G<br />
)) but not in the same subcycle as S G<br />
.<br />
There is the additional constraint that any S′ ∈ pfg(S G<br />
) must<br />
not have 1 as its last distance, because then S′ could not be<br />
a first grounded state from Lemma 5.4.<br />
We have the following useful corollary.<br />
Corollary 5.6. For any grounded state S that does not have<br />
1 as its second-to-last distance, there is exactly one S′ where<br />
S ∈ pfg(S′). This S′ has the same distance notation as S but with<br />
the second-to-last distance decremented and the last distance<br />
incremented.<br />
We now have the groundwork to tighten the bound for<br />
L(2b, b).<br />
Theorem 5.7. For b > 2, L(2b, b) < L≤(2b, b).<br />
Proof. This proof relies on the unique subcycle of<br />
multiplicity n/2. There are 2 states in this subcycle:<br />
From Theorem 5.1, we know that only S 1<br />
could ever be<br />
reached in a prime pattern, except if that pattern consists<br />
of just S 1<br />
and S 2<br />
. Consider pfg(S 1<br />
). From Lemma 5.5, each<br />
S qi<br />
∈ pfg(S 1<br />
) satisfies at least one of the following criteria:<br />
• The distance notation is<br />
• The second-to-last distance is 1 and the last distance<br />
is 2 − 1=1.<br />
• The second-to-last distance is 1 and the last distance<br />
is 2 − 1+2=3.<br />
The third possibility is actually the same as the first<br />
in this case. From Lemma 5.4, the second possibility does<br />
S 1<br />
is also the leaving state, so consider the possible<br />
states S i<br />
where S 1<br />
∈ pfg(S i<br />
). Because S 1<br />
does not have 1 as its<br />
second-to-last distance, Corollary 5.6 applies and the only<br />
state S where S 1<br />
∈ pfg(S) has distance notation<br />
This is actually S′ 1<br />
.<br />
Therefore, if only one state is to be missed in each<br />
subcycle, the subcycle containing S′ 1<br />
must both immediately<br />
precede and immediately succeed S 1<br />
. The only prime<br />
pattern that satisfies this consists of only that subcycle<br />
minus one state and S 1<br />
, so it has length n. For any b > 2,<br />
this is not as long as the longest possible prime pattern, so<br />
we miss out on S 1<br />
. Therefore, for b > 2, L(2b, b) < L ≤(2b, b).<br />
6. Concluding Remarks<br />
The cases where n = 2b are not the only cases where<br />
L
AN ANALYSIS OF A NOVEL NEURAL NETWORK<br />
ARCHITECTURE<br />
Vatsal Varma<br />
Abstract<br />
Artificial Intelligence is a rapidly growing field in computer science, and the pinnacle of this field is the Artificial Neural<br />
Network (ANN). Modeled after neuronal connections in the brain, neural networks have proved exceptional in locating<br />
and discriminating amongst patterns in vast datasets. Each neural network contains a multivariate function, which is<br />
known as the error function. Using a different optimization function, the neural network attempts to reach a minimum<br />
of its error function by reaching the respective minima of its weights and biases. This study aims to determine the effects<br />
of four different neural network architectures (NNA) on their overall convergence rates holding all other variables<br />
constant. The architectures are based on different types of neural networks: The Deep Residual Network (DRN), the<br />
Multilayer Perceptron Network (MLP), the Extreme Learning Machine (ELM), and one novel design dubbed as the<br />
Encoded Learning Machine (EncLM). A previous study used Boolean functions to determine the rate of optimization,<br />
and the novel design topped out of the tested networks. However, this study utilizes the Modified National Institute of<br />
Standards and Technologies (MNIST) Dataset, a dataset of images of handwritten digits. Each of the networks was run<br />
over the 60,000 images for one epoch, and within that epoch, was optimized every 100 images using backpropagation.<br />
It was determined that the MLP and DRN were the weakest networks for fast optimization as they took the longest to<br />
converge. The EncLM was once again the fastest architecture to converge upon a satisfactory result.<br />
1. Introduction<br />
1.1 – Neural Networks<br />
An artificial neural network (ANN) is an abstraction of<br />
the biological nervous system, using artificial neurons and<br />
axons to create a web and a means to solutions unfound.<br />
The popularity of such networks stems from their ability<br />
to adapt, learn and generalize. Due to these abilities,<br />
artificial neural networks can solve many computational,<br />
classification and pattern-recognition problems via a<br />
learning-based algorithm.<br />
In this study, every neural network is constructed and<br />
implemented with three factors remaining constant: the<br />
optimization method, the framework of the network, and<br />
the data used by each network. The data that each of the<br />
neural networks being tested will use are derived from the<br />
Modified National Institute of Standards and Technology<br />
(MNIST) dataset. The dataset in question is around<br />
60,000 images of handwritten digits, each of which has<br />
intrinsic properties that the network must derive to attain<br />
a successful output. Before specifying the steps taken to<br />
process the image, some notation needs to be defined. Let<br />
(w i<br />
, h i<br />
, l i<br />
), where w represents width, h represents height<br />
and l represents length, be defined by dimension d i<br />
. Each of<br />
the handwritten digits comes in an uncompressed format<br />
of d u<br />
= (28, 28, 1). Using a program, each one of those<br />
images was compressed to a size of d c<br />
= (24, 24, 1) to make<br />
computation and pooling easier for the neural networks.<br />
There were two parts to each of the networks tested<br />
in this study: the convolutional neural network, and the<br />
feed forward network. The convolutional neural network<br />
formed the back end of each of the networks, as it allowed<br />
further compression of the handwritten digits into a onedimensional<br />
input vector with meaningful data readable<br />
by the feed forward layer. The feed forward layer forms<br />
the front end of the network. The one-dimensional output<br />
vector determined by the convolutional neural network is<br />
then used as the input vector for the feed forward network.<br />
The feed forward architecture is what is being tested in<br />
this study. Thus, before describing the intricacies of each<br />
network, it is important to know how each network works<br />
mathematically.<br />
Before delving into the mathematics of neural<br />
networks, a few notation issues must be sorted out. Let n L j<br />
define each neuron in the feed forward layer. Let o L define j<br />
the activation of the neuron in layer L at position j and<br />
β L define the bias of the neuron at layer L and position j.<br />
j<br />
Similarly, let ī L define the net input a neuron receives.<br />
j<br />
L<br />
Next, let w 1 ,L 2<br />
j,k<br />
define the weight of the link between neuron<br />
j of layer L 1<br />
and neuron k of layer L 2<br />
. Finally, let σ define<br />
the activation function of the neuron.<br />
Figure 1. An artificial neuron model.<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 67
Each neuron in the feed forward network is derived<br />
from the McCulloch & Pitts neuron model (Fig. 1).<br />
The model describes neurons as synaptically linked to<br />
each other, and each neuron may have multiple links<br />
to multiple other neurons. Each link to a neuron holds<br />
a specific value, called its weight w, as described earlier.<br />
That value represents the importance of the link to the<br />
neuron the information is going to. To exemplify, say<br />
there existed a neural network with two layers L 1<br />
and L 2<br />
.<br />
Layer L 1<br />
has two neurons, and L 2<br />
has one neuron. In this<br />
network, there would only be two links, with weights, w 1<br />
= w 1,2 and w = 1,1 2 w1,2.<br />
If w = 0 it would mean that the input<br />
2,1 1<br />
of the neuron ī 2 would remain unaffected by the output<br />
1<br />
o 1 . This is also reflected in the way the net input of each<br />
neuron is calculated.<br />
The input of each neuron in successive layers is<br />
calculated based on the sum of the product of the output<br />
of each neuron and the respective weight of the link<br />
propagating that output.<br />
A neuron not only holds its net input, but also is<br />
responsible for calculating its net output, which is a<br />
function of its net input ī and its bias β. Each neuron’s<br />
output is calculated as follows where σ is representative<br />
of the sigmoid function, and this is true for both the<br />
convolutional neurons and the feed forward neurons.<br />
In this model, three activation functions are used:<br />
sigmoid σ(n), hyperbolic-tangent h(n) = tanh(n), and<br />
exponential linear units ε(n), a modification of the rectified<br />
linear units function.<br />
The sigmoid activation function suppresses each of the<br />
outputs within a range of (0,1). The hyperbolic tangent<br />
serves a similar purpose and suppresses the outputs within<br />
a range of (−1,1). The Exponential Linear Units serves a<br />
different purpose. It is used on the convolutional neurons<br />
within the convolutional neural network part of the<br />
entire network. Since the convolutional neural network<br />
(CNN) is tasked with compressing an image, its inherent<br />
purpose is to process each pixel of an image. The value<br />
of each pixel locus is the atomic number of the element<br />
present at that location, and zero otherwise. This is not<br />
adequately processed by the sigmoid or hyperbolic tangent<br />
functions, as when the pixel values become larger and<br />
larger, the hyperbolic tangent and sigmoid functions will<br />
become more and more saturated. Furthermore, negative<br />
pixel values are typically regarded as zero. That is why the<br />
Convolutional Neural Network utilizes the exponential<br />
linear units activation function instead of another.<br />
There are two types of neurons in the CNN, the<br />
convolutional neuron n C<br />
and the pooling neuron n P<br />
. The<br />
n C<br />
neurons operate in a similar fashion to the feed forward<br />
neurons, but the n P<br />
neurons have a different purpose. Each<br />
of the n P<br />
neurons take a (2, 2) section of the image and<br />
finds the largest value within its section, and sets that value<br />
as its output. This essentially carries the most important<br />
pixel value for the next layer of processing. As the network<br />
progresses layer by layer, the image is compressed further<br />
and further until it becomes a one-dimensional input<br />
vector for the fully connected layer.<br />
The CNN, like the feed forward network, is built in<br />
layers; however, the way those layers are designated is<br />
completely different from the feed forward network.<br />
The CNN operates through filters and convolutions.<br />
Essentially, a filter is a set of weights which are applied<br />
to sections of the image to create a net input for the<br />
convolutional neuron that filter is sending its data to. A<br />
convolutional layer can be described with a dimension d i<br />
where its length represents the number of filters that layer<br />
has. For example, the image of dimensions d c<br />
= (24, 24, 1)<br />
is convoluted upon to create a layer of size d L<br />
= (24,24,3).<br />
This means that there are three filters in that convolutional<br />
layer, each responsible for one (24, 24) section of that<br />
layer. To further explain how filters work, let there be<br />
three filters f 0<br />
, f 1<br />
and f 2<br />
, for the layer discussed above. Each<br />
filter acts upon the entire depth of the input image, thus<br />
the length of the filter must be the length of the previous<br />
image. Say the dimension of the filter was d f<br />
= (3, 3, 1).<br />
Filter f 0<br />
would operate on consecutive three by three by<br />
one sections of the input image. In the next convolutional<br />
layer, each filter would operate on consecutive three by<br />
three by three sections of the previous layer, and so on.<br />
Between convolutional layers exists a pooling layer, which<br />
further compresses the layer. For example, if the layer<br />
was of size (24, 24, 13) the max pooling layer would be of<br />
size (12, 12, 13). That is essentially how a CNN operates.<br />
Successive layers will convolute upon the previous layer’s<br />
output image, and slowly pool the image down to a<br />
manageable size. That output vector will then be used as<br />
an input vector for the feed forward network.<br />
All of the above tools, when put together, can be<br />
68 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
used to deliver an output of the network that classifies<br />
the handwritten digit with its respective value. This is<br />
known as the feed-forward stage. Initially that output will<br />
be meaningless, and it will remain so until the network<br />
is trained. The training vector t → is built using already<br />
optimized geometries to train the network. t → will be of the<br />
dimension of the input image. To then train the network,<br />
the error of the network is calculated using t → and the<br />
Mean Squared Error formula (MSE), and then that error is<br />
backpropagated throughout the network.<br />
If t → defines the correct/preferred output of the network,<br />
and → ρ is the actual output of the network, MSE can be<br />
calculated as follows.<br />
(3)<br />
Backpropagation has three main procedures: first, to<br />
determine the effect that a neuron’s output has on the<br />
error; then to determine the effect that the neuron’s bias<br />
and net input has on that same error; and finally, using<br />
the value calculated by the net input, to determine the<br />
effect that each link’s weight has in that error. Through<br />
backpropagation, the network is attempting to minimize<br />
the error function MSE(t → , → ρ) by calculating its negative<br />
gradient.<br />
To do this, the network calculates the partial derivative<br />
of each weight and bias with respect to the error function.<br />
Symbolically, this can be represented as for bias<br />
and for weights. By applying the chain rule we can<br />
expand these basic equations to finish the implementation<br />
of the entire backpropagation rule. The first step is to<br />
calculate a δ value for each neuron which indicates the<br />
direction the neuron’s output needs to step for it to reach<br />
a minimum. The δ is calculated differently depending on<br />
whether the neuron in question is an output neuron or<br />
not.<br />
(4)<br />
In these functions the sigmoid activation function can<br />
be replaced by any other activation function discussed<br />
earlier in the Introduction.<br />
Using this δ the network is able to calculate the<br />
effects of the weights and biases on the total error of the<br />
network and take a step down the gradient of the error<br />
function. The equations below define the backpropagation<br />
algorithm where η β<br />
is the constant that determines the<br />
size of the step the bias β must take, and η w<br />
is the constant<br />
that determines the size of the step the weight w of each<br />
individual link must take.<br />
(5)<br />
(6)<br />
All of the above will remain constant in this study.<br />
The only variable will be the neural network architecture,<br />
or, in other words, the number of connections and how<br />
each network is linked together. What changes, then, is<br />
how the gradients δ are calculated. If different neurons<br />
are connected, then their outputs will differ based on the<br />
outputs of the neurons they are connected to. Thus, what<br />
happens if the neurons are connected in specific ways?<br />
How does that architecture change the performance of<br />
the network? Most importantly, is it possible to hybridize<br />
two architectures and obtain properties of both? This<br />
study is designed to test the differences between various<br />
neural networks based purely on their architecture.<br />
Using the image dataset, each neural network will be<br />
run for the entirety of 60,000 iterations, during which<br />
data corresponding to the neuron will be collected every<br />
iteration, and data corresponding to the network will<br />
be collected only when the total error is backpropagated<br />
(every 100 iterations). Data will include the neuron biases<br />
and activations, link weights and the overall network<br />
error. The aim of this study is to find the quickest and most<br />
efficient NNA convergence rate based on its architecture<br />
and architecture only. Furthermore, using the knowledge<br />
gained from the three initial networks, a novel NNA by the<br />
name of the Encoded Learning Machine, which employs<br />
principles of various networks in attempt to obtain faster<br />
and more efficient convergence, can be implemented.<br />
2. Computational Approach<br />
The designing and visualization of these NNAs took<br />
place in two parts. First, each neural network was designed,<br />
implemented and tested in Java. Then, using Mathematica<br />
and Microsoft Excel, the data were visualized and each<br />
neural network was heuristically evaluated and given a<br />
score relative to the other networks. A high score meant<br />
the network converged faster than the other networks;<br />
a lower score meant that either the network’s error was<br />
high, or the network’s accuracy was low. However, before<br />
delving into the object-oriented implementation of these<br />
NNAs, an overview of the structures must be given.<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 69
2.1 – Structure Overview<br />
sets. They have further been used to obtain "higher-quality<br />
contact prediction” data for proteins (Wang, 2017). The<br />
DRN most likely will not show its full potential during this<br />
study due to all networks only being 4 layers deep, when<br />
DRN architectures can be upwards of 100 layers.<br />
Figure 2. A Multilayer Perceptron (MLP) model<br />
visualized in STELLA.<br />
There are four different NNAs used in this study:<br />
The Multilayer Perceptron, the Deep Residual Network,<br />
the Extreme Learning Machine, and finally the Encoded<br />
Learning Machine. The MLP is one of the most basic<br />
implementations of a neural network. Generally, an MLP<br />
consists of an input layer, usually represented as a vector,<br />
an output layer, and n hidden layers (Fig. 2). To clarify,<br />
layers are simply objects that hold an array of neurons.<br />
Each neuron is then connected with every neuron in the<br />
layer in front of it with links starting from the input layer<br />
and ending at the output layer. The MLP architecture is a<br />
simple and efficient structure that has been proven to be<br />
able to organize and classify data. It is often used as a part<br />
of a larger neural network.<br />
Figure 3. A Deep Residual Network (DRN) model<br />
visualized in STELLA.<br />
The DRN is quite like the MLP, but instead of only<br />
connecting consecutive layers, it can include connections<br />
that span more than one layer at regular intervals<br />
throughout the network (Fig. 3). The theory behind<br />
this network is that previous input is being forward<br />
propagated many layers to prevent loss of data and enhance<br />
generalization capabilities. DRN architectures have proven<br />
adept at generalizing images and other complicated data<br />
Figure 4. An Extreme Learning Machine (ELM)<br />
model visualized in STELLA.<br />
The ELM is basically a network with an input layer,<br />
a hidden layer and an output layer. The only catch is<br />
that there exist random forward links from each neuron<br />
which can connect to any other neuron in the network<br />
if the prospective neuron is not in a layer behind the<br />
neuron requesting connection (Fig. 4). This network was<br />
designed to reduce the slow training speed of other types<br />
of neural networks (Ding, 2015). It has also proven to be<br />
better at generalization and have a faster learning rate<br />
(Ding, 2015). Furthermore, due to the stochastic nature<br />
of the connections, the number of hidden layers becomes<br />
arbitrary, therefore making it pointless to initialize this<br />
network with multiple hidden layers like the others. This<br />
network is actually trained differently from the rest of the<br />
networks when applied in other cases; however, in this<br />
study it remains like a network with random connections.<br />
The design was simply an experiment to analyze how a<br />
neural network with such connections would behave<br />
when trained by backpropagation.<br />
Finally, the EncLM is a hybridization of two different<br />
types of architectures: an autoencoder and the ELM. An<br />
autoencoder takes an input vector and transposes it to a<br />
layer identical to it (Fig. 5). This essentially means that<br />
it encodes the same information in a different pattern,<br />
meaning that more intricate aspects of the data can be<br />
seen by the network. Furthermore, due to the speedy<br />
performance, but higher average error, of the ELM,<br />
it became the second part of this network. Due to the<br />
autoencoder, it was theorized that if the neural network<br />
was able to process the same information coming from<br />
more neurons, it would be able to step down the gradient<br />
faster and more efficiently.<br />
70 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
Figure 5. An Encoded Learning Machine (EncLM)<br />
model visualized in STELLA.<br />
A convolutional neural network’s job is to compress a<br />
two or three-dimensional input tensor, such as an image,<br />
into a one-dimensional vector that can be processed by the<br />
front end neural network. Each convolutional network<br />
is based on the idea of filters, where each filter “learns”<br />
to differentiate certain aspects of that tensor from other<br />
aspects. For handwritten digits, filters might be used to<br />
recognize loops or edges within the numbers. Each layer<br />
in the convolutional neural network depends on the filter<br />
assigned to it. Each convolutional neural network in this<br />
study had 10 filters of size five pixels by five pixels. At the<br />
end of the convolutional network, a pooling layer was<br />
constructed that normalized and compressed the previous<br />
outputs so that the front end networks had less data to<br />
analyze, and therefore saved memory on the computer.<br />
For this network, one convolutional layer of dimension<br />
d L<br />
= (24, 24, 10) was used. Ten filters of size (5, 5) pixels<br />
were applied to this layer. After this layer determined its<br />
output, the next layer was the max pooling layer, which<br />
condensed the output of the convolutions into a smaller<br />
space and allowed faster forward propagation of the<br />
network. Each neuron in the CNN was connected to the<br />
neurons in the previous layer based on the size of its filter.<br />
The feed forward network had 4 layers, each with 32<br />
neurons as the initial layer, 20 neurons, 16 neurons and<br />
finally 10 neurons as the output layer. The input neurons<br />
of the feed forward layers were connected to the output<br />
neurons in the max pooling layer of the convolutional<br />
neural network.<br />
This is a brief overview of how the architectures were<br />
set up in this study.<br />
2.2 – Data Generation<br />
The initial objective of the NNA implementation was<br />
to write the algorithm that converted the MNIST data into<br />
a readable form. Using a small Python script, each image<br />
of the 60,000 images in the dataset was converted into a<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
one-dimensional input vector and put into a file readable<br />
by the Java ParseCSV class. Once this was complete, each<br />
one of those images needed to be assigned a classification<br />
vector → t . By obtaining the correct value for each image,<br />
the ParseCSV class was able to create classification vectors<br />
for each image. Each vector would be used to train the<br />
network after it determined the output on the input image.<br />
After generating the datasets, the NNAs had to be<br />
implemented as well. To keep implementation easy<br />
and understandable, each neural network is based on<br />
a superclass, NeuralNetwork.java, which allowed each<br />
subclass, corresponding to the neural networks in this<br />
study, to be initialized by defining how they are run,<br />
trained, and connected. All neural networks were run by<br />
forward propagation, which involves taking all the layers<br />
of a network, starting from the input layer, and calculating<br />
the output or activation o L of each neuron within that<br />
j<br />
layer via equation 2 until it reached the final neuron,<br />
which, when its output was calculated, would represent<br />
the network’s integer answer to → b.<br />
While a run operates, the neural network writes<br />
data to its respective csv file. Data are written either<br />
every activation cycle, or every optimization cycle.<br />
Every activation cycle, each neuron’s activation o L j<br />
and its net input ī L is written to a file by the name of<br />
j<br />
”NetworkNameNeuronData.csv” where NetworkName<br />
would be replaced by the acronyms given to each network<br />
in this study. Furthermore, since the weights and biases<br />
are updated every optimization cycle, those are written in<br />
a smaller file by the name of “NetworkNameNetworkData.<br />
csv”. This file also contains the network error data that will<br />
be crucial to the analysis of these NNAs. During every 100<br />
iterations, or optimization cycle, the network is trained<br />
through backpropagation. This is done in two steps: first,<br />
recursively going backwards along the various links and<br />
neurons and creating a δ value for each neuron according<br />
to equation 4; then, again recursively back propagating<br />
along the links and neurons to update the biases β L and j<br />
L<br />
weights w 1 ,L 2 L L<br />
j,k<br />
according to the δ j<br />
and o j<br />
of the neuron (see<br />
equations 5 and 6).<br />
This summarizes, in brief, the constants and variables<br />
to which each neural network will be subject.<br />
3. Results<br />
Each network has a vast amount of data to analyze and<br />
visualize. Therefore, to make this section more organized,<br />
it will be split into two subsections: error & accuracy,<br />
which will discuss the evaluation, speed and efficiency of<br />
the networks, and visualization, which will discuss what<br />
is being visualized and why it was chosen to represent the<br />
neural network.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 71
3.1 – Error and Accuracy<br />
The objective of this study was to determine the highest<br />
performing network based on two parameters, accuracy a<br />
and error e. By those two metrics, a heuristic score can be<br />
assigned to each of the networks, where the higher score is<br />
the better score, based on the function H as shown below.<br />
As accuracy increases, the score increases, and as error<br />
decreases the score increases. In other words H(a, e) ∝ a<br />
and H(a, e) ∝ 1/e.<br />
Before delving into the meaning of each one of those<br />
numbers, definitions of error and accuracy are required.<br />
Error is the average difference between the perfect output,<br />
and the actual output of the network. It can range from<br />
−∞ to ∞. Accuracy is the percentage of the images the<br />
network classified correctly while training. It can only<br />
range from 0 to 100. Finally, the prediction percentage of<br />
the network is how sure the network is of its answer.<br />
Table 1. Error and Accuracy Averages for the Two<br />
Trials.<br />
TRIAL DRN ELM MLP ENCLM<br />
1-Error 0.030128 0.005262 0.034142 0.006175<br />
2-Error 0.038433 0.016085 0.039846 0.03341<br />
1-<br />
Accuracy 0.1374 0.118417 0.113433 0.125283<br />
2-<br />
Accuracy 0.153573 0.095943 0.124493 0.217846<br />
Avg.<br />
Accuracy 0.145487 0.10718 0.118963 0.171565<br />
Avg.<br />
Error 0.034281 0.010674 0.036994 0.019793<br />
Scores 424.395 1004.12 321.574 866.796<br />
For each of these networks, interesting observations<br />
can be gathered. Of course, if more trials were performed,<br />
the data would reflect the actual performance of the<br />
network more accurately. According to Wang et. al’s<br />
paper, the DRN was particularly adept at accurately<br />
classifying various things, especially images (Wang, 2017).<br />
Seeing the scores, this seems to be true, as the DRN has the<br />
second highest accuracy of all of the network architectures<br />
tested in this study. This means that, out of every 100<br />
predictions, about 15 predictions were correct, even if<br />
they were predicted by a low margin. That low margin is<br />
indicated by the relatively high error value that the DRN<br />
has. In other words, the DRN can guess correctly, but it is<br />
quite unsure about its guesses. For example, if an image of<br />
a three was input into the DRN, it may guess it correctly,<br />
but its prediction percentage for an eight would also be<br />
relatively high.<br />
Next, the ELM architecture was the highest scoring<br />
of all of the networks. The random connections of the<br />
architecture seemed to have helped it achieve the low<br />
error. However, its high score is deceiving. The accuracy<br />
of the ELM is the lowest of all of the network architectures<br />
tested in this study, a mere 10.718% across 60,000 images.<br />
Its very low error, coupled with its low accuracy, can only<br />
be attributed to one occurrence; the network was trained<br />
in a manner that associated several features with the wrong<br />
values. For example, if a three was input into the network,<br />
it would guess that eight was the answer because of the<br />
similar features, with a very high prediction percentage. It<br />
was sure of its answer, even if that answer was incorrect.<br />
Third, the MLP architecture was the lowest scoring<br />
of all of the networks. The orderly connections of the<br />
architecture as seen in Fig. 2 seem to have inhibited the<br />
network from learning the nuances present within the<br />
structures of each of the handwritten digits. The MLP<br />
had both the lowest accuracy as well as the highest error.<br />
For example, if a three was put into the network, it would<br />
be unsure of what the output should be, and would guess<br />
based on whatever it found familiar, leading to a guess of<br />
eight, nine, and sometimes, three.<br />
Fourth and finally, the new hybrid network<br />
architecture, EncLM, was the second highest scorer<br />
of the four tested networks across the two trials. The<br />
orderly connections that form its first two layers, instead<br />
of inhibiting the network’s performance, seem to have<br />
enhanced it. The worst performer combined with the best<br />
performer created a hybrid network with new capabilities.<br />
It had the highest accuracy, with the second lowest error.<br />
Comparing this with the other architectures, if a three was<br />
put into the EncLM architecture, the architecture would<br />
guess three and be fairly sure of its guess, meaning that it<br />
would have a high prediction percentage.<br />
From these numbers, this is a summary of the<br />
observations that can be made. The performance of each<br />
of the networks can further be broken down when looking<br />
at their performance over time.<br />
3.2 – Visualization<br />
For each one of the networks, the accuracy over time<br />
a(t) and the error over time e(t) were plotted (Fig. 6, 7, 8,<br />
9). The trends visible in each of the graphs indicate the<br />
aspects of the networks discussed above, but they show the<br />
networks’ training speed as well.<br />
A more accurate network will have an a ′(t) slope larger<br />
than most networks. In this case, by observation, the DRN<br />
has the fastest increasing accuracy. Perhaps, after several<br />
iterations over the same data, the DRN will have the<br />
72 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
highest accuracy out of all of the networks. One thing that<br />
the a(t) graph indicates about the network is its potential<br />
to learn within the limited vision it was given. Obtaining<br />
a higher average accuracy is often indicative of the faster<br />
training of the network. It is true in all cases that the<br />
accuracy value goes up as the number of iterations goes<br />
on. Where the real differences in the networks can be seen<br />
is in the e(t) graphs.<br />
The faster training network will have an e ′(t) slope<br />
value greater in magnitude to all of the other networks.<br />
That network will reach the error threshold, where the<br />
error begins to flatten, much more quickly than the other<br />
networks. During backpropagation, such a network is<br />
more likely to take the correct step down the gradient<br />
of its error function −∇MSE(t → , → ρ). Furthermore, another<br />
property of the e(t) graphs is the stochastic nature of the<br />
values, or how far they deviate from a proper curve. It is<br />
here where the differences are quite noticeable between<br />
networks.<br />
As the networks trained, the data corresponding<br />
to each error and accuracy were collected every 100<br />
iterations. Using Wolfram Mathematica, those data were<br />
then visualized.<br />
Figure 8. Trial one error data for each network.<br />
Figure 9. Trial two error data for each network.<br />
4. Discussion<br />
Figure 6. Trial one accuracy data for each network.<br />
Figure 7. Trial two accuracy data for each network.<br />
MATHEMATICS AND COMPUTER SCIENCE<br />
Ultimately, the goal of the study is to prove that the<br />
new hybrid network architecture is viable for use in<br />
various situations. Furthermore, this study can open up a<br />
new area of neural network research, where properties of<br />
two different architectures, whether they be mathematical<br />
or structural, can be hybridized to obtain a hybrid network<br />
that reflects the desired properties of both networks. In<br />
the case of the EncLM, the high accuracy of the DRN<br />
architecture and the fast training of the ELM architecture<br />
were the desirable properties. Over both studies, the<br />
EncLM expressed both of these properties, becoming a<br />
fast training and highly accurate network.<br />
As with any hybridization, unexpected results come up.<br />
Those results manifested themselves in the form of the<br />
error graphs e(t) of each network.<br />
Each error graph looks similar, but varies in one aspect,<br />
the deviation of the error in between each iteration. The<br />
MLP has the least deviation, and the EncLM has the most<br />
deviation. This deviation is akin to exploration. The more<br />
the error deviates, the more the network is exploring its<br />
individual error function to locate its minima. Half of this<br />
is luck. The network might reach optimized values that<br />
could drop the error to a very low value. The other half is<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 73
exploration, or how much the neural network is willing<br />
to deviate from certain relative minima to find the next<br />
lowest possible error. This feature is indicative in the<br />
lowest error values observed within each network. The<br />
stochastic connections within the EncLM and ELM gave<br />
the two networks error values that were nearly a sixth of<br />
the more orderly DRN and MLP architectures. A random<br />
set of connections, it seems, enables a network to see the<br />
input data as a whole, rather than seeing it in layers. This<br />
allows the network to traverse its respective error function<br />
rapidly. However, the one drawback is that the network<br />
cannot determine with certainty whether the output it<br />
creates is the correct one. For the orderly networks, this<br />
was their strength, especially in the DRN. Even with its<br />
high error, it was able to accurately classify each digit.<br />
The two properties of the DRN and ELM, when<br />
combined, seem to have amplified each of their individual<br />
effects. The exploratory nature of the EncLM is enhanced<br />
by the DRN’s orderly connections, and the accuracy of the<br />
network overall is higher than all other networks in the<br />
study.<br />
5. Conclusion<br />
The EncLM is, ultimately, a hybrid network architecture,<br />
employing tools from both orderly connected networks as<br />
well as stochastic connected networks. The end result is<br />
very satisfactory. The error it achieves is comparable to<br />
the error the ELM achieved, and the accuracy is higher<br />
than all other networks.<br />
Overall, the novel architecture proved to be an<br />
intriguing development in neural network architectures,<br />
as it furthered the idea of speedy and efficient convergence<br />
to a global minimum. It is safe to say that the novel<br />
architecture is viable in all aspects when compared to the<br />
architectures tested in this study. To further this study,<br />
more research will be needed to determine whether the<br />
properties of the EncLM can be further generalized to<br />
more complex and larger datasets, maybe involving larger<br />
and more intricate images than handwritten digits. If this<br />
is possible, research also needs to be done to determine<br />
the convergence rates of the other architectures in this<br />
study on that same dataset to determine whether a more<br />
organized structure like that of the MLP, or DRN will<br />
be able to notice the complicated patterns present in the<br />
new dataset, or whether a similar pattern of stochastic<br />
dominance in this study will extrapolate onto that dataset.<br />
Furthermore, the possibilities of architecture mixing<br />
could potentially have uses in business, industry and other<br />
fields that require the management of more than one task<br />
at the same time. Another study could be carried out to<br />
determine the effects of mixing more architectures on a<br />
similar dataset. This could be used to determine the hybrid<br />
architecture that provides the best possible results for any<br />
problem.<br />
6. Acknowledgments<br />
The author would like to thank Mr. Robert Gotwals for<br />
his sincere and expertful management and his fascinating<br />
insights into various tools and means that were used in<br />
this paper, including Excel, Mathematica and LaTeX. The<br />
author would also like to thank his mentor Mr. Keethan<br />
Kleiner for interesting insights and guidance throughout<br />
this project. Appreciation is also extended towards<br />
the North Carolina School of Science and Math for its<br />
investment into every one of its students.<br />
7. References<br />
Wang, S., Sun, S., Li, Z., Zhang, R., & Xu, J. (2017).<br />
Accurate de novo prediction of protein contact map by<br />
ultra-deep learning model. PLoS Computational Biology,<br />
13(1) doi:http://dx.doi.org/10.1371/journal.pcbi.1005324<br />
Ding, S., Zhao, H., Zhang, Y., Xu, X., & Nie, R. (2015).<br />
Extreme learning machine: Algorithm, theory and<br />
applications. The Artificial Intelligence Review, 44(1),<br />
103-115. doi:http://dx.doi.org/10.1007/s10462-013-<br />
9405-z<br />
74 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> MATHEMATICS AND COMPUTER SCIENCE
EFFECTS OF RELATIVITY ON QUADRUPOLE<br />
OSCILLATIONS OF COMPACT STARS<br />
Abhijit Gupta<br />
Abstract<br />
In the present age of space-based photometry, telescopes such as K2 and TESS are providing pulsation frequencies of<br />
stellar objects to unprecedented accuracy, requiring equally precise theoretical models correlating these observations to<br />
mass- and composition-dependent characteristics of stars. At this precision, relativistic models are required for compact<br />
objects such as white dwarfs and neutron stars. We model these stars as polytropes using the Tolman-Oppenheimer-<br />
Volkoff equation, and compute relativistic nonradial stellar pulsations around this equilibrium state. Outside the stellar<br />
surface, we integrate the Zerilli equation to locate resonant quasinormal modes, where ingoing gravitational radiation<br />
vanishes. We compare the frequencies of a subset of these modes to their corresponding pressure-modes in the Newtonian<br />
limit, as a function of the strength of relativity inside the star. Our results contribute to our understanding of the impact<br />
of general relativity on stellar oscillations, and can be used to determine the conditions under which the Newtonian<br />
approximation is justified.<br />
1. Motivation<br />
1.1 – Asteroseismology<br />
Although stars generally evolve on extremely long<br />
timescales, they are not static but pulsate periodically<br />
around an equilibrium. The frequencies of these<br />
oscillations inform us about internal characteristics of the<br />
star, suchv as mass, radius, pressure, and density. While<br />
these variables cannot be directly measured, telescopes can<br />
detect luminosity deviations that stellar pulsation cause.<br />
The frequencies of these oscillations are the frequencies of<br />
the stellar pulsations.<br />
Asteroseismology, the study of these stellar pulsations,<br />
involves two components: theoretical calculations<br />
and experimental observations. Theoretical programs<br />
assume a particular equilibrium state, and then model<br />
perturbations on this system. Only pulsations that satisfy<br />
boundary conditions at both the interior and stellar surface<br />
can possibly occur. Each pulsation can be described by a<br />
frequency and spherical harmonic degree and mode. The<br />
experimental observations measure periodic luminosity<br />
oscillations of stars over long periods of time. A Fourier<br />
transform is performed, and after filtering, spikes in<br />
the frequency curve are used to determine potential<br />
eigenfrequencies (Fig. 1).<br />
Figure 1. Experimental results from the K2 mission.<br />
The top panels show the standard and phasefolded<br />
light curves. The bottom panel shows the<br />
amplitude and residual spectrum after the pulsation<br />
frequencies are removed. Red vertical lines indicate<br />
observed pulsation frequencies (Bowman, D. M. et<br />
al., <strong>2018</strong>)<br />
Given these experimentally determined frequencies,<br />
programs can be run to determine the predicted central<br />
density, central pressure, total mass, radius, and many<br />
additional stellar variables. Asteroseismology presents an<br />
additional method to calculating these variables, alongside<br />
existing procedures. The combination yields stronger<br />
approximations than any method individually.<br />
1.2 – Compact Objects<br />
In recent years, space-based telescopes such as NASA’s<br />
Kepler spacecraft and Transiting Exoplanet Survey<br />
Satellite (TESS) are providing pulsation frequencies of<br />
stellar objects with unprecedented accuracy. Equally<br />
precise theoretical models correlating these observations<br />
to mass and composition-dependent characteristics of<br />
stars is required to make full use of these satellites. Present<br />
theoretical models have reduced error to less than 1 part<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 75
in 10 7 , roughly equivalent to the observational accuracy of<br />
these telescopes (Christensen- Dalsgaard & Mullan, 1993).<br />
However, when studying highly dense compact objects,<br />
general relativity can have a noticeable impact on the<br />
stellar structure and pulsations, requiring more rigorous<br />
models.<br />
While most stars are not substantially affected by general<br />
relativity, a class of compact objects require general<br />
relativistic corrections to accurately model the pulsations<br />
to the desired accuracy due to their extreme densities.<br />
Among compact objects, there are two main classes of<br />
stars: white dwarfs and neutron stars. White dwarfs are<br />
the remnants of low-mass to medium-mass stars that have<br />
exhausted their hydrogen and helium supplies. These<br />
stars are composed of heavier elements such as carbon<br />
and oxygen, and support themselves against gravitational<br />
collapse with electron degeneracy pressure. The density of<br />
a white dwarf is some 10 6 times greater than that of our<br />
Sun.<br />
Even more extreme are neutron stars, formed by the<br />
supernova explosions of stars not quite large enough to<br />
produce black holes. Neutron stars are similar to white<br />
dwarfs except are composed almost entirely of neutrons<br />
and supported with neutron degeneracy pressure instead<br />
of electron degeneracy. Physicists are still unsure exactly<br />
what types of matter are present at the very center of a<br />
neutron star, where density is the highest. Neutron stars<br />
are believed to be the densest macroscopic objects in the<br />
Universe, with densities about 10 15 times higher than that<br />
of the Sun.<br />
Relativistic corrections have small but noticeable impacts<br />
on white dwarfs, but are essential to study the pulsation<br />
frequencies of neutron stars. By better understanding the<br />
pulsations of neutron stars, we gain a better understanding<br />
of their interiors. Recent research even suggests that the<br />
matter in a neutron star may be the strongest material<br />
in the Universe, 10 billion times stronger than steel<br />
(Caplan, Schneider, & Horowitz, <strong>2018</strong>). Relativistic<br />
asteroseismology can assist in evaluating the different<br />
models attempting to describe the neutron star interior<br />
by providing accurate experimental data on neutron star<br />
properties.<br />
In this paper, we analyze how general relativity impacts<br />
the stellar pulsations of compact objects. By understanding<br />
when the Newtonian approximation is justified for a<br />
given error tolerance, we can improve the computational<br />
efficiency of theoretical asteroseismology without<br />
decreasing accuracy. On the other hand, computational<br />
improvements to previously published algorithms make<br />
our results potentially more accurate than existing<br />
results. Additionally, this research has applications to<br />
understanding the yet unknown physics governing the<br />
dense neutron star cores.<br />
2. Stellar Equilibrium<br />
The radius-dependent characteristics of compact objects<br />
affect their stellar pulsations, so an accurate model of the<br />
equilibrium state is required before computing stellar<br />
pulsation eigenfrequencies and other characteristics. To<br />
simplify calculations, a polytropic model is used in both<br />
the Newtonian and relativistic calculations (Knapp, 2011).<br />
A polytrope is a star where pressure (p) and density (ρ) are<br />
continuous with respect to radius, and are related by the<br />
equation of state:<br />
(1)<br />
where κ is the constant of proportionality, and n is the<br />
polytropic index. A polytropic index between 0.5 and 1<br />
generally models a neutron star well, while white dwarfs<br />
are modeled with a polytropic index of 3.<br />
2.1 – Newtonian Equilibrium<br />
In flat spacetime, the Lane-Emden equation describes<br />
the relationship between radius and density for polytropic<br />
stars, derived from the equation of hydrostatic equilibrium<br />
and the mass-continuity equation (Knapp, 2011)<br />
(2)<br />
where θ is defined by ρ = ρ c<br />
θ n , ρ c<br />
being the central density.<br />
ξ is the dimensionless radius defined by<br />
where G is the universal gravitation constant. The<br />
boundary conditions for this differential equation are θ(0)<br />
= 1 and θ′(0) = 0. For n = 0, n = 1, and n = 5, analytic<br />
solutions are available. For any other polytropic index,<br />
numerical integration to θ = 0 is required to analyze the<br />
equilibrium conditions of the star. Specifically, the Lane-<br />
Emden equation can be separated into two coupled firstorder<br />
ODEs using:<br />
(4) (5)<br />
Adaptive step-size fourth-order Runge-Kutta numerical<br />
integration is run on the system until the first step where θ<br />
< 0. Newton’s Method is then used to locate a more precise<br />
ξ where θ = 0. At this point, pressure and density become<br />
0, marking the outer edge of the star (Fig. 2).<br />
(3)<br />
76 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
are used when analyzing the stellar pulsations.<br />
2.3 – Comparison<br />
We compare the results of the Lane-Emden Equation<br />
and TOV Equation for a neutron star with typical<br />
characteristics. While the shapes of the curves are similar,<br />
there is a noticeable difference in radius and mass (integral<br />
of density with respect to radius) in Newtonian and<br />
relativistic spacetime (Fig. 3).<br />
Figure 2. θ vs. ξ for varying n. n = 0 has θ decline the<br />
fastest, while n = 5 decreases asymptotically but<br />
never reaches θ = 0. Neutron stars have n ≈ 1, and<br />
white dwarfs have n ≈ 3.<br />
2.2 – Relativistic Equilibrium<br />
While the Newtonian model is accurate in predicting<br />
the oscillation frequencies of main sequence stars, general<br />
relativity is needed to accurately describe compact objects<br />
with immense densities. The quantity σ approximates<br />
how relativistic a star is:<br />
(6)<br />
The greater σ, the greater the impacts of general relativity<br />
on both the equilibrium and stellar oscillations. For a white<br />
dwarf star, σ ≈ 0.001, while for a neutron star, σ ≈ 0.1.<br />
In this paper, we shall consider all stars in Schwarzschild<br />
spacetime, where spherical symmetry is assumed and there<br />
is no stellar rotation or magnetism involved (Hartle, 2003).<br />
The Schwarzschild metric tensor describes the spacetime:<br />
(7)<br />
where e −λ(r) = 1 − 2M (r)/r. e ν relates to the mass of the star,<br />
but cannot be analytically represented for a relativistic<br />
polytropic star. This metric tensor is given in geometric<br />
units (c = 1, G = 1) and in standard Schwarzschild<br />
coordinates (t, r, θ, Φ).<br />
A relativistic equivalent of the Lane-Emden equation,<br />
the Tolman-Oppenheimer-Volkoff (TOV) Equation, takes<br />
into account curved spacetime in describing polytropic<br />
stars. It calculates P, ρ, and ν as a function of radius. The<br />
TOV Equation can be written as three coupled first-order<br />
ODEs (Tooper, 1964).<br />
(8) (9) (10)<br />
With boundary conditions M(0) = 0, ν(R) = 1 − 2M/R,<br />
and p(0) = p 0<br />
, we solve this system very similarly to the<br />
Lane-Emden equation. The numerical results of the<br />
equilibrium analysis, radius-dependent p, ρ, ν, λ, and M,<br />
Figure 3. Comparison of solutions to Lane-Emden<br />
Equation and TOV Equation for neutron star with<br />
polytropic index n = 1. The TOV Equation predicts<br />
smaller radius and mass.<br />
The TOV Equation is used in all relativistic stellar<br />
pulsation calculations as the equilibrium model. Relativistic<br />
effects can be attributed both to differences in the<br />
equations governing stellar equilibrium, and differences in<br />
the equations governing stellar pulsations.<br />
3. Stellar Pulsations<br />
To analyze stellar pulsations, a perturbation is applied<br />
and propagated through the polytropic equilibrium state.<br />
Only under certain eigenfrequencies will the solution be<br />
continuous throughout the star. These oscillations can<br />
be both radial or nonradial, and each have a spherical<br />
harmonic degree l and mode m.<br />
Furthermore, the oscillations can be grouped into<br />
families of modes, depending on their restoring forces. The<br />
two most important classifications are Pressure Modes<br />
(p-modes) and Gravity Modes (g-modes). P-modes are<br />
high frequency modes whose deviations from equilibrium<br />
are counteracted by pressure changes in the convective<br />
zone. G-modes are low frequency modes, counteracted by<br />
mass movement in the radiative zone. In this research, we<br />
focus on p-modes, although our methods apply to g-modes<br />
as well.<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 77
For a specific spherical harmonic degree, spherical<br />
harmonic mode, and mode classification, there are<br />
multiple energy eigenmodes with ascending mode number<br />
k. The three variables, l, m, and k, along with the mode<br />
classification, fully describe a particular stellar pulsation.<br />
Multiple pulsations can occur simultaneously in a star,<br />
with resonant modes resulting from superposition (Fig. 4).<br />
perturbation variables y 1<br />
through y 4<br />
representing fractional<br />
changes in radius, pressure, gravitational potential, and<br />
gravitational acceleration. The solutions to the system are<br />
independent of all stellar equilibrium factors, except the<br />
polytropic index, allowing this dimensionless analysis. The<br />
differential equations for these variables can be written as<br />
one matrix equation (Unno, 1989).<br />
(13)<br />
The 1/x term in front of the matrix causes potential<br />
singularities in the integration, and also requires further<br />
emphasis closer to x = 0, where the system changes faster.<br />
To improve computational accuracy, we apply a change of<br />
variables from x to ln(x), yielding this simpler form:<br />
Figure 4. P-mode propagation for two harmonics.<br />
The number of reflections is the degree. The resonant<br />
modes result from a superposition of component<br />
waves travelling in opposite directions (Tosaka, n.d.)<br />
These stellar pulsations have separable time, angle, and<br />
radius dependence, given by:<br />
(11)<br />
(12)<br />
where f(t,r,θ,Φ) is a perturbation function, ω is the<br />
frequency, P lm<br />
(cosθ) is the associated Legendre polynomial,<br />
and N is a normalizing factor. By calculating f l<br />
(r), the radiusdependent<br />
perturbation for a specific eigenfrequency, the<br />
overall nature of the oscillations can be understood. While<br />
all degrees from 0 to ∞ could occur, in reality only the first<br />
few have substantial amplitude. l = 2 is the first degree<br />
at which gravitational radiation occurs in the relativistic<br />
model, making it the most optimal case study.<br />
4. Newtonian Quadrupole Oscillations<br />
4.1 – Pulsations Inside the Star<br />
In Newtonian spacetime, a set of 4 homogeneous firstorder<br />
differential equations describe the perturbations<br />
of radial displacement, pressure, gravitational potential,<br />
and gravitational acceleration. Physically, these relations<br />
are derived by maintaining continuous variables and<br />
appropriate boundary conditions.<br />
The system of differential equations originally is<br />
dimensioned, but can be made dimensionless, with<br />
(14)<br />
A * , U, V g<br />
, and c 1<br />
are dimensionless stellar equilibrium<br />
quantities as defined in Equations 15-18 below (Unno,<br />
1989). Although they contain ρ and p, all can be simplified<br />
to dimensionless form using ξ and θ. A * is the Eulerian<br />
pressure perturbation, c 1<br />
is an inverse scaled average<br />
density, and U and V g<br />
are common stellar variables.<br />
(15)<br />
(16)<br />
(17)<br />
(18)<br />
x is the dimensionless radius, ranging from 0 to 1. ω refers<br />
to the frequency of the oscillation being tested, and is made<br />
dimensionless by multiplying the dimensioned frequency<br />
by . ρ c<br />
and p c<br />
are the central density and pressure,<br />
respectively.<br />
The system of differential equations has central and<br />
surface boundary conditions, defined below. These<br />
conditions ensure the solution is physically acceptable at<br />
both boundaries (Unno, 1989).<br />
(19)<br />
(20)<br />
The differential equations are singular at both<br />
boundaries due to division by zero-valued variables. At<br />
the center of the star (x = 0), ln(x) is not defined, and<br />
at the outer surface of the star (x = 1), pressure is zero<br />
and V g<br />
and A * approach ∞. To handle this issue, we use<br />
the Magnus Multiple Shooting Scheme (Townsend &<br />
78 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
Teitler, 2013). Two arbitrary solutions satisfying the<br />
boundary conditions are created on both boundaries, and<br />
integrated to x = 0.5. They are inserted into a matrix, and<br />
the determinant is computed. Eigenfrequencies are found<br />
when the determinant of this square matrix is 0. Adaptive<br />
step-size fourth-order Runge-Kutta integration is used to<br />
integrate the system, and Newton’s Method is used during<br />
root-finding to locate where det(M) = 0 with quadratic<br />
convergence.<br />
4.2 – Algorithmic Roadmap<br />
In this section, we explain the specific steps taken to<br />
accurately compute the resonant modes of Newtonian<br />
polytropic stars. The code used to implement this<br />
algorithm was written in Python 3.<br />
1. The central pressure and density of the star are<br />
provided. The polytropic index n is given as well.<br />
From these, fourth-order Runge-Kutta integration is<br />
used on the Lane-Emden Equation (Eq. 2).<br />
2. For a test frequency and spherical harmonic degree,<br />
the perturbation variables are calculated at both<br />
boundaries using boundary conditions (Eq. 19-20).<br />
Two possible solutions on each end are integrated to<br />
r = 0.5R using fourth-order Runge-Kutta integration<br />
(Eq. 14). The equations are treated in matrix form for<br />
improved computational efficiency.<br />
3. Using the Magnus Multiple Shooting Scheme, the<br />
determinant of a 4x4 square matrix of partial solutions<br />
is calculated. Each row is a single integration<br />
from the previous step, and all 4 solutions are used. A<br />
determinant of 0 corresponds to an eigenfrequency.<br />
4. Steps 2 and 3 are repeated keeping the spherical<br />
harmonic degree constant and varying the test frequency.<br />
Newton’s Method is used to locate where<br />
det(M) = 0 with quadratic convergence. The derivative<br />
required for Newton’s Method is approximated by<br />
sampling 2 points slightly above and below the test<br />
frequency. Newton’s Method is run until a certain<br />
threshold accuracy is obtained.<br />
5. Steps 2 to 5 are repeated for each spherical harmonic<br />
degree. In this paper, results for l = 2 are shown,<br />
although others can be calculated with this algorithm.<br />
l = 2 is of particular importance because it accounts<br />
for the majority of gravitational radiation in the<br />
relativistic system.<br />
5. Newtonian Model Results and Discussion<br />
perturbation is the second-harmonic pressure-mode. We<br />
refer to the fundamental or lowest frequency mode in a<br />
family as the first-harmonic. l = 2 is chosen because it is the<br />
lowest spherical harmonic degree for which gravitational<br />
waves occur in the relativistic model.<br />
Figure 5 shows the results of this calculation. The<br />
four graphs left to right and top to bottom are radial<br />
perturbation, pressure perturbation, gravitational potential<br />
perturbation, and gravitational field perturbation, y 1<br />
to y 4<br />
in the above calculations. Radial displacement and pressure<br />
perturbations are largest near the center of the star, and all<br />
four perturbation variables approach zero near the surface<br />
of the star.<br />
Figure 5. Dimensionless Perturbations as a function<br />
of radius for l = 2 2 nd Harmonic Pressure-Mode for n=3<br />
Polytrope. x = 0 is center of star, x = 1 is stellar surface.<br />
While the perturbation dynamics are interesting, the<br />
eigenfrequency at which the pulsation occurs is generally<br />
more important, as it can be readily observed from Earth.<br />
The eigenfrequency for a Newtonian polytrope is solely a<br />
function of n among the equilibrium characteristics, and is<br />
also dependent on the spherical harmonic and particular<br />
mode.<br />
Prior calculations by Christensen-Dalsgaard and Mullan<br />
have yielded the first few p-mode eigenfrequencies for l =<br />
1, l = 2, and l = 3 to high precision (Christensen-Dalsgaard<br />
& Mullan, 1993). We compared the results of our method,<br />
described in Section 1.3, against these literature values.<br />
As a sample, Table 1 below shows a comparison of our<br />
calculations against theirs for the first 5 eigenfrequencies<br />
of a star with polytropic index n = 3 and spherical harmonic<br />
degree l = 2.<br />
We can visualize the normalized perturbations of a<br />
polytrope with index n = 3 as a function of dimensionless<br />
radius x. Although n = 3 best represents a white dwarf,<br />
a neutron star’s pulsations could be seen with n = 1. We<br />
use n = 3 for ease of comparison to prior calculations<br />
for main-sequence stars, also well approximated with n<br />
= 3. The spherical harmonic is l = 2, and this particular<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 79
Table 1. Dimensionless Frequencies of low harmonic<br />
pressure-modes (l = 2) for n = 3 Polytrope<br />
Harmonic Literature Calculated Rel. Error<br />
Fundamental 3.90687 3.90687 1.2491x10 -7<br />
2nd Harmonic<br />
3rd Harmonic<br />
4th Harmonic<br />
5th Harmonic<br />
5.169468 5.169469 7.6588x10 -8<br />
6.439991 6.439990 4.5185x10 -8<br />
7.708951 7.708951 1.8080x10 -10<br />
8.975891 8.975891 3.1879x10 -8<br />
With higher harmonics, the perturbation variables<br />
have a higher spatial frequency in the interior of the star,<br />
and have more zeroes and relative extrema. This makes<br />
numerically simulating these scenarios more complex, and<br />
less accurate than lower harmonics for equal number of<br />
integration steps. With increased integration steps, our<br />
model is sufficiently accurate even for higher harmonics.<br />
Table 2 uses the same polytropic equilibrium as Table 1,<br />
and the same spherical harmonic degree l = 2.<br />
Table 2. Dimensionless Frequencies of high harmonic<br />
pressure-modes (l = 2) for n = 3 Polytrope<br />
Harmonic Literature Calculated Rel. Error<br />
31st Harmonic<br />
41.5192 41.5221 6.8743x10 -5<br />
32nd Harmonic<br />
42.7630 42.7664 7.8022x10 -5<br />
33rd Harmonic<br />
44.0065 44.0104 8.7698x10 -5<br />
34th Harmonic<br />
45.2497 45.2541 9.7599x10 -5<br />
35th Harmonic<br />
46.4927 46.4977 1.0758x10 -4<br />
Although the error in the higher harmonics is<br />
approximately 100 times larger in magnitude than the<br />
error in the lower harmonics, it is still 1 part in 10,000<br />
or less. Given the strong match for both low and high<br />
eigenfrequencies, this code can be used to calculate<br />
frequencies for higher harmonics than previously<br />
reported ((Christensen-Dalsgaard & Mullan, 1993) goes<br />
to 50 th ). However, these higher harmonics require greater<br />
energy, and thus occur at smaller amplitudes in real<br />
compact objects. Their study is useful for understanding<br />
patterns in stellar pulsations, but not for experimental<br />
asteroseismology.<br />
6. Relativistic Quadrupole Oscillations<br />
6.1 – Perturbation Metric<br />
Similar to the Newtonian case, we use a polytropic model<br />
of the equilibrium structure. A perturbation is applied, and<br />
as a result of the motion, the geometry of spacetime around<br />
the relativistic star is no longer described by Equation (7).<br />
Rather, the new metric, involving the perturbation metric<br />
h uv<br />
, becomes<br />
(21)<br />
In even-parity Regge-Wheeler gauge, the perturbation<br />
metric takes the form (Thorne & Campolattaro, 1967):<br />
(22)<br />
The variable μ is the dimensionless radius of the star<br />
(ranging from 0 to 1), Y = e iωt * Y lm<br />
is the time dependence<br />
multiplied by the spherical harmonic of the perturbation.<br />
H 0<br />
, H 1<br />
, and K are functions of r only. The Regge-Wheeler<br />
gauge is preferred for only introducing two terms outside<br />
the main diagonal. Substituting into Equation (21), we<br />
obtain:<br />
(23)<br />
6.2 – Perturbations Inside the Compact Object<br />
Inside the star, the perturbed fluid is described by a<br />
displacement ξ α , where:<br />
(24) (25)<br />
(26)<br />
The three fluid perturbations have separable timeand<br />
radius-dependence, allowing for calculations done<br />
at a specified time to represent the system with the<br />
necessary transformations. The variables W and V are<br />
fluid perturbation variables that must be solved for to<br />
describe the nonradial stellar pulsations. Five variables<br />
are dependent on radius, H 0<br />
, H 1<br />
, K, W, and V. The first<br />
three relate to the initial spacetime perturbation, and W<br />
and V describe fluid perturbations (Lindblom & Detweiler,<br />
1983).<br />
Einstein’s Field Equations can be applied to the<br />
spacetime metric given in Equation (23) to give differential<br />
equations for each perturbation variable. Using these<br />
relations, we can eliminate one variable, creating a system<br />
of four differential equations. Following Detweiler and<br />
Lindblom, H 1<br />
is eliminated instead of H 0<br />
, to avoid possible<br />
singularities (Lindblom & Detweiler, 1985). To simplify<br />
the resultant equations, X is defined as a function of W,<br />
V, and H 0<br />
:<br />
80 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
(27)<br />
The four first-order differential equations for H 1<br />
, K, W,<br />
and X, are (Lindblom & Detweiler, 1985):<br />
(28)<br />
(29)<br />
(30)<br />
in the interior of the star. We are mainly interested in<br />
solutions composed only of outgoing waves, as these<br />
represent resonant oscillation and the energy radiated<br />
from the star. These frequencies are called Quasi-Normal<br />
Modes (QNMs), and include the relativistic equivalent of<br />
Newtonian p-modes.<br />
To find these specific eigenfrequencies, we analyze<br />
the perturbation variables outside the compact object to<br />
determine the gravitational radiation produced. In the<br />
exterior of the star, the fluid perturbations W, V, and X<br />
are zero and the 2 metric perturbations H 1<br />
and K can be<br />
combined to obtain the single second-order differential<br />
equation known as the Zerilli equation (N. Andersson &<br />
Shutz, 1995).<br />
with the effective potential V Z<br />
given by:<br />
(32)<br />
(31)<br />
Equations (28) to (31) can be expressed in matrix form<br />
similar to Equation (14), and are handled computationally<br />
in this manner. A major difference between the<br />
Newtonian and relativistic calculations is that the<br />
relativistic calculations are dimensioned while Newtonian<br />
is dimensionless. Only 2 of the 4 linearly independent<br />
solutions to this system are well-behaved at the center of<br />
the star (at r = 0). The perturbed pressure must vanish at r<br />
= R, so X(R) = 0. From these conditions, a single acceptable<br />
solution is specified for each frequency ω.<br />
At the central boundary, r = 0, the differential equations<br />
are singular, as they contain multiple 1/r terms that tend<br />
the function to infinity. Since the numerical integration<br />
cannot be started at r = 0, a power-series approximation<br />
is used to determine an appropriate starting condition<br />
slightly away from the center, following the procedure<br />
described in (Lindblom & Detweiler, 1983) and (Lindblom<br />
& Detweiler, 1985).<br />
The power series approximations are used to r = 0.01R.<br />
Then, the differential equations are integrated using<br />
fourth-order Runge-Kutta integration to r = 0.5R. There<br />
are two linearly independent solutions, labelled Y 1<br />
and Y 2<br />
.<br />
Similarly, the three solutions from the exterior of the star<br />
are iterated to the midpoint of the interval, giving Y 3<br />
, Y 4<br />
,<br />
and Y 5<br />
. A linear combination of these five solutions exists<br />
that makes each variable H 1<br />
, K, W, and X continuous at the<br />
midpoint. With five solutions for four variables, there is<br />
an extra degree of freedom. This additional degree allows<br />
for free-scaling of the solution.<br />
6.3 – Perturbations Outside the Compact Object<br />
Given any spherical harmonic degree and frequency,<br />
we can find the unique solution for the radial dependent<br />
variables H 1<br />
, K, W, and X that define the perturbations<br />
The tortoise coordinate r * is defined by:<br />
(33)<br />
(34)<br />
The Zerilli equation is notable because it provides<br />
a Schrödinger-type equation for even-parity Regge-<br />
Wheeler perturbations of Schwarzschild geometry. This<br />
presents simplifications to the analysis of wave equations<br />
(Fackerell, 1971). The Zerilli function is defined in terms<br />
of the perturbations H 0<br />
(r) and K(r)<br />
(35)<br />
where the functions a(r), b(r), g(r), h(r), and k(r) are<br />
functions of the frequency, spherical harmonic degree, and<br />
mass and radius of the compact object, given in (Lindblom<br />
& Detweiler, 1983). We recover H 0<br />
with the following<br />
equation, similar to the relation defined between V and X<br />
in Equation (27).<br />
(36)<br />
Using Equation (35), we obtain initial conditions for Z (r * )<br />
and dZ (r * ) /dr * . For a given (r * , Z) coordinate, Equation<br />
(32) can be used to calculate d 2 Z/dr *2<br />
, and in this manner<br />
we propagate Z through r * . In practice, we integrate Z from<br />
r * = R * to r * = 25ω −1 (Lindblom & Detweiler, 1983). Far<br />
away from the star, the Zerilli function can be expressed as<br />
a combination of 2 components, namely the ingoing and<br />
outgoing contributions. These individual solutions may be<br />
asymptotically expressed as power series.<br />
(37) (38)<br />
The solution Z − represents purely outgoing gravitational<br />
radiation, while Z + represents purely ingoing waves. The<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 81
constants a j<br />
and the complex conjugates ā j<br />
are recursively<br />
defined in (Chandrasekhar & Detweiler, 1975). A solution<br />
to the Zerilli equation will be given by a constant linear<br />
combination of Z + and Z − .<br />
(39)<br />
For Quasi-Normal Modes, all the gravitational radiation<br />
is outgoing, so the particular solution Z should be a multiple<br />
of Z − , with no parts Z + . At r = 25ω −1 , the Zerilli equation<br />
numerically integrated is matched onto the asymptotic<br />
series, with j max<br />
= 2. We determine values of constants<br />
β(ω) and γ(ω), and use Newton’s method to search for ω<br />
such that γ(ω) = 0. The eigenfrequencies found are those<br />
of quasinormal modes, a subset of which correspond to the<br />
Newtonian Pressure-Modes.<br />
(40)<br />
We compare the difference in frequencies of these<br />
corresponding modes Equation (40) against the relativity<br />
parameter σ defined in Equation (6) to understand the<br />
effects of general relativity on pulsation frequencies of<br />
compact objects, the ultimate goal of this research.<br />
6.4 – Algorithmic Roadmap<br />
A similar approach is taken here compared to the<br />
Newtonian model (See Section 4.2). However, there are<br />
some key differences in the implementation. Instead<br />
of using the Lane-Emden Equation, the Tolman-<br />
Oppenheimer Equation is used (Eq. 8-10). After<br />
integrating the Equations (28)-(31) in the interior of the<br />
star, the Zerilli function and its derivatives are computed<br />
at r = R (Eq. 32-36). Runge-Kutta integration is used to<br />
iterate the Zerilli function far from the star where it is<br />
matched onto the asymptotic power series expansions and<br />
the coefficients β(ω) and γ(ω) are calculated (Eq. 39). γ(ω)<br />
replaces det(M) in the Newtonian model, and we proceed<br />
as before locating eigenfrequencies for various spherical<br />
harmonics.<br />
7. Relativistic Model Results and Discussion<br />
We calculate the normalized perturbations of a<br />
polytrope with n = 3 as a function of dimensionless radius<br />
r. Although n = 1 is most optimal for a neutron star, we<br />
use n = 3 initially to best compare to the Newtonian model.<br />
The spherical harmonic degree is l = 2, and this particular<br />
perturbation is the second harmonic pressure mode.<br />
Figure 6 shows the results of this calculation. The four<br />
graphs left to right and top to bottom are perturbation<br />
variables X, W, K, and X 0<br />
. Recall K and X 0<br />
represent metric<br />
perturbations (Eq. 23). The shape of these curves closely<br />
match the shapes of y 3<br />
and y 4<br />
in the Newtonian section<br />
(Fig. 5). X and W are different variables than y 1<br />
and y 2<br />
,<br />
explaining the differences in the shapes of the top two<br />
panels between Figure 5 and 6.<br />
Figure 6. Perturbation variables calculated for<br />
a specific pulsation, with corresponding Zerilli<br />
variable integrated outside the compact object using<br />
the Zerilli equation.<br />
These results show strong qualitative similarities to<br />
our previous Newtonian results, indicating the relativistic<br />
model is successful in predicting the general behavior of<br />
the interior perturbation variables. The single discernible<br />
frequency and sinusoidal shape of the Zerilli function<br />
indicate these methods can locate quasinormal modes<br />
fairly accurately. Further research is ongoing to search<br />
for exact quasinormal eigenfrequencies. Until then, we<br />
cannot numerically comapre quantitative results between<br />
the Newtonian and relativistic models. Nonetheless,<br />
our model successfully replicates the Newtonian model<br />
behavior within curved spacetime as well.<br />
8. Acknowledgments<br />
I would like to thank Mr. Reece Boston (UNC-Chapel<br />
Hill), Dr. Charles Evans (UNC-Chapel Hill), and Dr.<br />
Jonathan Bennett (NCSSM) for their continued support<br />
and guidance throughout this research project.<br />
9. References<br />
Bowman, D. M., Buysschaert, B., Neiner, C., P ́apics, P.<br />
I., Oksala, M. E., & Aerts, C. (<strong>2018</strong>). K2 space photometry<br />
reveals rotational modulation and stellar pulsations in<br />
chemically peculiar a and b stars. A&A, 616, A77. Retrieved<br />
from https://doi.org/10.1051/0004-6361/<strong>2018</strong>33037 doi:<br />
10.1051/0004-6361/<strong>2018</strong>33037<br />
Caplan, M. E., Schneider, A. S., & Horowitz, C. J.<br />
(<strong>2018</strong>, Sep). Elasticity of nuclear pasta. Phys. Rev. Lett.,<br />
121, 132701. Retrieved from https://link.aps.org/<br />
doi/10.1103/PhysRevLett.121.132701 doi: 10.1103/<br />
PhysRevLett.121.132701<br />
Chandrasekhar, S., & Detweiler, S. (1975). The quasinormal<br />
modes of the schwarzschild black hole. The Royal<br />
Society.<br />
82 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
Christensen-Dalsgaard, J., & Mullan, D. J. (1993). Accurate<br />
frequencies of polytropic models. Royal Astronomical<br />
Society.<br />
Fackerell, E. D. (1971). Solutions of zerilli’s equation for<br />
even-parity gravitational perturbations. The Astrophysical<br />
Journal.<br />
Hartle, J. B. (2003). Gravity: An introduction to einstein’s<br />
general relativity (1st ed.). San Francisco: Addison-Wesley.<br />
Knapp, J. (2011). Polytropes.<br />
Lindblom, L., & Detweiler, S. L. (1983). The quadrupole<br />
oscillations of neutron stars. The Astrophysical Journal.<br />
Lindblom, L., & Detweiler, S. L. (1985). On the nonradial<br />
pulsations of general relativistic stellar models. The<br />
Astrophysical Journal.<br />
N. Andersson, K. D. K., & Shutz, B. F. (1995). A new<br />
numerical approach to the oscillation modes of relativistic<br />
stars.<br />
Thorne, K. S., & Campolattaro, A. (1967, Sep). Nonradial<br />
pulsation of general-relativistic stellar models. i.<br />
analytic analysis for l ≥ 2. The Astrophysical Journal,<br />
149, 591. Retrieved from http://adsabs.harvard.edu/<br />
abs/1967ApJ...149..591T doi: 10.1086/149288<br />
Tooper, R. F. (1964). General relativistic polytropic fluid<br />
spheres. The Astrophysical Journal, 140(434). Tosaka, W.<br />
C. C.-B.-S.-. G. (n.d.).<br />
Townsend, R., & Teitler, S. (2013). Gyre: An open-source<br />
stellar oscillation code based on a new magnus multiple<br />
shooting scheme.<br />
Unno, W. (1989). Nonradial oscillations of stars (2nd ed.).<br />
Tokyo: University of Tokyo Press.<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 83
EFFECT OF ELLIPTIC FLOW FLUCTUATIONS ON THE<br />
TWO- AND FOUR-PARTICLE AZIMUTHAL CUMULANT<br />
Brian Lin<br />
Abstract<br />
We incorporate finite elliptic flow fluctuations for the 2-particle and 4-particle azimuthal cumulants. Starting from<br />
expressions that include transverse momentum conservation, we consider three potential v 2<br />
distributions: a Gaussian<br />
distribution, a Bessel-Gaussian distribution, and a power law distribution. For the Bessel-Gaussian distribution, we find<br />
the results are sensitive to the size of fluctuations, and c 2<br />
{4} values at large multiplicity range from 0 to significantly<br />
negative. Therefore, the 4-particle cumulant c 2<br />
{4} with transverse momentum conservation can be used to study elliptic<br />
flow fluctuations in both small and large systems.<br />
1. Introduction<br />
In the Pb+Pb and p+Pb collisions in heavy ion colliders,<br />
evidence indicates a nearly perfect fluid is produced in this<br />
system of quarks and gluons. The collective flow phenomenon<br />
that arises from these collisions is predicted very well by<br />
the use of hydrodynamics. Caused by the collision’s initial<br />
geometric anisotropies, we observe azimuthal anistropy in<br />
the produced particles, which is the clearest indicator of the<br />
collective flow phenomenon (Nagle and Zajc, <strong>2018</strong>).<br />
The study of relativistic heavy ion collisions originates<br />
from the desire to learn more about the basic origins of matter<br />
and, in particular, a new form of QCD matter: the Quark-<br />
Gluon Plasma (QGP). Understanding of the QGP will reveal<br />
fundamental properties of matter in high-temperature and<br />
high-density systems, such as systems existing in the core of<br />
neutron stars and theorized to have existed in the early stages<br />
of the Big Bang (Jacak and Steinberg, 2010).<br />
Elliptic flow is an essential observable that can reveal<br />
the equation of state of the QGP among other important<br />
characteristics of dense matter, so accurate measurement of<br />
elliptic flow has impactful theoretical implications (Snellings,<br />
2011). However, momentum conservation and jet quenching<br />
add non-flow effects to measurements of elliptic flow, so our<br />
measured elliptic flow values contain non-flow effects. Thus,<br />
the field of relativistic heavy ion collisions utilizes cumulants,<br />
which suppress the effects of non-flow factors and emphasize<br />
the true effects of collective flow. Elliptic flow reflects the<br />
initial geometric anisotropies of the overlapping nuclei. Even<br />
at the same centrality, the elliptic flow for each event differs<br />
due to fluctuations. Therefore, we need to consider effects<br />
of fluctuation on multi-particle cumulants (Bilandzic et al.,<br />
2011).<br />
The clearest way to remove non-flow effects from elliptic<br />
flow coefficients is to analyze the azimuthal cumulants<br />
associated with the collisions. However, even these azimuthal<br />
anisotropies differ as the overlap between the two nuclei<br />
varies. Thus, the measured elliptic flow coefficient is expected<br />
to follow a probability distribution. This paper calculates the<br />
effect of elliptic flow distributions on two-particle and fourparticle<br />
cumulants, which have been calculated assuming<br />
global transverse momentum conservation.<br />
We calculate new expressions for c 2<br />
{2} and c 2<br />
{4} by<br />
incorporating three predicted distributions of elliptic flow<br />
that originate from geometric anisotropies between events.<br />
We primarily analyze the the effect of the distribution<br />
characteristics on the values of the azimuthal cumulants.<br />
2. Methods<br />
The two- and four-particle azimuthal cumulants are<br />
functions of elliptic flow, v 2<br />
, so we determine new expressions<br />
for c 2<br />
{k} by incorporating the v 2<br />
fluctuations as probability<br />
density distributions, P(v 2<br />
). The single-event average 2-particle<br />
and 4-particle azimuthal correlations are defined as follows,<br />
where the brackets represent averaging over all particles in<br />
the event:<br />
The average 2-particle and 4-particle azimuthal correlations<br />
over many events may be written as follows, where we<br />
denote two averages, first over all particles in an event and<br />
then over all events:<br />
The single-event average 2-particle and 4-particle<br />
azimuthal cumulants are defined as:<br />
Due to fluctuations in the average azimuthal cumulants<br />
between events, we calculate the event-averaged 2-particle<br />
and 4-particle cumulant values, denoted and<br />
respectively, by incorporating a probability distribution<br />
on v 2<br />
as follows:<br />
84 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
3. Results<br />
3.1 – Gaussian Formulas<br />
We begin by incorporating a Gaussian distribution of<br />
v 2<br />
, as follows:<br />
We now introduce the formulas derived earlier<br />
(Bzdak and Ma, <strong>2018</strong>) using the assumption of transverse<br />
momentum conservation (TMC). This assumption has a<br />
key contribution for small systems because the last particle’s<br />
momentum is restricted. The effect of TMC diminishes as<br />
the number of particles, N, increases because the effect of<br />
one particle’s momentum also decreases with N.<br />
2.1 – TMC Formulas<br />
For the Gaussian distribution, we have:<br />
3.2 – Bessel-Gaussian Formulas<br />
The Bessel-Gaussian distribution that we give v 2<br />
is<br />
defined as:<br />
where I n<br />
(x) denotes the Bessel function of the n-th kind.<br />
For the Bessel-Gaussian distribution, we have:<br />
The original formulas (Bzdak and Ma, <strong>2018</strong>) contained<br />
the variable v 2<br />
(p), which we denote for brevity v 2<br />
. In<br />
both instances, this represents the elliptic flow value at a<br />
specific momentum p. We see the 2-particle and 4-particle<br />
cumulant as a function of other variables such as transverse<br />
momentum p, number of produced particles N, and the<br />
expected value of the square of transverse momentum<br />
over the full phase space . We define<br />
3.3 – Power Law Formulas<br />
We continue by incorporating the power law<br />
distribution of v 2<br />
, given as:<br />
For the Power-Law distribution, we have:<br />
2.2 – General Distribution<br />
We denote<br />
We plot the three distributions under the condition that<br />
= 0.05.<br />
When we incorporate the probability distribution P(v 2<br />
)<br />
we integrate in terms of v 2<br />
. Thus, for a general distribution,<br />
we may express<br />
Figure 1. Plotted above are sample probability<br />
distributions where = 0.05.<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 85
The probability distributions are plotted so that the<br />
expected value of elliptic flow remains 0.05. We define w =<br />
for the Bessel-Gaussian distribution. When w = 0,<br />
the Bessel-Gaussian distribution reduces to the Gaussian<br />
distribution. We see the Gaussian, Bessel-Gaussian with w<br />
= 0, and Power Law curves are all very similar, while the<br />
Bessel-Gaussian curve with w = 2 differs from the other<br />
three curves.<br />
3.4 – An Example of Numerical Comparisons<br />
Our graphs take a reasonable value for as 0.05.<br />
Additionally, we assume = 0.025 and = 0.25<br />
(GeV/c) 2 . This allows us to solve for the unknown σ in the<br />
Gaussian distribution, σ and in the Bessel-Gaussian<br />
distribution, and α in the Power Law distribution. The<br />
Gaussian σ value turns out to be around 0.0564, while α<br />
≈ 313.<br />
For this section, we tune the Bessel-Gaussian mean<br />
and width to that of the Gaussian (i.e. we control both<br />
and ). Our solutions are ( , σ) = (0, 5.64 × 10 −2 )<br />
when we equate the variances of the Bessel-Gaussian and<br />
Gaussian distributions, and ( , σ) = (1.57 × 10 −2 , 5.42 ×<br />
10 −2 ) when we equate the variances of the Bessel-Gaussian<br />
and Power Law distributions.<br />
The Gaussian and Bessel-Gaussian with w=0<br />
distributions are identical, and the power law distribution<br />
is very similar to them. Thus, these three distributions<br />
result in an upward shift of the 2-particle cumulant.<br />
Because of the similarity between all three distibutions, all<br />
three lead to essentially the same 2-particle cumulant (fig.<br />
2).<br />
The event-averaged 4-particle cumulant approaches<br />
for large N. Specifically for the<br />
Gaussian distribution, from section 3.1, the Gaussian<br />
approaches 0 regardless of σ. Because we set the<br />
variance of the Bessel-Gaussian distribution equal to<br />
that of the Gaussian distribution, and is similar to that<br />
of the power law distribution, the behaviors of all three<br />
distributions are very similar. More general features of the<br />
Bessel-Gaussian will be shown in section 3.5.<br />
Figure 2. The event-averaged 2-particle cumulant<br />
(top panel) and 4-particle cumulant<br />
(bottom panel) for all three distributions (Gaussian,<br />
Bessel-Gaussian, and Power Law distributions)<br />
plotted as a function of event multiplicity N. The<br />
results obtained without elliptic flow fluctuations<br />
are shown in comparison in black. Additionally,<br />
experimental results obtained from the ATLAS<br />
collider are shown in green.<br />
3.5 – Effects of Relative Fluctuation Size<br />
The Bessel-Gaussian probability distribution, defined<br />
in Section 3.2, may be rewritten in terms of two instead<br />
of three variables. Denoting u = /σ and w = /σ, we<br />
may rewrite the Bessel-Gaussian distribution in terms of<br />
u and w as:<br />
We define the relative fluctuation of v 2<br />
as:<br />
We can show that r(v 2<br />
) = r(u) is a function of w only, and<br />
so to manipulate the Bessel-Gaussian distribution, we only<br />
need to manipulate w. Specifically, when w = 0, the relative<br />
fluctuation of v 2<br />
reaches a maximum of .<br />
In the large event multiplicity limit, the 4-particle<br />
Bessel-Gaussian cumulant is given by:<br />
while the cumulant neglecting elliptic flow fluctuation is<br />
86 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
. Both these values depend on two parameters:<br />
and σ. However, their ratio, which approaches<br />
, is dependent only on w. The ratio of<br />
these two values is shown as the dashed curve, and the<br />
relative fluctuation of v 2<br />
is shown as the solid curve (fig. 3).<br />
When we equate the variance of the Bessel-Gaussian<br />
distribution to that of the Gaussian distribution, we<br />
obtain w = 0, and the maximum relative fluctuation (fig.<br />
3). The corresponding large-N limit of for w = 0<br />
is 0. On the other hand, for large w, relative fluctuations<br />
are small (fig. 3). In that limit, the Bessel-Gaussian eventaveraged<br />
4-particle cumulant will approach the results<br />
obtained through TMC that neglected v 2<br />
fluctuation, i.e.<br />
at large N, and so the relative<br />
fluctuation for small w approaches 1.<br />
Because both curves (fig. 3) are solely functions of w,<br />
the ratio<br />
at large N may be expressed as a<br />
function of the v 2<br />
relative fluctuation only. The 4-particle<br />
cumulant ratio against r(v 2<br />
) is shown as the dashed curve<br />
in Figure 4. As the relative fluctuation increases, we see<br />
the ratio decrease from 1 to 0, i.e. the goes from<br />
significantly negative to 0 for large N.<br />
For the 2-particle cumulant, we observe its large event<br />
multiplicity limit to be . Additionally, the 2-particle<br />
cumulant neglecting elliptic flow fluctuation approaches<br />
, so at large N, the ratio between the two may be<br />
written as<br />
Figure 4. The ratio between Bessel-Gaussian and<br />
non-fluctuation cumulants at large N for both c 2<br />
{2}<br />
and c 2<br />
{4}.<br />
Therefore, we see as the relative fluctuation increases,<br />
the ratio increases from 1 to 4/π.<br />
To visualize our results, we plot in Figure 5 the Bessel-<br />
Gaussian 2-particle and 4-particle cumulants for varying<br />
relative fluctuations of v 2<br />
under the condition that<br />
= 0.05. These relative fluctuations were chosen so that<br />
r(v 2<br />
) = 0.523, 0.466, 0.319, and 0 correspond with w values<br />
of 0, 1, 2, and ∞ respectively.<br />
Figure 3. Relative fluctuation of v 2<br />
for the Bessel-<br />
Gaussian distribution (solid) and the ratio between<br />
the large-N limits of the Bessel-Gaussian and nonfluctuation<br />
cumulants (dashed) as functions of w.<br />
Figure 5. The event-averaged (first) and<br />
(second) for the Bessel-Gaussian distribution<br />
as a function of event multiplicity N. Results with<br />
various amounts of relative fluctuations of v 2<br />
are<br />
shown. Again, ATLAS results for the 4-particle<br />
cumulant are shown in green.<br />
PHYSICS<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 87
4. Conclusions<br />
When elliptic flow fluctuations are included in the<br />
calculations of two and four-particle azimuthal cumulants,<br />
there is a definite shift in the cumulant. For the two-particle<br />
cumulant, there is an increase for the Gaussian, Bessel-<br />
Gaussian, and power law distributions of v 2<br />
fluctuations.<br />
Meanwhile, for the four-particle cumulant, we observe a<br />
large positive shift so that its value is close to 0 for large<br />
event multiplicity when we incorporate Gaussian or<br />
Power Law elliptic flow distributions. The Bessel-Gaussian<br />
distribution allows for variation of the relative fluctuation<br />
of v 2<br />
. When the relative fluctuation is small, the 4-particle<br />
cumulant tends towards a significantly negative value at<br />
large event multiplicity, approaching results obtained<br />
previously without including v 2<br />
fluctuations. When the<br />
relative fluctuation is large, the cumulant goes to zero<br />
at large event multiplicity, approaching results from<br />
the Gaussian and power law elliptic flow distributions.<br />
Therefore, the c 2<br />
{4} observable may be used to probe the<br />
fluctuation of elliptic flow in both small and large systems.<br />
5. Acknowledgements<br />
Firstly, we would like to acknowledge Dr. G.L. Ma of<br />
Fudan University for his patient, insightful mentorship.<br />
We would like to gratefully thank Dr. Z.W. Lin for<br />
valuable discussions and feedback. We also acknowledge<br />
Dr. J. Bennett for engaging in weekly discussions and<br />
offering advice as well as the NCSSM Foundation for<br />
providing the necessary support and resources to carry out<br />
his research.<br />
6. References<br />
Nagle, J. L., & Zajc, W. A. (<strong>2018</strong>). Small System Collectivity<br />
in Relativistic Hadronic and Nuclear Collisions. Annual<br />
Review of Nuclear and Particle Science, 68 (1), 211-235.<br />
Snellings, R. (2011). Elliptic flow: A brief review. New Journal<br />
of Physics, 13(5), 055008.<br />
Jacak, B., & Steinberg, P. (2010). Creating the perfect liquid<br />
in heavy-ion collisions. Physics Today, 63(5), 39-43.<br />
Bzdak, A., & Ma, G. (<strong>2018</strong>). A remark on the sign change<br />
of the four-particle azimuthal cumulant in small systems.<br />
Physics Letters B, 781, 117-121.<br />
Bilandzic, A., Snellings, R., & Voloshin, S. (2011). Flow<br />
analysis with cumulants: Direct calculations. Physical Review<br />
C, 83(4).<br />
88 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> PHYSICS
AN INTERVIEW WITH DR. VALERIE ASHBY<br />
From left, Navami Jain, BSS Editor-In-Chief; Emily Wang, BSS Editor-In-Chief; Dr. Jonathan Bennett, BSS Faculty<br />
Advisor; Dr. Valerie Ashby, Dean of Trinity College of Arts & Sciences at Duke University; Kathleen Hablutzel, Publication<br />
Editor-In-Chief; and Jackson Meade, BSS Essay Contest Winner<br />
What drew you to chemistry?<br />
That’s an easy answer. My dad was a math and science<br />
teacher; he taught chemistry and various versions of math<br />
in high school… so science was never scary to me. It just<br />
seemed like what we did… The second thing is I had a great<br />
high school chemistry teacher. I actually did something I<br />
don’t recommend to my own Duke students, which is to<br />
decide what you’re going to major in before you arrive.<br />
Leaving high school, I said that I’m going to be a chemistry<br />
major and I’m not going to change my major, because I had<br />
heard these stories about how college students hit their<br />
first hard course or their second hard course and they shift<br />
their major. I decided I was not going to do that. The good<br />
news was that I loved it, even at the college level… that’s<br />
how I decided I was going to major in chemistry. Science<br />
was always my thing.<br />
So you’re in more of an administrative position now. Do<br />
you ever wish you could go back to the lab?<br />
Oh, you mean every 30 seconds? I wish for you that every<br />
job you have is your favorite job. And I have led this crazy,<br />
lovely life where every single job that I have held has been<br />
my favorite job at that moment. When I was a faculty<br />
member, it was my favorite job. Who I am and what I do<br />
have overlapped my entire life. That’s a gift that I get to<br />
be who I am in my job. I am a teacher, that's who I am.<br />
Even though I’m out of the classroom, that’s still who I am.<br />
The way that it presents itself now is through inspiring<br />
other teachers, encouraging other faculty, and mentoring<br />
students...I have office hours with students every Friday<br />
even though I’m not teaching. They come and talk to me<br />
about their lives and I get to do the thing that I love…<br />
I also miss running my old research group. I kept my<br />
research group at UNC when I took this job… I graduated<br />
my last PhD students from UNC Chapel Hill last year.<br />
For the first time in twenty years I haven’t had my own<br />
research group. I’m so busy that I don’t have time, but<br />
I miss training graduate students and I miss creating<br />
knowledge. There’s something about waking up every<br />
day trying to do something that nobody else has ever done<br />
and answering a question that remains open, and then<br />
teaching other people how to do that… it is so much fun.<br />
FEATURED ARTICLE<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 89
We were wondering how the scientific and problemsolving<br />
skills you’ve gained as a chemist have translated<br />
into other roles such as your current role?<br />
It was absolutely great training. When you do scientific<br />
research, it is team-based with vertically integrated teams;<br />
so a professor, a postdoc, graduate students, undergrad<br />
students, and then high school students who come in the<br />
summer or during the academic year. That team-based<br />
approach and learning how to work with every level of<br />
that team are great training for what I do here.<br />
I have an administrative team and it’s a vertically integrated<br />
team… When you run a research group, you’re not just<br />
doing science, you’re doing people - people who spend a<br />
lot of time together in close proximity. Teaching graduate<br />
students how to navigate being in a group that has a<br />
personality and a culture… I had to manage all the finances<br />
of the group, so I learned how to do big budgets for grants.<br />
You learn how to write, you learn how to communicate<br />
- so many different parts of running a research team. It’s<br />
like a small business if you're doing science... So what<br />
do I do in my present job? I run the finances - they’re<br />
my responsibility. Human resources, the well-being of<br />
students, faculty, and staff, making sure that we’re being<br />
collaborative and collegial - all my responsibility. It’s<br />
absolutely great training and I think I use all of that now.<br />
My day-to-day life is really all of those skills that you learn<br />
about being in a team and managing people.<br />
And my job is to raise money. If you’re going to do science,<br />
you better know how to raise money. You may know<br />
who Joe DeSimone is - I was his first PhD student so we<br />
have known each other for a very long time and one of<br />
my favorite Joe quotes is “Val, a vision without funding<br />
is just a hallucination.” And as a scientist, if that’s not<br />
your mindset, you can’t actually do your science. This<br />
enterprise doesn’t run without funding, so being a little bit<br />
entrepreneurial is important... for this job.<br />
While at UNC you worked with an NSF grant to<br />
increase the number of underrepresented minority<br />
students who receive doctoral degrees in STEM fields.<br />
What were some of your more effective policies and<br />
what challenges have you personally faced as a minority<br />
woman in STEM?<br />
Quite frankly, I never paid any attention to being a woman<br />
or being underrepresented. Now, that’s a luxury. People<br />
treated me so well it was never my experience. Now when<br />
I advocate for women and underrepresented people I have<br />
to say to them “I haven’t had a bad experience. My goal<br />
is for you not to. And if you have, my goal is to help you<br />
with it.” My PhD advisor was incredible - some people<br />
have trouble with that. The reason I want to help so many<br />
people is because I have had such a wonderful experience.<br />
I always say to people if somebody tried to offend me at<br />
some point or did something, I just didn’t take it in… it just<br />
never affected me. So that’s my history with that.<br />
I loved working in that program and it had a model that<br />
worked already and my job was to not break it and to<br />
try to expand it. It’s a cohort model of students and it’s<br />
everything from making sure students are onboarded into<br />
their departments. It’s very isolating to be a grad student.<br />
Especially if you are an underrepresented student, you<br />
could be the only one in the program. If you’re not in a<br />
group that welcomes you and has a great culture, it can feel<br />
even more isolating. We were the place where students<br />
could come when they hit roadblocks...Sometimes we<br />
were the place that would support them in going to talk<br />
about their research... we would pay for travel for them<br />
to go to conferences... we would help them engage with<br />
faculty and collaborators... So many different ways. It<br />
was quite successful and we were able to expand it into<br />
the humanities, because all grad students need support for<br />
different reasons.<br />
What do you think is the future for women in STEM,<br />
and what can we do to make sure that the STEM fields<br />
are inclusive for all people?<br />
That’s a great question. When you look at the number of<br />
women faculty that we have in each one of our disciplines,<br />
we are not very different from most universities... we have<br />
more women who are humanists than social scientists<br />
and scientists. I think 23-27% of our science faculty are<br />
women. 50% of the graduate students are women... but<br />
the numbers just don’t translate into the faculty for several<br />
reasons... so we have a lot of work to do here for women<br />
in science. Part of that is making sure that we have a<br />
culture that is welcoming, but also that we are thinking<br />
about how families and having children affects women<br />
and men differently. It’s serious when you’re a scientist<br />
because you have to be in the lab, right? There are several<br />
family-friendly things that we can do… but making sure<br />
that people have the mentorship that they need is really<br />
important… [and] making sure the climate is such that we<br />
are equally supportive of every single person. That’s not<br />
trivial to pull off.<br />
What can you do [referring to Navami, Emily, and<br />
Kathleen]? Stay in. Don’t quit. If you love it, stay in. Even<br />
if it gets hard just stay in there. Find some great mentors…<br />
I have four mentors that I’ve had for more than twenty<br />
years, including my PhD advisor. They keep me going.<br />
When it got hard, I wanted to quit. And they kept me<br />
going. Get good mentors. What can you do [referring to<br />
Jackson]? What you do is more important than what they<br />
do. All of my mentors are men. That actually is just what<br />
90 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> FEATURED ARTICLE
happened in my life. I’m not saying it’s a good or bad thing.<br />
But you being equally as supportive is important… I’ve got<br />
four of them [mentors] and they’ve been incredible. They<br />
were just the right people for me…<br />
If you love it, don’t let anything keep you from doing it.<br />
Your part is to find what you love and don't give anybody<br />
the power to take you out of doing what you are supposed<br />
to be doing.<br />
What advice do you have in general for STEM majors?<br />
Get some sleep is what I tell my Duke students. Just relax.<br />
It’s okay. It can be pretty intense. Have some fun. I’m<br />
serious about that. I think the reason I love what I was<br />
doing and what I have always done is because I have a<br />
balanced life. The sooner you start taking care of your<br />
whole self and form that habit, the better.<br />
The problem with being an independent scientist is that<br />
you’re independent, which is the same problem I have<br />
with this job… nobody’s telling me when to come to work<br />
every day and nobody’s telling me when to go home. The<br />
problem is that if you are a crazy workaholic, you can do<br />
this 24/7. As an independent scientist, you are actually<br />
working for yourself because you’re running your own<br />
small business. When do you not work because everything<br />
you're doing is for you and your group? Start practicing<br />
now being more balanced. The other recommendation I<br />
would have, at least my experience with STEM majors, is<br />
to make sure you really get a great liberal arts education.<br />
You’re going to be smart enough; that’s not the question.<br />
This navigating across culture, ethics, language… that’s<br />
actually going to make you a more creative scientist. You<br />
never know where you’re going to land in this world,<br />
right? You might be doing your science on the other side of<br />
the world. You need to feel an appreciation for differences<br />
in culture and religion. Get a great liberal arts education<br />
with depth in your science and I think it sets you up in a<br />
beautiful way.<br />
Can you tell us about a time that you failed and what<br />
you’ve learned from that experience?<br />
When I failed? Sure - you want to talk about last week or<br />
yesterday or 20 minutes ago? [laughs]<br />
So in graduate school at UNC, you can get a high pass, you<br />
can get a pass, you can get a low pass. That’s the grading<br />
scale. So I took a mechanistic organic chemistry class and<br />
I got an L. And what that means is that if you’re in a PhD<br />
program you get bumped out of the PhD program down<br />
to the Master's program. Let me give some context to you.<br />
We don’t admit Master's students typically into chemistry.<br />
Because you can go from a B.A. or B.S. to a PhD and almost<br />
FEATURED ARTICLE<br />
nobody gets a Master’s degree intentionally and stops.<br />
So I got bumped down to the Master's and had to earn<br />
my way back into the PhD program, meaning that I had<br />
to pass. So having a good mentor is a good thing, because<br />
right there I would have been gone and everything after<br />
that would not have been possible had my PhD advisor<br />
not said, “Val this is not a big deal. You weren’t prepared<br />
because you didn’t know you were going to graduate<br />
school.” And watching somebody else not flinch is really<br />
good. He was so supportive. He said “This is not a problem.<br />
We’re going to do what we need to do here. We’re gonna<br />
pretend like this didn’t happen and we’re gonna keep you<br />
moving as if you’re on the PhD track.” So I took my PhD<br />
comps.<br />
And I did all of the hourly exams - we took them on<br />
Saturdays; you have to pass a certain number before you<br />
qualify to take the actual oral exam. And then after I took<br />
my comps I had to request in a letter to be readmitted. And<br />
I did and there I was. And it was as if it didn’t happen…<br />
Thank goodness for mentorship, because when your head<br />
is not in the right place, your mentor can keep your feet<br />
moving until your head catches back up…<br />
The beauty for me of that failure is that when a student<br />
comes in here and they have had an academic failure they<br />
don’t think I’ve had one, right? Because they think you<br />
can’t really do the Dean stuff, can you? What I get to say to<br />
them is, it turns out, you can. You’re fine. You can recover.<br />
And then I tell them my story.<br />
I mentor students who think that their first failure is the<br />
end of the road. Turns out you can get a C in physics and<br />
still be the Dean. Perfection is not required.<br />
For sports, Duke or UNC?<br />
Oh - so I’m glad you asked me this. So Duke. I have to tell<br />
you my story - this is so fun. So I hated Duke because I had<br />
two UNC degrees and not only that I had an undergraduate<br />
degree and when you have an undergraduate degree<br />
from UNC the hate is deep. It’s like genetic. I was such a<br />
Duke hater that I would root for anybody playing Duke<br />
because I just wanted Duke to lose and badly, with shame.<br />
[laughs] So when one of my mentors suggested that I<br />
interview for this job, I said to him, “How am I going to<br />
be able to do this?” And he said, “Val, get over yourself.”<br />
And he is a UNC alum and he said this is a great job and<br />
it’s a great place and you’re going to love the people,<br />
you’re going to love the students. And all of that stuff<br />
is going to go away the moment you show up and meet<br />
people. And in my first interview, I walked out and I said<br />
if they offer me this job I’m taking it. And I just found<br />
my people sitting right there at the table and it was just<br />
stunning to me… It’s a serious lesson for me on diversity.<br />
<strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> | <strong>2018</strong>-<strong>2019</strong> | 91
It’s easy to not like people from a distance. The moment<br />
I know you, the game is over. Everything I told myself<br />
about you is no longer true. You just become another<br />
person, and that’s what I found. I sat at that table and I<br />
thought “I love these students.” I love the ideals and<br />
the values and I’m like, “These are my people.” I love<br />
this place. I’m all in Duke. I’m fiercely competitive in<br />
sports and I love great coaching. Duke 100%. On the<br />
weekends, I’m in full Duke gear. It drives my friends<br />
insane. [laughs] But it was surprisingly easy. The people<br />
made all the difference and I love this place. I really do.<br />
So this isn’t a newfound hatred for UNC, it’s a newfound<br />
understanding?<br />
It’s a newfound understanding and I never thought you<br />
could love both of those places. I so appreciate what UNC<br />
has done for me. I love how UNC grew me and supported<br />
me and got me here. And I love that these guys have<br />
accepted me but I also love what we do here - it’s pretty<br />
doggone special and those students are incredible. I get to<br />
love both.<br />
BROAD STREET SCIENTIFIC<br />
The North Carolina School of Science and Mathematics Journal of Student STEM Research<br />
ncssm.edu/bss<br />
VOLUME 8 | <strong>2018</strong>-<strong>2019</strong><br />
92 | <strong>2018</strong>-<strong>2019</strong> | <strong>Broad</strong> <strong>Street</strong> <strong>Scientific</strong> FEATURED ARTICLE