Connections between simulations and observation in climate ...

Studies in History and Philosophy of Modern Physics 41 (2010) 242–252Contents lists available at ScienceDirectStudies in History and Philosophyof Modern Physicsjournal homepage: www.elsevier.com/locate/shpsbConnections between simulations and observation in climate computermodeling. Scientist’s practices and ‘‘bottom-up epistemology’’ lessonsHélene GuillemotCentre Alexandre Koyré, EHESS—CNRS, Pavillon Chevreul, 57, rue Cuvier, 75005 Paris, Francearticle infoKeywords:Climate modelsModelingComputer simulationObservation dataScientific practicesValidationEvaluationabstractClimate modeling is closely tied, through its institutions and practices, to observations from satellitesand to the field sciences. The validity, quality and scientific credibility of models are based oninteraction between models and observation data. In the case of numerical modeling of climate andclimate change, validation is not solely a scientific interest: the legitimacy of computer modeling, as atool of knowledge, has been called into question in order to deny the reality of any anthropogenicclimate change; model validations thereby bring political issues into play as well. There is no systematicprotocol of validation: one never validates a model in general, but the capacity of a model to account fora defined climatic phenomenon or characteristic. From practices observed in the two research centersdeveloping and using a climate model in France, this paper reviews different ways in which theresearchers establish links between models and empirical data (which are not reduced to the lattervalidating the former) and convince themselves that their models are valid. The analysis of validationpractices—relating to parametrization, modes of variability, climatic phenomena, etc.—allows us tohighlight some elements of the epistemology of modeling.& 2010 Elsevier Ltd. All rights reserved.When citing this paper, please use the full journal title Studies in History and Philosophy of Modern Physics1. IntroductionAnthropogenic climate change has been studied in climatescience laboratories over the past thirty years. Over the pasttwenty years, it has become the object of global expertise with thepublication of IPCC reports to which thousands of scientists havecontributed. More recently, the issue of climate change hasclimbed to the top of the international political and diplomaticagenda, leading to major evolutions in considerations of economy,geopolitical equilibrium, North–South relations, consumption andlifestyles. Even if the problem has become increasingly political,scientific expertise retains its central role. From the beginning,climate change was a largely ‘‘science-driven’’ issue, mostlyresulting from climate modeling and simulations of future climatechange (Heymann, 2009).Considering the importance of these political and economicstakes, climate modeling has not given rise to much research inthe social sciences. In particular, there are few studies of thevalidation and evaluation of models through comparison withobservational data, even though this is an essential domain thatE-mail address: guillemot@damesme.cnrs.frendows models with their scientific character as well as with thecredence we lend to simulations. 1 Indeed climate scientists spendmuch of their time testing their models, comparing them tomeasurements of the real climate, and these evaluations constitute,in their opinion, the principal guarantees of the skill ofmodels, of their scientific validity and of the reliability of futureclimate projections.This work is in large part based on a study carried out between2003 and 2006 in the two most important institutions thatdevelop and use climate models in France, the Laboratoire de1 I should define the terms model, simulation and numerical experiment as theyare used in this article, because these definitions vary depending on the discipline,language, uses and authors. Here, I adopt the terminology currently used byclimate modelers. The word model (climate model, Global Circulation Model,meso-scale model, etc.) designates a program that is meant to run on a computerand carry out algorithms step by step. Simulation designates the results producedby a model’s output, i.e. the product of the calculations performed by thecomputer by means of that model in simulating a particular climatic configurationthat is provided as input (for example, the simulation of the climate in the 20thcentury with such increase of CO 2 levels with the Météo-France’s model, orsimulation of the last glacial maximum. . .). A numerical experiment is a simulationwith the objective of virtually exploring the climate’s behavior by varyingparameters or representation of phenomena.1355-2198/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.doi:10.1016/j.shpsb.2010.07.003

H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252 243Météorologie Dynamique (LMD) of the CNRS, in Paris, and theCentre de Recherche de Météo-France (the French organism forweather forecast) in Toulouse. This study primarily relied on twotypes of sources: internal laboratory records (activity reports,planning reports, internal journals, Web sites, colloquia, workshopreports, etc.) and extensive interviews with scientists. InSection 2 of this article, I draw a brief review of studies oncomputer modeling and climate models, and try to frame the roleof observation in validation of climate models. In Section 3, after aquick survey of the observational data used in climate science,and an insight of the fundamental difficulties that faces validationof climate models, I seek to describe and analyze what sorts ofrelationships modelers establish between data and model simulations,based on observing actual scientific practices throughseveral examples: top-down validation, local evaluation ofparametrization against field experiments, validation of climaticeffects of a small scale processes, evaluation of a model’s capacityto simulate particular phenomena. . . In Section 4, I attempt toderive a few epistemological lessons on climate modeling andwhat distinguishes it from more traditional theoretico-experimentalsciences in regards to their relationship to data.2. Studies on numerical models, climate modeling andobservation, a brief review2.1. Models between theory application and the creation ofknowledgeComputer modeling, which emerged after World War II, hassince the 1970s and 1980s permeated nearly all sectors of science,technology, industry and economy. Yet, it was only recently that itstarted eliciting any notable research activity in philosophy,history of science or science and technology studies (STS). For along time, discourse on models privileged a ‘‘semantic’’ approachthat focused on theories and according to which a model is firstand foremost a representation that gives meaning to a mathematicalformalism. Over the past few years, however, there hasbeen a renewed reflection on models. The ‘‘cult of theory’’ hasbeen criticized by authors who assert models’ superiority overtheories as representations of the world (Cartwright, 1983, 1999;Sismondo, 1999). According to Morgan and Morrison (1999),models consist of elements other than just theories and data, and‘‘it is precisely because the models are partially independent ofboth theories and the world that they have this autonomouscomponent and so can be used as instruments of exploration inboth domains’’. As autonomous agents, models can therebybecome active ‘‘mediators’’ between theories and the world, andmay have more to teach us than theory does on its own. Owing tothis mediating function, the modeling of complex physicalsystems—systems that are poorly understood even when theoriesof underlying processes exist—cannot be reduced to calculationsthat apply known laws: ‘‘The computer simulations (. . .) involvesa complex chain of inferences that serve to transform theoreticalstructures into specific concrete knowledge of physical system.(. . .) This process of transformation is also a process of knowledgecreation, and it has its own unique epistemology’’ (Winsberg,1999, p. 275).Nevertheless, the cognitive approach to models in undoubtedlyinsufficient. The computer’s massive influence, the emergenceof new scientific objects, modeling practices that are moreand more heterogeneous, as well as their increasing use inestablishing expertise, makes ‘‘studying the activity of modelingin its institutional, technical and political environment, andwithout dissociating cognitive and social elements’’ (Armatte &Dahan, 2004, p. 245) all the more necessary. According toKnuuttila et al. (2006), a convergence is already discerniblebetween philosophers (up to this point focused on theories) andresearchers in science studies (traditionally oriented towards thelaboratory), who are sharing interest in the role of modeling andsimulation in scientific practices.2.2. What is at stake in studying climate modelsUntil very recently, climate modeling attracted few epistemologists,in part because it seems to fall under application, not underfundamental physics, bringing into play established physicstheories and loosely formalized heterogeneous elements. Moreover,it is embedded in expertise and embroiled in major politicalstakes. One would think that the latter characteristics would elicitthe interest of the social sciences. This has been the case, but to alimited extent: the sociopolitical stakes inherent in climatemodels seem to have a contradictory influence in this regard.In the context of climate change debates, models are oftenbrought center-stage and placed into question—how can theypurport to predict the future of climate? Political controversiesare frequently transposed to the scientific, and even epistemological,field: climate change skeptics shed doubt not only onprojections but on the models themselves and seek to ‘‘stigmatizemodeling as inferior science on philosophical grounds’’ (Norton &Suppe, 2001, p. 67). Much of the time, such polemics focus on therelationship between models and observational data (Edwards,1999). Computer models are contrasted with ‘‘sounds science’’,found on data and solid theories, and the possibility of verifyingmodels’ projections against data is questioned (Oreskes et al.,1994). In response to these criticisms, Paul Edwards has shownthat if models and data are effectively in a relationship ofinterdependence—models being ‘‘data-laden’’ and data ‘‘modelfiltered’’—thisrelationship is not circular, but symbiotic, eachgaining advantage from the other. In a detailed analysis, Norton &Suppe (2001) maintain that interdependence exists betweentheories and experiments as well, and that the absence ofcertainty, simplifying hypotheses, etc. are not the prerogative ofmodels, nor do they constitute definitive obstacles to knowledge.These authors conclude that models may be trusted for the samereasons and to the same extent that traditional experiments areendowed with credibility or trusted.If, despite these debates, climate modeling as a new mode ofproduction of knowledge has attracted a small number of studies(comparatively to the economic or political side of the problem), itcould be that the political role of these models makes their studymore delicate, possibly risky. The research conducted in thisdomain has shown how the representation of uncertainty(Shackley & Wynne, 1996), the estimate of climate sensitivity(Van der Sluijs et al., 1998), and recourse to flux adjustments(Shackley et al., 1999) all result in negotiations or in scientiststaking into account the expectations of policy makers. This has attimes placed researchers in an uncomfortable position: emphasizingthe co-construction between science and politics risksbeing co-opted by a critique of the objectivity of climate scienceand the validity of its results. Even when it comes to ‘‘defendingthe indeterminate character of the climate sciences’’ (Shackleyet al., 1999, p. 35), any reference to uncertainty or ambiguity canbe turned into a weapon in the hands of critics of the fight againstglobal warming (Edwards, 1996). Forsyth (2003, p. xiii) analyzedit well when he wrote: ‘‘these are controversial times for writingabout environmental science and politics’’. The political stakescan lead to closing black boxes more so than to analyzing theircontent; to prematurely transforming ‘‘matters of concern’’ into‘‘matters of fact’’ (Latour, 2004).

H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252 245synergy between the various scientific tools used in the study ofthe atmosphere: theory, modeling, observation, and instrumentation.In other laboratories that participate in the development of aGCM together with LMD (and that studied other components ofthe climate such as the ocean, the biosphere, etc.), the experimentalor instrumental dimension is as important as modelingand precedes it. The same can be said for Météo-France. All theselaboratories have strong and autonomous identities, focused onan object of research tackled using several tools. Traditionally,their scientists were often more interested in understandingclimate (or ocean. . .) mechanism and developing parametrizationclose to the physics of the process than in building coupled globalmodels (Dahan Dalmedico & Guillemot, 2008). Modelers attachimportance to links between observation and modelisation, andtry to develop these relationships—even if they constituted adistinct community, separated from communities conductingobservation, and if communication difficulties exist between thetwo. That culture of observation constitutes a shared value amongFrench laboratories in this domain. 33.2. Corrected, homogenized, and reanalyzed observational dataWhat are the observational data with which simulations areconfronted? Three categories can be distinguished: data from vastnetworks of meteorological stations, on land, and across theoceans; data provided by instruments on satellites (meteorological,oceanographic, or scientific satellites), and finally, datacollected during field experiments focused on a specific region,phenomenon, and time period. Dozens of field campaigns havebeen organized since the 1970s, requiring massive human andtechnical investments (airplanes, ships, balloons), and bringingtogether numerous organizations (from meteorology, but alsotransportation, defense, forestry, . . .).It is important to insist on one essential characteristic of thesedata—where they come from: they have been heavily reconstructed(Edwards, 1999). In the case of the climate sciences, ‘‘rawdata’’ are not usable. What a satellite instrument is measuring isvery much removed from the physical parameters which areneeded by meteorologists or climate scientists: for example, inorder to obtain the surface temperature of a point on the Earth’ssurface from a signal captured by a meteorological satellite, nofewer than three intermediate models are necessary that combineinitial data with numerous other factors (Kandel, 2002). Informationcoming from surface stations are not as indirect, but they aremore heterogeneous and scattered, and therefore always have tobe completed and harmonized with the aid of computer models.On the other hand, obtaining broad and homogenous data seriesover many decades (which is indispensable for detectingsignatures of climate change), demand important reconstructionefforts: data must be corrected and re-calibrated to take intoaccount changes in measuring technique or detectors. Météo-France developed a program that could locate the ‘‘ruptures inhomogeneity’’ in measurements it had collected over a century(these ruptures were largely due to modifications in theinstrument’s surroundings). This allowed Météo-France to reconstructa homogenous evolution of climate in France throughoutthe 20th century (Moisselin et al., 2002).Finally, an increasingly common type of data is the ‘‘reanalysisdata sets’’, which takes the interdependence between models and3 We do not find in our study the distinction observed by Sundberg (2007)between experimenters’ and modelers’ aims. On the one hand, we have notstudied the community of researchers working on observations—only thecommunity of modelers. On the other hand, the French modeler’s objective is(at least) as much to improve the understanding of climatic processes as tocreating representations of processes that can be implemented in models.data even further. These are series of meteorological data from thepast entirely reprocessed and completed with a climate modelthrough a procedure inspired by the ‘‘variational assimilation’’technique, which is routinely used in weather forecast in order tointroduce observational data in near real-time into projectionmodels—in other words, a kind of dynamic interpolation of pastobservations in a model. Thus reanalyzed data fill in the gaps andcorrect the shortcomings of observations: the data becomecomplete, homogenous, and coherent. Veritable hybrids ofobservation and modeling, these reanalyses add new uncertaintiesthat are linked to the shortcomings of the initial data, to theuncertainties of the model, and most of all, to the sparsity ofobservations that are artificially filled in by modeling (Bengtssonet al., 2004). 4 When using them to validate models, modelers tryto take into account these uncertainties by, for example, lendingless credence to the reanalysis of past periods than they grant tothose of more recent periods (interview with A, Météo-France,January 2004).3.3. Fundamental difficulties in the validation of modelsBefore turning to validation practices, let us quickly describeclimate models. A numerical climate model (like a weatherforecast model) is a Global Circulation Model (GCM): it seeks tosimulate atmospheric circulation, represented by a three-dimensionalgrid, over the course of time. Within each grid cell and foreach time step, the computer calculates the parameters that arecharacteristic of the atmosphere’s state from their values in thepreceding time step by running the algorithms that constitute themodel. The model is composed of two large parts: a part thatdescribes the movement of air masses, whose algorithms arederived from fluid dynamics equations; modelers call this thedynamic part. The other part, the so-called physics part, calculatesthe forcing of atmospheric circulation; it deals with verticalexchanges between atmosphere and outer space or with theEarth’s surface 5 (the ocean, continental land mass, or ice). Theseexchanges—of radiation, energy, water, etc.—occur at a muchsmaller scale than that of the model (which is on the order ofseveral hundred kilometers) and they are represented by ‘‘parameterizations’’that statistically reproduce, at the scale of a gridcell, the climatic effects of the phenomena under consideration.The parameterizations, veritable smaller models nested withinthe main model, are extremely diverse; certain more or lessdirectly arise from physical theories (when it comes to radiation,for instance, the parameterizations synthesize quantum physicscalculations), others are more empirical or phenomenologicalrepresentations (vegetation ecosystems, for example) (Guillemot,2007; Sundberg, 2007).Meteorological models and climate models are thereforerather similar. However, they are used differently. In everydayforecasts, the atmosphere’s future state is calculated in a4 The reanalysis can be classified depending on the relative influence of theobservational data and the model :‘‘An A indicates that the analysis variable isstrongly influenced by observed data, and hence it is in the most reliable class (e.g.,upper air temperature and wind). The designation B indicates that, although thereare observational data that directly affect the value of the variable, the model alsohas a very strong influence on the analysis value (e.g., humidity, and surfacetemperature). The letter C indicates that there are no observations directlyaffecting the variable, so that it is derived solely from the model fields forced bythe data assimilation to remain close to the atmosphere (e.g., clouds, precipitation,and surface fluxes). ‘‘(Kalnay et al., 1996).5 Calculations of dynamic part are carried out on the three-dimensional grid,whereas physics part may be seen as a juxtaposition of air columns that do notinteract. Dynamic and physics parts deal with different kinds of variables, andhave different time steps (the time step of the physics part being larger). Cf. LMD(2005).

246H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252deterministic fashion, starting from its state as measured a fewdays or hours earlier. 6 For longer ranges of a few weeks, theatmosphere’s dynamic is chaotic, and only statistical propertiesare predicted. 7 Weather is not predicted for November 30, 2077;instead, averages (or variabilities) are provided for temperature,precipitation, etc., for the period between 2070 and 2100. Thisbrings us to the first difficulty of evaluating climate models. Formeteorological models, validation seems rather simple, at least inprinciple: forecasts can be compared the very next day to thecharacteristics of actual weather. In climate modeling, however,the simulations only have a statistical value, and the comparisoncan only be based on data series for present or past climate.Validating projections of future climate is even more difficult andindirect (I will return to this later).The evaluation of climate models faces another fundamentaldifficulty: the entanglement of climate processes in the modelmakes validating the representation of a particular phenomenonextremely delicate. When modeling first began, the manageabilityof models gave rise to certain illusions about their capacity toexplain the climate’s mechanisms. Joseph Smagorinsky, the‘‘father’’ of the first GCM, wrote that ‘‘The main advantage indiagnosing model simulations is that we know a great deal aboutthe mathematical distortions we have introduced, and right orwrong, we have all the variables defined everywhere and all of thetime’’ (cited in Nebecker, 1995, p. 178). Jule Charney, anotherimportant pioneer of climate modeling who founded and directedthe American Numerical Meteorology Project in 1948, noted withoptimism: ‘‘When a computer simulation successfully synthesizesa number of theoretically predicted phenomena and is in accordwith reality, it validates both itself and the theories’’ (cited inNebecker, 1995, p. 180).However, as models became more complex, it was necessary toobserve that Charney’s statement could backfire: the apparentvalidation of a model through observation could result from thecombination of a false hypothesis and a faulty representation inthe model, which together give a good climate simulation due towhat modelers call ‘‘error compensation’’. A simulation couldthereby appear ‘‘correct but for the wrong reason’’. Modelvalidation proved to be more delicate than theory validation:theories are also under-determined by experiments, in accordancewith Duhem’s thesis, but the difficulty is pushed to a higherdegree in models, which bring into interaction numerousintricated hypothesis and representations of processes.As models incorporated a growing number of elements andprocesses, scientists were confronted with the difficulty ofinterpreting their simulations, of attributing a cause to aphenomenon and understanding multiple feedbacks. The largenumber of interacting mechanisms, which constitutes thespecificity of models and makes numerical simulations andexperiments possible, is also what makes their validationparticularly arduous (Lenhard & Winsberg, this issue). 8 Moreand more, the researchers’ work came to consist of understandingwhat happens in their model, which parameter influences suchand such result and how. The model itself, ‘‘almost as difficult tounderstand as the real climate’’, became an object of study.‘‘Climate modeling means working on what comes into the model6 Nowdays, meteorological agencies also predict the weather 4–10 days inadvance based on probabilistic ‘‘ensemble’’ forecasts.7 These differences in scale can introduce certain differences between weatherforecast models and climate models. For example, ocean currents or polar icedynamics are represented in climate models, but not necessarily in meteorologicalmodels. Reciprocally, local details are important for weather forecasting, but notfor climate modeling.8 Lenhard & Winsberg (this issue) develop a similar analysis of what they call‘‘confirmation hollism’’ caused by the ‘‘entranchment’’ of climate models, whichmakes them ‘‘analytically impenetrable’’.and what comes out of it, it means understanding what there is inthe model, how it reacts, what are its characteristics’’ explains aresearcher at the LMD (interview with B, LMD, July 2003).3.4. Evaluation on all scalesModelers tend to distrust the term ‘‘validation’’ and prefer touse expressions like ‘‘evaluation’’—the norm in IPCC reports(Randall et al., 2007). The concept of validation was criticized in aScience article (Oreskes et al., 1994) that maintained thatnumerical climate models could be neither verified nor validatedin a rigorous fashion, but could only be confirmed under certainconditions. The article gave rise to many debates, and modelers,prompted to a modesty of sorts, declared more humbly that it wasa question of ‘‘evaluating the model’s performance’’ or of showingthat a model is ‘‘sufficiently good to be useful’’.Beyond these nuances, the first essential point is that a modelis not evaluated in general. What is evaluated is its capacity toaccount for a particular climatic characteristic or a definedphenomenon. Evaluations are performed at all temporal andspatial scales and at all levels of the model: they might concerncharacteristics of average global climate, a geographically limitedphenomenon, the model’s capacity to represent a specific feedback,etc. The type of validation depends on how the model isused: a model might be appropriate for medium-term simulations(for the study of interannual variability, for instance), but not forsimulations spanning one hundred years. Consequently, there isno systematic protocol for evaluating models. An evaluationsupposes a preliminary question, the definition of a problem forwhich an appropriate procedure for confronting models and datamust be imagined. ‘‘It requires astuteness and creativity more sothan respect for a method’’, summarizes a LMD researcher(interview with C, LMD, July 2003). Modelers must know howto make use of everything at their disposal in order to ‘‘constrainthe system’’ (interview with D, Météo-France, January 2004). Forexample, the eruption of Mount Pinatubo in 1991, presentedresearchers with a unique opportunity to validate the representationof aerosols in models.Despite the plurality of evaluations, two greater types ofapproaches are distinguishable, one ‘‘top-down’’ and one ‘‘bottomup.’’ Top-down validations consist of comparing simulations toglobal data series. This is an older type of validation: models wereinitially validated in relation to average climate, by comparing amap of the simulated climate with a map of climatic averagesbased on observational data, in order to see if the model correctlyreproduced large climatic characteristics (temperature, wind,main phenomena). Subsequently, it was also necessary to validatethe climate’s variability: seasonal cycles, interannual and decennialvariability, monsoons, extreme event frequency, El Nino. . .Model evaluation concerned itself with larger and larger variabilitydomains. In this way, Météo-France studies the capacity ofits model Arpege to reproduce the variability spectrum ofprecipitation, including extreme precipitation, by studying thedistribution of rain intensity down to the daily scale.Since the 1990s, another kind of validation (bottom-upvalidation) has been used for testing parameters against fieldexperiments. The procedure consists of studying a climatephenomenon in those geographic locations where it is predominant(for example, the monsoon in India, or winter storms in theNorth Atlantic) and using this case study to test the manner inwhich this phenomenon is represented (or ‘‘parametrized’’) in themodel (Chaboureau & Bechtold, 2002). This methodology hasbeen used to test many parametrizations through numerous fieldexperiments—for example parametrizations of clouds in severalGCM in Europe have been tested within the framework of

H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252 247program EUCREM and EUROCS. 9 The procedure can be decomposedinto three steps, within which intervene no less than threedifferent models. The first step consists of carrying out ameasurement campaign within a zone that is monitored byobservation stations. Then—this is the second step—a meso-scalemodel, also called ‘‘cloud resolving model’’ (with grid cellsmeasuring a few kilometers) is used to simulate weatherevolution in this zone during the time period under consideration:meso-scale specialists enter into their models parameters measuredat the beginning of this period, and define limit conditions,so that the simulated climate resembles the climate observedduring the campaign. In this way, the meso-scale model isvalidated by observations in a detailed fashion. Finally, climatemodelers test the studied parametrization in a simplified, onedimensionalversion of the model. This so-called ‘‘column’’ modelconsists of a single horizontal grid cell with all of the verticallayers superimposed; it contains all of the ‘‘physics part’’ of a GCMbut no dynamics, so, it is far less difficult to use. This onedimensionalmodel is provided in input with the external climatedata from the meso-scale model, then it runs and its simulation iscompared to that produced by the meso-model, which permitsmodelers to validate its parametrization, and eventually, toameliorate it (interview with B and E, LMD, July 2003 and July2005).To summarize: modelers compare the climate simulated by aone-dimensional model equipped with the parametrization to betested, to a climate simulated by the meso-model that is validatedby observation. In this sophisticated methodology, the meso-scalemodel plays the intermediary between the data and the largescalemodel, by providing this large-scale model with a completeset of ‘‘predigested’’ data reconstituted from observations. Suchstudies of parametrization through field experiments have multipliedover the past fifteen years or so due to the availability ofoperational meso-scale models (such as the Méso-NH modeldeveloped by Météo-France).In following these steps, what remains to be done is theimplantation of the new (or ameliorated) parameterization inthe GCM, and it is here that the real difficulties begin (joke themodelers), since one invariably obtains a simulated climate that iscatastrophic! This has to do with the error compensationmentioned above: even though the preceding model wasequipped with a parameterization that is farther from thephysical mechanism, the model’s errors and approximationscompensated each other in order to engender an accurate climate.The parameterizations in a model constitute ‘‘a family’’ (interviewwith A, Météo France, January 2004), and the implantation of anew parameterizations requires much research, testing, andcontrol.3.5. Evaluating long-term feedback between clouds and radiationNevertheless, validation cannot be reduced to global evaluationof a model by comparing it to global databases on the onehand, and validation of parameterizations by local case studies onthe other. In particular, one must evaluate the so-called ‘‘climaticeffects’’ of some parametrizations. Small-scale processes representedby parameterizations have consequences that cannot beperceived during field study, but that become apparent in thelong-term and over longer distances. These climatic effects alsomust be reproduced through parameterization. One of the mostimportant being the feedback between clouds and radiation.Clouds are the most influential element of climate, and their9 See EUROCS website: http://www.cnrm.meteo.fr/gcss/EUROCS/EUROCS.html.effects are particularly complex and difficult to model, speciallytheir interaction with radiations—which they can reflect orabsorb depending on the wavelength, type of clouds, and otherparameters. While cloud formation, convections or precipitationscan be studied through local campaigns, the interaction betweenclouds and radiation are not proximately observable (interviewwith A, Météo-France, January 2004).It was the global warming issue that made necessary the studyof the long-term impacts of these parameterizations. Indeed,climatic changes are nothing other than distant and large-scaleconsequences of a small scale warming, provoked by an increasein green-house gas concentration. This warming provokes adisequilibrium in the atmosphere’s lower layers that interactswith numerous small-scale mechanisms, which are representedby parameterizations: turbulence, cloud microphysics, etc.(Le Treut, 1999). A number of studies have established thatdifferent cloud representations are primarily responsible for theuncertainty of models as well as for differences between differentmodel’s projections of climate change. It is therefore crucial tounderstand how cloud-radiation feedback evolves with climatechange, and to evaluate if models correctly reproduce thisevolution. Because they cannot observe future climate, climatescientists have to settle with exploring the correlations betweenclimatic variables with increase in temperature. However, theyface an additional difficulty: temperature increase principallyaffects atmospheric dynamics, which favors certain types ofclouds. This first-order effect masks the impact of temperatureincreases on cloud microphysics and on convection heights—inshort, on the physical evolutions of clouds which have important,long-term climatic impacts because they affect cloud–radiationinteraction (interview with F, LMD, July 2003).A key point in modeling is to wisely choose the variablesthat characterize the phenomenon under study, called ‘‘diagnostics’’by the modelers. During a stay at the Godard Institute forSpace Studies (GISS) in New York, a young LMD researcher (F)developed an original method for exploring cloud–radiationfeedback and its sensitivity to warming (Bony et al., 2004).Drawing on all sorts of observational data on clouds (fromsatellite databases to reanalyzed data, etc.) and using statisticalanalysis, F sought to establish the diagnostics that showed theeffect of temperature on cloud physics, in given dynamicconditions. Her new diagnostic has been used later to test thecapacity of cloud parametrization of the LMD climate model torepresent this feedback. Let us expose briefly the principle of thisdiagnostic.Instead of representing clouds according to longitude andlatitude, F decided to represent them along a single axis, as afunction of the vertical speed of the air. ‘‘In this way, we canrepresent on a single axis, on one side where the air rises, on theother, where it goes down’’, she explains ‘‘Over there are all ofthe convective clouds, here the low clouds [. . .] In classifying allof the zones as a function of vertical speed, we obtaincontinuous variation in properties of clouds and of the watervapor; everything is organized in a simple manner. . . It’smuch easier to analyze what is happening [. . .] Instead of lookingat things regionally, we try to classify them in a syntheticmanner’’.By coming up with this diagnostic (widely used ever since),which provides the specialist with a new synoptic panorama ofcloud organization, the scientist was able to extract newinformation from a flood of available data. The search for newways of seeing is described as creative and funny: ‘‘One canchoose to look at things in more detail, by looking at differentvariables, by carrying out statistical analyses (. . .) Whatinterests me is looking at things otherwise; in a somewhat moretwisted way. . .’’.

248H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–2523.6. Making simulations and data ‘‘speak the same language’’A diagnostic is a means of analyzing observational data, butalso the output of the model’s simulations. Diagnostics thereforeconstitute an evaluation method: applying the same diagnostic tosimulations and observations renders them commensurable andallows for the comparison of two schemes constructed inaccordance with the same criteria. Some climatologists enforcethis rule when evaluating models: ‘‘use observations and simulationsin the exact same way’’ (interview with F, LMD, July 2003).For example, they may choose to study the radiation sent backinto outer space by the outer layer of the atmosphere becausesatellite instruments measure this flux. However, they wouldavoid working on the radiation’s vertical distribution, which themodel can calculate, but which cannot be measured. In this case,researchers voluntarily limit the field of numerical experiments sothat it only includes those that can be directly compared toobservation.On this theme, one can give two additional examples of themodel–data relationship. LMD researchers who wished toevaluate their model’s reproduction of the monsoon in India,used a program for analyzing meteorological data that wasdeveloped by a researcher at the European Centre for Medium-Range Weather Forecasts in Reading (UK). This program selectsfrom data the parameters that characterize monsoon depressions,thereby constructing trajectories; subsequently, it produces astatistical representation of the properties of these perturbations.With the help of this program, LMD researchers have analyzedmonsoon depressions simulated by their models in a region inIndia over nine (simulated) years. Among those virtual depressions,they’ve identified those that have the characteristics of amonsoon, and have compared them to real observations of Indianmonsoons that are analyzed by the same program. Theirconclusion is two-fold: ‘‘(1) LMD’s model is capable of simulatingmonsoon depressions that exhibit realistic circulation characteristics,(2) the ‘‘tracking’’ method is powerful and offers a panoplyof robust quantitative statistical studies that make it possible toeffectuate a precise analysis of the simulated depressionsystems’’. 10 With this example, we can see that research doesnot limit itself to evaluating a model; it is also concerned with themethod of analysis used, and with monsoon displacement. Thiswork actually deals with three levels of analysis: the monsoon(whose statistical characteristics are under study); the model(which is to be evaluated in relation to its capacity to reproducethe monsoon), and finally, the common method of analysis(whose efficacy is under evaluation).Another example, which will also keep us in India: this time itis a matter of validating the model’s capacity to simulate atropical hurricane that hits the northeast of India on October 29,1999. First, researchers introduce into the model the atmosphericconditions of October 21, 1999, and then the model simulates theclimate of the following ten days with the zoom function pointedat India. A program made it possible to transform calculatedclimate characteristics by simulating the signal that would bedetected by the Meteosat satellite if it observed the phenomenon.Thus, researchers could easily compare ‘‘real’’ images of cyclonestaken by Meteosat in October 1999 with simulated satelliteimages (this comparative method is named ‘‘from model tosatellite’’.). In all of these examples, it is a matter of ‘‘makingmodels and measurements speak the same language’’. 1110 Extracts from LMD’s internal journal LMDZ Info no. 0 (July 2000), p. 12:‘‘caractérisation des dépressions atmosphériques simulées par le modeles decirculation générale du LMD (MCG LMD6)’’.11 Extract of ‘‘Rencontre Modeles—Données’’, LMD’s internal journal LMDZInfo no. 0 (July 2000), p. 7.4. Bottom-up epistemology of climate modeling4.1. Some epistemological remarks on the relationships betweensimulations and observationsIn order to be more through in this outline of the relationshipbetween data and simulations, I would have to give moreexamples, notably examples of how simulations of past climateare compared to paleoclimatic data. Besides, intercomparisonprojects organized by the Program for Model Diagnosis andIntercomparison (Atmosphere Model Intercomparison Project,Paleoclimate Model Intercomparison Project, Coupled ModelIntercomparison Project, etc.) play an essential role in theseevaluations, and in structuring the modeling community as well;they deserve to be discussed in details (see Lenhard & Winsberg,this issue). However, the cases described above, even if succinctand limited, do already point to a few reflections on a bottom-upepistemology of modeling that emerges from an analysis ofpractices.The first remark concerns the manipulation of data. Thecomputer does not merely engender virtual climates, it alsoallows researchers to manipulate observations or rather ‘‘to lookat things otherwise, in a somewhat twisted way’’ (as F pointedout) in order to uncover hidden correlations. Due to open accessto databases and reanalyses, playing with observations is anintegral part of the modeler’s work. We are quite far here from theopposition between fixed and untouchable data and readilymalleable simulations. It is almost the inverse: having alreadyundergone all sorts of transformations, the data can be sortedfurther and ‘‘twisted’’ in all directions since they come fromobservation and remain unswervingly linked to it. Simulations, onthe other hand, demand a more cautions handling: because theyhave no link to the real world, they are bound to always beingmeasured against data.Certainly, the fact that data are not given, that they arecalibrated, corrected, reduced, that they need theory andmodeling—all of this is not new and has been studied bynumerous authors (Galison, 1987; Hacking, 1989; among others).The general character of data manipulation was one of thearguments Norton and Suppe (2001) used to counter critics ofmodel validation. Knorr-Cetina (1992) remarks that ‘‘the theoreticalrelevance of laboratories’’ rests ‘‘upon the malleability ofnatural objects’’. Within the domain of the field sciences, Latour(2001) impressively analyzed the ‘‘circulating reference’’ (thesuccessive transformations of data) in pedology (science of soil).The work on climate observation that I have described recallsLatours description in this regard: Between science and its object,there is neither correspondence, nor gaps, Latour writes, but asuccession of small displacements, ‘‘cascade of transformations’’.Similarly, we have seen that cascades of transformations allowback and forth between data and climate simulations. ‘‘This chainhas to remain reversible. The succession of stages must betraceable, allowing for travel in both directions’’ (Latour (2001)p. 72). This claim applies of course to climate data. 12 In both cases,we find the same capacity of combination and comparison of data,allowing new transformations: ‘‘Hardly surprising, then, that inthe calm and cool office the botanist who patiently arranges theleaves is able to discern emerging patterns that no predecessor12 Without going into a fresh controversy that has opened up as this text issubmitted for publication, and which does not encroach on the problematic of thisarticle, the ‘‘ClimateGate’’ affair (the hacking of more than 1000 emails from theClimate Research Unit (UK) and the polemics that followed few days before theClimate Change summit in Copenhagen in December 2009), highlights the crucialimportance of accessibility and traceability of scientific data in a highly politicalcontext.

H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252 249could see (. . .) Scattered through time and space, these leaveswould never have met without her redistributing their traits intonew combinations’’ (Latour (2001), p. 38). If we stand in the leavesfor clouds, and the botanist for the climatologist, we would thinkwe are reading a description of F exploring cloud’s properties.Regarding the manipulation of data, there is no difference inprinciple between climate modeling and other sciences. Nevertheless,this data malleability is of a much higher degree in theclimate sciences, due to the extensive use of computers on alllevels.The second remark is that exchanges between models andobservations are more symmetrical than we would believe. Thiscirculation continually moves in both directions, in an iterativeback and forth, whereby both data and simulations are repeatedlycompared and discussed. Therefore, even if customarilymodels are evaluated based on observations, the inversealso occurs and models serve to validate methods of interpretingthe data. This is especially the case in paleoclimatology.The interpretation of data from the distant past is particularlydelicate because it relies on relationships that are themselvesdepending on the climate. For instance, reconstructing climatefrom pollen makes use of relationships (between properties ofpollen and parameters such as temperature and precipitation),which might have been very different in the past from what theyare today. If divergences between the model and the data arefound, they can be explained in many ways: model errors, aprocess poorly accounted for in modeling, but also errors in data,or moreover in data interpretation—or any combination of thesedifferent factors.Thus, contradictions between paleoclimatic simulations anddata has sometimes raised doubts about the latter: for example,an estimate of ocean surface temperatures during the last glacialmaximum was shown to be false after it happened to beincompatible with many models. The ‘‘reconciliation’’ betweenmodels and data is the result of dialogues and ‘‘iterative work’’between data specialists and modelers (interview with G,Laboratoire des Sciences du Climat et de l’Environnement, June2006)—this cooperation playing a notable part in paleoclimatology.The iterative dialogue between simulations and observationsis facilitated by the availability of powerful computers, whichmade it possible to make the two ‘‘speak the same language’’ to‘‘use them in the same manner’’ (according to the rule followed bysome modelers); in the examples given in Section 3, the commonlanguage is that of the diagnostic, of a method of analysis orinstrumental measurement. As in all sciences, here one obtainscharts, curves, and histograms—produced by the computer, andarising from simulations as from data—that constitute what iscompared, analyzed, and discussed, and what constitutes proof(Latour, 1989).The third characteristic noticed here proceeds from this backand forth: models are at once (or alternatively) validated byobservations and used to complete the data. They are the object toexplore and the tool of exploration; they inseparably serve tounderstand the climate and make predictions about it. Because ofthis, it is impossible to strictly distinguish a models constructionfrom its use. Scientists go back and forth between the developmentphase and the simulation phase, relying on the model tostudy and project the climate and on the climate to explore andameliorate the model, following a multitude of different modalities.(Heymann (2006) made a similar point for atmosphericchemistry modeling). If model development involves its utilization(even if just for the sake of testing), using the model generallyrequires participating in its construction, or at least sufficientlyknowing its content to be able to carry out sensitivity experimentsor studies of variability. Stated differently: the model is not(yet) a black box, a tool easily accessible to all users, a machine forproducing color images, in the words of a LMD scientist (whodreads such use of today’s models). The model remains an objectof research.There is a fourth point to raise. What is evaluated is thecapacity to reproduce or account for a phenomenon (themonsoon, for instance), or a climatic characteristic (for example,variability of precipitation). And this evaluation is never anisolated task: every combination of phenomena, available data,and urgent questions suggests certain methods of analysis tothe scientists’ imagination. A general protocol for validation doesnot exist—no more than a universal norm for experimentaltesting (Atten & Pestre, 2002). Norms of validation are definedat the same time as the facts to be validated, as well as theanalysis methods to which data and models will be subjected andthat allows for their comparison. We could say that theresearchers must construct the phenomenon to be validated.Evidently, the researchers do not construct the phenomenoncalled Indian monsoon; however, it is not the simulation ofthe monsoon ‘‘itself’’ that is under evaluation, rather the setof parameters characterizing the monsoon—that is effectivelyconstructed.Fifth remark: these modes of analysis, which bring outstructures, correlations, and retroactions in the models and thedata, are also what allows modelers to explain and understandphenomena by simplifying them. Evaluations also have a heuristicfunction. The computer’s combinatorial and sorting capacity, inallowing modelers to bring order to observations and to classifyclouds, helps triumph over tangle and multiplicity of data andovercomes (in part) the famous complexity of clouds and theirclimatic interactions. Here we recall Latour’s remark about thebotanist sorting leaves in her lab: ‘‘. . . a pattern emerges (. . .) andhere again, it would be astonishing if it were not the case.Invention almost always follows the new handle offered by a newtranslation or transportation’’ (Latour, 1999, p. 53).4.2. The question of realism in climate modelingThe last remark deals with a particularly delicate point, therealism of models, which deserve a longer exposition. Models donot aim to ‘‘imitate’’ the real climate (which would hardly makesense). Every model is incomplete, simplifying, and it is the resultof choices, definitions, and conceptions made by its authors.Modelers, as we have seen, seek to look at things according to newframes of references. Is there anything less realistic than cloudsthat are lined up as if ready for battle, or monsoons reduced totheir statistical characteristics? It is a matter of understandingclimatic mechanisms, increasing confidence in models andprojections of climate change, which does not entail producingrealistic representations. Simulations and observations must‘‘speak the same language’’, but this is not the language of thereal climate. 13 Even data do not constitute a faithful representationof the climate. The referent to which simulations arecompared is not ‘‘the climate’’, it is a set of remodeled data thatare carefully selected from the hundreds of thousands of data setsfurnished by instrument networks on land, sea, and in outerspace. Only an instrumented world is capable of providing thedata that are used to test models.Yet, the question of realism is at the heart of the image ofmodels that scientists present to policy makers and the public.13 That still recalls Latour’s analysis about pedology in Amazonia: ‘‘But theseacts of reference are all the more assured since they rely not so much onresemblance as on a regulated series of transformations, transmutations andtranslations’’ (Latour, 1999, p. 58) ‘‘It is not realistic; it does not resemble anything.It does more than resemble. It takes the place of the original situation, which we canretrace. . .’’ (Latour, 1999, p. 67)

250H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252Models are in part constituted from algorithms that reproducelaws of dynamics and thermodynamics and in part fromparametrizations that scientists continuously seeks to bring‘‘closer to the physics of the process’’. An increasing number ofphenomena and cycles that influence the climate are taken intoaccount. Spatial and temporal resolutions are improving due tomore powerful computers and means of observation, giving abetter geographical representation of the globe. It is becausecomputers allow ‘‘pragmatic interaction without theoreticalbackground’’ of model components (Küppers & Lenhard, 2007),because they allow all of these ‘‘formal heterogeneous systems’’(Varenne, 2007) to interact and work together, simulating theirtemporal development, that simulations have come to betterreproduce some climatic phenomena (for instantance, couplingwith vegetation helps explain strong Asian monsoon in theHolocene epoch). Progress in this domain seems to move towardsfiner resolution, detail, and proliferation of elements, for anincreasingly faithful representation of the climate. When climatescientists address the non-specialist public, they normally beginby presenting a sketch of this progress from primitive atmosphericmodels—bare globes with crude grids–from the 1970s totoday’s ‘‘earth systems’’ populated by all sorts of chemical andvegetal species, mountains, seas, cities, and rivers. This representationseem to be the exact opposite of that given by high-energyphysicists—the famous ‘‘quest for the ultimate particle’’: insteadof a reductionist race towards the elementary and towardsunification, climate modeling would advance towards a horizonof comprehensive representation of everything.The image is nevertheless too simple. While this path towardscomplexity is a serious trend in climate modeling, to which thestriving for realism is often linked, it is but one driving force ofmodeling, and it is not without problems. First, global EarthSystem models are not the only models used by researchers; theyare assisted by simpler models, ‘‘idealized’’ models, etc. Inaddition, there are scientific debates on the best means ofimproving models. Certain scientists think that we will betterunderstand and project the climate by improving parametrizationsof essential atmospheric phenomena as opposed to describingin detail various types of vegetation, while others think thatmodeling will benefit most from taking into account new factors.Other internal debates exist on the limits of GCMs, for example,on the question of to what point can these models be used on theregional scale to respond to growing social demand. Illustratingthe problematic character of the notion of realism, climatescientists use this term to designate a number of different things:they speak of the realism of a description of a process—of aparametrization, for instance—and of the realism of a simulation.Now, those two realisms are often initially incompatible (I evokedthe difficulty of implementing new parametrizations earlier)—thechoice to privilege one or the other depends on the objective ofmodeling, as well as on the culture and institutional frameworkof the laboratory (Shackley, 2001). The researchers’ work consistsof defining what counts, which cannot be reduced to anineluctable advance towards complexity or maximum resolution.However, this scientific work, and the question of realism inparticular, cannot be isolated from the political stakes involvedin climate modeling (Shackley et al., 1999). Climate projectionsrely on models that are expected to provide to policy makersincreasingly better representations of all climatic elements inorder to produce a higher quality and more precise projection,notably on the local scale. It is necessary to discuss the naivevision of models, to cast doubt on their realism, to underscoretheir limits and to challenge that conception of science–policyrelationship based on the ‘‘linear model’’ (Pielke, 2002; Sarewitz,2000), even if it can pose the risk being confused with (or used by)the opponents of climate change policy.4.3. Modeling and field science versus theory and experiment-basedsciences: similitudes and differencesTo what extent are the preceding epistemological remarksspecific to the relation between numerical models and observations?Could the same analysis be carried out in regards to therelationship between theory and experimental results? That dataare reconstructed, that validation norms are defined at the sametime as that which is to be validated, that sciences do not producemimetic images of reality, all of this results from very generalfindings. On the level of field observations, there do not seem tobe differences in principle between the older classical sciencesand climate modeling. Of course, differences in size and scaleexist, since the climate sciences depend on a veritable (technical,metrologic, administrative. . .) ‘‘mobilization of the world’’(Latour, 1989, p. 539), which has allowed for the developmentand maintenance of a dense and complex measurement network.Numbers representing objects and properties involved in climate(meteorological measurements on the ground, cloud measurementsfrom satellites, water measurements from ships, characteristicsof vegetation. . .) circulate, are concentrated andcombined, and this information coming from all over the worldis made available to models. In order to master the avalanche ofdata inside this gigantic ‘‘center of calculations’’, it is stillnecessary to classify, superimpose, invent new transformationsthat will reduce the number of these objects, summarize theirrelationships and transcribe them to our scale. Computationpermits these manipulations at a grand scale, and allowsfor virtually transforming the climate and experimenting onobservations. The change is by degree: the computer prolongs anold history, offering increased computational power and superimpositioncapacity (Latour, 1989). The power of computerscience endows models and data with unprecedented plasticity(Pickering, 1995) and allows for new connection between them.As we have seen, numerical computation has made possiblethe modeling of a system containing many heterogeneouselements and processes that interact nonlinearly, but in doingthis it has also made validation particularly arduous. A similarunderdetermination is found in observations: climate scientistshave at their disposal and abundance of data collected in massivequantities, which are at the same time more global, plural, andincomplete than experimental measurements. If experiments arecarefully tailored to testing a theoretical hypothesis, observationsin the climate domain are rather ‘‘ready to wear’’: it is necessaryto choose, select, and even adjust if needed. How is it possible tomove past this double underdetermination of models andobservational data? How can the infinite potentialities of themodel be confronted with the infinite realities of the field? Thefew examples we have seen have given an idea of it: by choosing adomain, a definition, a series of questions, and a mode of analysisthrough which the model’s and the real climate’s functioning canbe studied at once.More fundamentally, there is a difference between therelations that link theories and experiments on the one hand,and modeling and field observations on the other. All scientificdisciplines share the requirement of testing that permit to decidebetween statements. An experimental statement benefits from itscapacity to resist controversy by demonstrating that it is not asimple fiction, the instrument having been constructed preciselyso that ‘‘the experimental fact would be explained by the answerto the question posed’’ (Stengers, 1993, p. 154). This is not thecase for numerical modeling because models have to do preciselywith fiction. Like fictions, and unlike theories, models have thecapacity to ‘‘express content not immediately situated in thedimensions of true or false’’ (Barberousse and Ludwig, 2001, p. 23)they can ‘‘at once contradict accepted hypotheses and tell us

H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252 251something about the world’’ (Barberousse and Ludwig, 2001, p.36). No model has exclusivity on representing a phenomenon,since many representations coexist that meet different needs andthat are sometimes linked to choices made by the author. 14 As faras field observations are concerned, it is impossible to prove withcertainty the stability of established relationships to their subjects(this is one of the difficulties of validating simulations of futureclimate and interpreting data from the past). A valid explanationfor one terrain will not necessarily hold in another (it oftenhappens that a parametrization validated for one region of theglobe is not suitable for another because the dominant processeswere not identical). Situations cannot be purified, ‘‘no single causehas the general power to cause, each one is part of a history and itis from this history that it gets its identity as a cause’’ (Stengers,1993, p. 159). Any validation of a model with data from the fieldhas to take into account these multiple causes that vary in timeand space.5. ConclusionThe central importance of observation for modeling cannot beoveremphasized. There has been a parallel evolution betweenprogress made in modeling and the collection of larger and largerdata quantities, by satellites in particular. If climate modeling hasadvanced to a degree unparalleled in the environmental sciences,this is due to the exponential growth in computational power andto resolution of equations governing atmospheric circulation, butalso due to a unique and dense instrument network that is highlydeveloped and organized on an international scale.Let us return to the questions posed by the philosophersevoked at the beginning of this article. If models do not directlyrepresent the real world, if their numerical experiments are notabout the world, how can they bring knowledge about reality?What gives models their credential? What guarantees knowledgeobtained from simulations? We have looked for the responses inresearchers’ practices. What establishes scientists’ trust in theirown models, what guarantees the value of the knowledge theyproduce, is validation or evaluation by observational data, muchmore so than their relation to theories or experiments.Relationships to theories undoubtedly constitute one of theoriginalities of climate modeling, and distinguish it from othermodes of knowledge production. Its capacity to integrate heterogeneoustheories, milieus and elements, to span scales of time,space and complexity, to combine the singular and the contingent,along with its faculty for simulation, taking into accounthistorical evolution and fictional scenarios—all of this constitutesnew ways of understanding scientific knowledge peculiar tonumerical modeling. On the other hand, in its relations to the real(instrumented) world, climate modeling resembles other scientificdisciplines, thereby establishing its credentials and reliability.It is these relationships between models and observations, theselanguages invented by scientists and common to simulations anddata, these multiple original and robust links, that fasten climatemodeling to the material world, and that contribute to theimprovement of models and scientists’ understanding of climaticphenomena.ReferencesArmatte, M., & Dahan, A. (2004). Modeles et modélisations (1950–2000):Nouvelles pratiques, nouveaux enjeux. Rev Hist Sci, 57/2.14 In climate models, parametrizations are often identified by their author’sname—we hear of ‘‘convection pattern from X’’ or ‘‘boundary layer scheme of Y’’.Atten, M., & Pestre, D. (2002). Heinrich Hertz. L’administration de la preuve. Paris:PUF.Barberousse, A., & Ludwig, P. (2001). Les modeles comme fiction. Philosophie, 68,16–43.Bengtsson, L., et al. (2004). Can climate trends be calculated from re-analysis data?J Geophys Res, V109.Bony, S., et al. (2004). On dynamic and thermodynamic compônents of cloudchanges. Clim Dyn, 22, 71–86.Cartwright, N. (1983). How the laws of physics lie. Clarendon Press.Cartwright, N. (1999). Models and the limits of theory: quantum Hamiltonians andthe BCS model of superconductivity. In M. S. Morgan, & M. Morrison (Eds.),Models as mediators: perspective on natural and social science. Cambridge:Cambridge University Press.Chaboureau, J.-P., & Bechtold, P. (2002). A simple cloud parameterization derivedfrom cloud resolving model data: Diagnostic and prognostic applications. JAtmos Sci, 59, 2362–2372.Dahan Dalmedico, A., & Guillemot, H. (2008). Climate change: scientific dynamics,expertise and political challenges. In C. Paradeise (Ed.), Science and Soverainty.New York: Routledge.Edwards, P. (1996). Global comprehensive models in politics and policy making.Climatic Change, 32(2), 149–161.Edwards, P. (1999). Global climate science, uncertainty and politics : data-ladenmodels, models-filtered data. Sci Culture, 8(4).Edwards, P. (2006). Meteorology as infrastructure globalism. Osiris, 21, 229–250.Forsyth, T. (2003). Critical political ecology: the politics of environmental science. NewYork: Routledge.Galison, P. (1987). How experiments end. Chicago: University of Chicago Press.Galison, P. (1996). Computer simulations and the trading zone. In P. Galison, & D.Stump (Eds.), The disunity of science: boundaries, context and power. Stanford:Stanford University Press.Guillemot, H. (2007). Les modeles numériques de climat. In A. Dahan (Ed.), LesModeles du futur. Paris: La Découverte.Hacking, I. (1989). Representing and intervening. New York: Free Press.Hacking, I. (1992). The self-vindication of the laboratory sciences. In A. Pickering(Ed.), Science as practice and culture. Chicago: University of Chicago Press.Heymann, M. (2006). Modeling reality. Practice, knowledge, and uncertainty inatmospheric transport simulation. Hist Stud Phys Biol Sci, 37(1), 49–85.Heymann, M. (2009). Klimakonstruktionen: Von der klassischen Klimatologie zurKlimaforschung. NTM. Int J Hist. Ethics Nat Sci Technol Med, 17(2), 171–197.Kalnay, et al. (1996). The NCEP/NCAR 40-year reanalysis project. Bulletin of theAmerican Meteorological Society 448.Kandel, R. (2002). Les modeles météorologiques et climatiques. In P. Nouvel (Ed.),Enquête sur le concept de modele. Paris: PUF.Knorr-Cetina, K. (1992). The couch, the cathedral and the laboratory: on therelationship between experiment and laboratory in science. In A. Pickering(Ed.), Science as practice and culture. Chicago: University of Chicago Press.Knuuttila, T., et al. (2006). Computer models and simulations in scientific practice(Editorial). Sci Stud, 19(1), 3–11.Küppers, G., & Lenhard, J. (2007). From hierarchical to network-like integration: arevolution of modeling style in computer simulation. In J. Lenhard, G. Küppers,& T. Shinn (Eds.), Simulation: Pragmatic Construction of Reality, Sociology of theSciences, vol. 25). Dordrecht: Springer.Kuhn, T. (1977). The essential tension: selected studies in scientific traditions andchanges. University of Chicago Press.Latour B. (1989). La science en action (réédition Gallimard 1995).Latour, B. (2001). L’espoir de Pandore. Paris: La Découverte. English version (1999):Pandora’s hope. Essays on the reality of science studies. Harvard University Press.Latour, B. (2004). Why has critique run out of steam? From matters of fact tomatters of concern. Crit. Inquiry, 30(2), 25–248.Lenhard, J., & Winsberg, E. Holism, entrenchment and the future of climate modelpluralism, this issue, doi:10.1016/j.shpsb.2010.07.001.Le Treut, H. (1999). Le défi de la modélisation du climat. In: INSU, 30 ans derecherche en sciences de l’Univers 1967–1997, 1997–1999, Lettre du ProgrammeInternational Géosphere–Biosphere-Programme Mondial de Recherche sur leClimat (IGBP-WCRP) 9.Lewis, J.-M. (1998). Clarifying the dynamics of the general circulation: Phillips’s1956 experiment. Bull Am Meteorol Soc, 79, 39–60.LMD (2005). Instruction manual for atmosphere global circulation model. /http://web.lmd.jussieu.fr/lmdz/manuelGCM/main/node3.htmlS.Miller, C. (2001). Scientific internationalism in American foreign policy: The case ofmeteorology, 1947–1958. In P. Clark Miller (Ed.), Changing the atmosphere:expert knowledge and environmental governance. Edwards: MIT Press.Moisselin, P., et al. (2002). Les changements climatiques en France au XXe siecle.Étude des longues séries homogénéisées de données de températures et deprécipitations. La Météorologie, 38, 45–46.Morgan, M. S., & Morrison, M. (Eds.). (1999). Models as mediators: perspective onnatural and social science. Cambridge: Cambridge University Press.Nebeker, F. (1995). Calculating the weather: meteorology in the 20th century.International geophysics series, 60. Academic Press.Norton, S., & Suppe, F. (2001). Why atmospheric modeling is good science. In C.Miller, & P. Edwards (Eds.), Changing the atmosphere: expert knowledge andenvironmental governance. Cambridge: MIT Press.Oreskes, N., et al. (1994). Verification, validation, and confirmation of numericalmodels in the earth sciences. Science, 263, 641–646.Pickering, A. (1995). The Mangle of Practice. Time, Agency and Science. University ofChicago Press.

252H. Guillemot / Studies in History and Philosophy of Modern Physics 41 (2010) 242–252Pielke, R., Jr. (2002). Policy, politics and perspective. Nature, 416, 368.Randall, D. A., et al. (2007). Climate Models and Their Evaluation. In Solomon (Ed.),Climate change 2007: the physical science basis. contribution of working group i tothe fourth assessment report of the IPCC. Cambridge: Cambridge UniversityPress.Rochas, M., & Javelle, J.-P. (1993). La météorologie. Paris: éditions Syros.Sarewitz, D. (2000). Science and environmental policy: An excess of objectivity. InR. Frodeman (Ed.), Earth matters: the earth sciences, philosophy and the claims ofcommunity (pp. 79–98).Shackley, S., & Wynne, B. (1996). Representing uncertainty in globate climatechange and policy : Boundary ordering devices and authority. Sci Technol HumValues, 21, 3.Shackley S. et al. (1999). Adjusting to policy expectations in climate changemodeling: an interdisciplinary study of flux adjustments in coupled atmosphere–oceangeneral circulation models, MIT Joint Program Science and Policyof Climate Change, Report no. 48, mai 1999.Shackley, S. (2001). Epistemic lifestyles in climate change modeling. In P. ClarkMiller (Ed.), Changing the atmosphere: expert knowledge and environmentalgovernance. Edwards: MIT Press.Sismondo, S. (1999). Models, simulations and their objects. Sci Context, 12(2), 247–260.Stengers, I. (1993). L’invention des sciences modernes. Paris: La Découverte.Sundberg, M. (2007). Parameterizations as boundary objects on the climate arena.Soc Stud Sci, 37, 473–488.Van der Sluijs, J., et al. (1998). Anchoring devices in science for policy: the case ofconsensus around climate sensitivity. Soc Stud Sci, 28(2).Varenne, F. (2003). La simulation conc-ue comme expérience concrete. In J. P. Müller(Ed.), Le statut épistémologiquedelasimulation,éditions de l’ENST (pp. 299–313).Varenne, F. (2007). Du modele a la simulation informatique. Paris: Vrin.Winsberg, E. (1999). Sanctioning models: the epistemology of simulation. SciContext, 12, 2.Winsberg, E. (2003). Simulated experiments: methodology for a virtual world.Philos Sci, 70, 105–125.