Combinatorial and High-Throughput Screening of Materials ...

ACS Combinatorial Science
Figure 15. Advanced QSAR approach in heterogeneous catalysis.
Composition space (left), physicochemical descriptor space (middle)
and performance space (right). 140
performance categories. By a preprocessing procedure variables
with a significant projection onto the response space can be
identified, irrelevance eliminated, as well as redundancy and
multicollinearity addressed. Afterward the data-driven selection
strategies have been applied to obtain an estimate of the
differences in variance and information content of various attributes
including a comparison of relative importance. Using probabilistic
neural networks various semiempirical QSAR models have been
created. Machine learning models as neural networks and classification
trees have been used by Corma et al. for comparison of the
fitting performance to that of support vector machines for the
modeling and prediction of zeolite synthesis, when the gel molar
ratios are used as synthesis descriptors. 141 It was found that support
vector machines models show very good prediction performances
and generalization capacity in zeolite synthesis prediction. Overfitting
problems sometimes observed for neural networks may be
overcome by this method. The efficiencies of several global
optimization algorithms comparing methods such as GA, evolutionary
strategies, simulated annealing, taboo search and GA
hybridized with knowledge discovery procedures have been compared
using the Selox benchmark by Farrusseng, Maschmeyer,
and co-workers. 142 Selox stands for the selective catalytic oxidation
reaction of carbon monoxide in the presence of hydrogen
and is investigated for fuel cell applications to relate catalyst
composition and reaction temperature effects on the conversion
and selectivity of CO oxidation. In taboo search, the word taboo
originally coming from the Polynesian language Tongan and
indicating things, which cannot be touched, because they are
sacred, a list of taboo candidate solutions are not repeated in the
subsequent iteration of a global optimization algorithm and are
updated in each iterative step. The authors find that an adjustment
of the size of parameter space during different Design of
Experiment (DoE) stages on the iterations through the highthroughput
optimization cycle bear the risk that synergistic
variables are discarded at an early stage, because their interactions
were not detected. Thus becoming rapidly trapped on a local
optimum or performing too lengthy an optimization may occur.
If a certain number of necessary experiments are performed, with
the global optimization algorithms mentioned above the final
(global) optimum is more efficiently achieved and guaranteed,
but only a poor understanding of the chemical system in the form
of knowledge is obtained. On the other hand, the search space
has to have a certain minimum size to apply global optimization
algorithms, so that they are not appropriate for every problem.
A completely different approach for DoE has been suggested
by Schunk, Sundermann, and Hibst. 143,144 This is the structure
oriented library design, which was tested with regard to their
usefulness and applicability in catalyst screening in two different
research projects with different degree of exploratory character,
namely the ammonoxidation reaction for the conversion of C 6
feedstocks to adiponitrile intermediates and the oxidation of
acrolein and methacrolein to the corresponding acids. Compared
REVIEW
to descriptor-based approaches here the additional information
taken into account for discrimination of the catalyst candidates
is limited to structural parameters estimated by X-ray diffraction
techniques prior to and after catalytic testing. In this case
the choice of the target structure is regarded as crucial to the
success of the program. In the first step an initial library was set up
mainly consisting of candidates structurally well characterized
and known to be catalytically active and selective in the reactions
given above. From this initial phase, a lead candidate of highly
complex structure was identified by testing. Despite the identification
of a high amorphous oxide content with crystalline
portions in this lead composition, again a structure-based approach
was chosen to develop this candidate further. After deconvolution
of the lead compound into binary, ternary and quaternary oxides
and inclusion of other element combinations from structural
databases forming the same structure types these potential
candidates of less complex nature were taken as reference points
of compositional designs to build a multidimensional grid of
highly crystalline multinary candidate materials. A variation in
stoichiometry, phase composition and crystalline nature increases
the diversity of the library constituents further on a
systematic basis. Classical DoE elements such as composite
designs facilitate the evaluation of the results and the approach
may be combined with even more complex methods as artificial
neural networks (ANN) or GA. Problem in this approach may
arise from differences between bulk and surface structure, as the
catalytic activity is exclusively related to the latter, but the crystal
(bulk) structure has been here the decisive design parameter.
Perhaps one can argue that both are related, that a certain bulk
structure enables a specific surface structure, which gives the
catalytic activity.
Another general problem in catalytic studies is that not only a
single objective function has to be optimized, a common scenario
is the evaluation of response surfaces for conversion as well as
selectivity to a number of alternative products. A GA based
multiobjective DoE aided a high-throughput approach to optimize
the combinations and concentrations of a noble metal-free
catalyst system active in the selective catalytic reduction of NO
by propene. 145 A direct comparison of single- and multiobjective
design of combinatorial experiments was also performed for the
optimization of a solid catalyst system active in the selective
oxidative dehydrogenation (ODH) of propane. 146 The basis of
the multiobjective optimization was a strength Pareto evolutionary
algorithm (SPEA-2), that is, a Pareto-based algorithm
operating toward two objectives, minimization of the distance
toward the Pareto-optimal set and maximization of the diversity
within this set. Elitism is implemented in SPEA-2 by an additional
archive population composed of the best 24 nondominated
individuals during the search, which is used in combination
with the regular population as mating pool. Due to the SPEA-2
features it is not surprising that in a Pareto plot of selectivity
versus conversion (Figure 16) the SPEA-2 solutions are well
distributed over Pareto space. The most important element
combinations found in the high conversion and high selectivity
region are highlighted. To have a fair comparison, a new
population has been created, denoted as the archive population
of the single objective approach. Similar to the way the multiobjective
archive population was created, the single objective
archive population consisted of the 24 best catalysts found so far.
The major difference compared to the SPEA-2 archive is that
no Pareto-selection was performed and the catalysts were only
selected with respect to their yield. For the single objective approach
593 dx.doi.org/10.1021/co200007w |ACS Comb. Sci. 2011, 13, 579–633