the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 03 Sep 2026
How models represent target systems in the world, and how we (need to) communicate this
Abstract. Humans are a representing species: we construct conceptual, numerical and material models that represent the world in certain respects. However, how scientists communicate about how their models represent the world, sometimes reflects that it is unclear by what virtue models can represent and can be used to make claims about the world. This becomes most apparent and problematic when models are used to make claims about a target system in communications between scientists, the public, and consultants.
After introducing four example models of different kinds from one geoscientific discipline, recent accounts of model representation from philosophy of science will be summarized and exemplified. Some of these centralize the representational relation between models and their target phenomena, whereas others centralize the role of the modellers in the process of using the model as representation for the target. The latter group of accounts convincingly demonstrate the role of the scientist, their skills and their background knowledge in the representation. This works in both ways: the modellers and other experts put trust in their models, not because the models are perfect, but because they have background knowledge and experience. However, the receiver, with different expertise, does not have this.
Following examples from the geoscientific literature, suggestions emerge how more nuanced claims can be made with models, and which background assumptions on both the modelling and the target systems, and the role of required skills, need to be explicated. While scientists may worry that putting a scientist in the picture could reduce the trust in their model, for effective communication the target audiences need understanding at a level that links to their background and training to be able to evaluate the model results and integrate these with their own experience and knowledge.
- Preprint
(784 KB) - Metadata XML
-
Supplement
(33 KB) - BibTeX
- EndNote
Status: open (until 30 Oct 2026)
-
EC1: 'Ed query about egusphere-2026-4944', David Crookall, 04 Sep 2026
reply
-
CEC1: 'Reply on EC1', John K. Hillier, 16 Sep 2026
reply
Publisher’s note: the content of this comment was removed on 30 September 2026 since the comment was posted by mistake.
Citation: https://doi.org/10.5194/egusphere-2026-4944-CEC1
-
CEC1: 'Reply on EC1', John K. Hillier, 16 Sep 2026
reply
-
RC1: 'Comment on egusphere-2026-4944', Anonymous Referee #1, 08 Oct 2026
reply
Review of
“How models represent target systems in the world, and how we (need to) communicate this”
by M. G. Kleinhans.
Submitted to Geoscience Communication.
Summary
The author argues that geoscientists communicate about models as if correspondence between model and target (dynamic similarity or isomorphism) suffices to justify inferences about the target. Drawing on agent-centered accounts of scientific representation (DDI, DEKI), he argues that the modeler’s purpose, skill, and background knowledge are part of the representational relation. Four estuary models illustrate the accounts. An informal survey of 52 papers in EGU, AGU, and Copernicus journals is offered as evidence that modeler agency is rarely communicated, and two anecdotes from science-policy settings illustrate the consequences. The paper concludes that modelers should state the respects and degrees in which a model represents its target and should make their background knowledge explicit, both within science and when communicating with other audiences.
The topic is important and within the scope of Geoscience Communication. Introducing the philosophy of scientific representation to geoscientists in accessible form is worthwhile, and the author’s willingness to critique his own papers is commendable. The writing is lively.
There are four main problems. First, the empirical claim of Section 4.1 rests on an informal, single-coder survey with undefined criteria, and the supplementary table does not consistently support the text. Second, the recommendations are too general to act on. Third, relevant literature is missing, notably from climate modeling, where modeler judgment in tuning has been documented explicitly. Fourth, several conceptual statements are inconsistent or inaccurate.
In view of these problems, I recommend major revisions.
Major comments
1. The survey of 52 papers (Section 4.1 and supplement)
This is the only systematic evidence offered for the central claim that modeler agency is rarely communicated, so it needs to be sound.
Sampling. The selection rules differ by journal: GMD “highlight articles” excluding a digital twin and scenario construction; JAMES “most recently published”; keyword searches elsewhere; in ESurf, excluding repeat authors and the author’s own papers. Publication years range from 2006 to 2026, and Climate of the Past contributes two papers, from 2008 and 2014. None of this is described in the manuscript. The protocol belongs in the text.
Coding. The column headings (“model reality”, “explicit key in represent”, “benchmark / data”, “modeller agency”) are not defined, and there is one coder with no reliability check.
Text and table disagree. Line 299 says all papers address shortcomings; the table codes 37 of 52. Lines 305–306 say all numerical papers use isomorphism and all experimental papers use similarity, but Eltner (a cellular automaton) and Gonzalez-Munos (analog experiments) are coded for neither. Lines 308–309 say direct application to the target is more frequent in experiments; the table gives 7 of 13 (54%) versus 17 of 40 (43%), a weak difference at n = 13.
Errors in the table. Six papers have the value 2 in the benchmark column, so the summary fraction in row 4 (0.77) overstates the share of papers with benchmarks (34 of 52, or 0.65). The link for Ma et al. points to the JAMES home page rather than the article.
The agency coding appears too strict, and the quotations in the table contradict it. Sockness et al. write that “the model was not designed to be a scale model of those areas. Instead, it was designed to be a process model…”. That is precisely the statement of purpose and of representational respects that the Conclusions (lines 373–375) recommend, yet it is coded as showing neither agency nor an explicit key. Other examples: “We selected C = 2 to achieve sufficient stabilization” (Gärtner); “we found it necessary to use a line search algorithm” (Popov); “we assume no infiltration” (Gauvain); “their reliability in extrapolation… has not yet been convincingly demonstrated” (Coucheney). These are first-person statements of modeler choices and limits. If the author means something narrower, such as disclosure of the background knowledge on which trust rests, he should define it and recode. Otherwise the statement that “only two papers showed minor signs of communicated author agency” (line 313) is not supported.
Some entries are not models by the paper’s own definition (lines 72–73 and 140–152: a model permits intervention). The table itself notes “[not really a model to intervene on]” for Tarling; Pantaleo is a “sketch (not to scale)”; de Nardis is a structural interpretation of tomography. Exclude or justify.
Unattributed quotations (lines 251–253, “coastal modelling papers best not referenced”). Readers cannot check these, which sits awkwardly in a paper about transparency. Use quotations from the documented sample instead.
The author should choose between two options: present the survey as illustrative only and remove the quantitative statements, or turn it into a content analysis with a stated sampling procedure, a codebook, a second coder, and reported agreement. The second option would make the paper considerably stronger.
2. Missing literature
The DDI account originates with Hughes (1997), who is not cited. Figure 2 and Section 3.2 should cite him. For DEKI, Frigg and Nguyen (2020) is the fuller source.
Climate modeling offers direct evidence on the paper’s question. Mauritsen et al. (2012), Hourdin et al. (2017), and Schmidt et al. (2017) document how modelers tune global models, which targets they tune to, and why, and argue that this should be disclosed. These papers show that at least one modeling community does communicate modeler agency explicitly, and they show what that communication looks like. They also discuss the risk, noted in the abstract, that disclosure can be used to undermine trust. The survey seems to have missed this body of work.
Parker (2020) and Baumberger et al. (2017) on adequacy for purpose and confidence building are directly relevant to “in which respects and degrees”.
The sociology of trust in models complicates the paper’s premise that modelers know the limitations and outsiders lack the background (lines 16–18 and 247). MacKenzie’s “certainty trough” (1990), applied to climate modelers by Lahsen (2005), found that modelers themselves sometimes conflate model and world, while users at an intermediate distance are often the most confident. This deserves discussion.
Knowledge in modeling is distributed across communities and institutions (Edwards, 2010; Winsberg, 2010). Trust is built through intercomparison projects, shared code, documentation, and observational benchmarks, not only through individual expertise. The Ems-Dollard example supports this: trust was built “over considerable time by a large community” (lines 364–365). The claim that knowledge is carried “by the scientists” (line 350) should be broadened accordingly.
Model hierarchies (Held, 2005; Jeevanjee et al., 2017) are a widely used way of making explicit which respects of a target a model is meant to capture. The four estuary cases form a hierarchy of this kind, and the connection would help the reader.
On communicating model limits to non-specialists: Thompson and Smith (2019) and Saltelli et al. (2020).
Box is quoted but not cited (Box, 1976, 1979). The von Neumann quotation needs a source. Sterrett (2002) is cited but missing from the reference list. Suárez (2003) and Frigg and Hartmann (2020) are listed but not cited.
3. The recommendations
The Conclusions advise indicating “precisely in which respects and degrees a model represents a target” and communicating background knowledge. A reader of Geoscience Communication will want to know how. I suggest a short checklist or table that a modeler could apply in a paper or a briefing: the purpose of the model; whether the target is a specific system or a kind; which model properties are claimed to carry over to the target, and by what key; what was calibrated or tuned, and against what; what background knowledge (field data, geological reconstructions, earlier experiments) supports the inference; and known misrepresentations. Applying it to one of the four cases, with a revised version of one of the quoted claims, would show the reader what better practice looks like.
The abstract raises the concern that showing the modeler’s role may reduce trust, but the paper does not discuss this. Akerlof et al. (2012) and the tuning literature above suggest this concern is real in politicized settings. Weisberg et al. (2021) addresses it only in part. The trade-off should be discussed.
4. Conceptual points
Definition of a model. Lines 140–152 state that figures, paintings, and maps are not models because they do not permit intervention. Yet Figure 1 presents image data (panel a) and a cave painting (panel b) as “model representations of lions”, and lines 34–35 call the cave paintings “masterful modelling”. A causal loop diagram (de Haas et al., 2018) permits intervention only in thought, as would a map. The boundary should be drawn consistently, or the intervention criterion dropped.
Isomorphism in numerical models (lines 177–190). The statement that a model represents the target “by having an object for each object in the target” does not hold for most geoscience models. Discretization and parameterization of unresolved processes break any one-to-one correspondence: a parameterization represents the statistical effect of many unresolved objects. Partial isomorphism or homomorphism (da Costa and French, 2003) is closer. This point strengthens the author’s thesis, because parameterization is where much modeler judgment enters. Also, “there is nothing essentially physical about these models” (line 179) overstates the case; the equations encode conservation laws.
Scaling of mantle flow (lines 163–165). Analog models of mantle convection are not scaled by the Reynolds number. Inertia is negligible (Re ≪ 1), and the governing parameter is the Rayleigh number, with a very large Prandtl number. Please revise.
Relation between the two groups of accounts. Lines 236–238 note that agent-centered accounts still require the modeler to have good reasons to think the key captures relevant similarity or isomorphism. This is an important point and should come earlier. The abstract and Section 3 present the two groups as alternatives; it would be clearer to say that agent-centered accounts add to correspondence rather than replace it.
Box (lines 27–33). The assertion that Box’s statement is “wrong in both parts” is rhetorical, and the supporting argument restates Box’s own point. Box (1976) wrote that “since all models are wrong the scientist must be alert to what is importantly wrong”, which is close to the author’s message. The real target is the use of the aphorism as a platitude that deflects scrutiny of the respects in which a model is wrong. Also, the claim that all models are useful because building them is a learning experience conflates usefulness to the modeler with usefulness for inference about the target.
Passive voice (lines 258–261). The term is “passive voice”, not “passive tense”. Of the two quoted examples, only the first is passive; “The simulations show…” is active voice with a nonhuman subject. The issue is agentless phrasing, not voice. The hypothesis is plausible but untested, and readers will expect evidence or a citation. The manuscript itself alternates between “we” (lines 128 and 192) and “I”; consistent first person would model the advice.
Statues (lines 335–338). The analogy is tangential, and the claim about how Greeks and Romans regarded cult statues is contested among classicists. Some readers may also find the religious comparison off-putting. I suggest removing it.
Digital twins (lines 147–152). As usually defined (e.g., Bauer et al., 2021), a digital twin of the Earth system contains a model with data assimilation, so the remark targets a straw version. Calling the term an oxymoron adds little. Shorten.
5. Figures
Figure 1: panel (f) is labeled a numerical model but shows LEGO figures. The text (lines 51–52) calls it a “visualization of a numerical model”; explain what model is visualized. The description of panel (e) at lines 50–51 (“here representing linear stability analysis of periodic pattern formation in their initialization stage”) is obscure. Panel (e) is adapted from Pieri et al. (2015), a Wiley journal; please confirm reuse permission. Panel (a) has no credit.
The lion figure is engaging, but it does not show the reader how the four estuary cases map onto the DDI and DEKI elements. A table or figure that does this (target, denotation, what is demonstrated, the key, what is imputed, and where modeler judgment enters) would make the argument concrete.
Figure 2: cite Hughes (1997) in the caption.
6. Disclosure
The disclosure at line 263 of involvement in two of the four cases is welcome. The author’s role as “intermediator” in the Ems-Dollard program (line 368) should also be noted in the Competing interests section.
Minor comments
L29. Give a URL or citation for the GMD aims and scope.
L76–77. “on which models are only partly be based”: delete “be”.
L83–86. The point-source delta example is unclear. A point source is a river mouth; “sediment does not rain out of the sky” does not explain what is unrealistic. Presumably it is radial progradation in an unconfined basin from a fixed source.
L89–90. This sentence is garbled; please rewrite.
L113. “The models is”: “The model is”.
L121–122. “added” appears twice.
L128. “Here we employ”; elsewhere “I”. Be consistent.
L132. “correspondence relation between the target system in the world”: insert “the model and”.
L143. Baar et al. (2022) should be 2023.
L157–158. The cat and lion bone example needs a sentence of explanation (allometric scaling).
L162. Name the Shields number.
L167–168. Name the classic example (Froude versus Reynolds similarity).
L173–175. “dogmatism”: support with a citation or soften.
L175–176. Cite a source for analog modeling being on an equal footing with numerical modeling in tectonics.
L226. “between the mud flats and saltmarshes in estuaries”: “between the model materials and the mudflats and saltmarshes”.
L236. “model representation… do not”: “does not”.
L277. “Haas et al.”: “de Haas et al.”.
L281. “Kleinhans et al. (2022) goes”: “go”.
L285. “the reconstructions is”: “are”.
L302. “Soil science”: “SOIL”.
L305. “Formulations how”: “Formulations of how”.
L339. Comma after “on the other hand”.
L385. “pertained communication”: “pertained to communication”; “communicating model”: “communicating models”.
References. The Li et al. (2026) URL is a university proxy; use https://doi.org/10.1029/2026MS005737. For Morgan and Morrison (1999), cite the chapter (Morrison and Morgan, “Models as mediating instruments”). Add DOIs where available (Gupta, Pieri, Sterrett, Suárez 2003, Weisberg et al. 2021).
Supplementary table
Add a legend that defines each column and the coding rules.
Add columns for publication year and model type (numerical, analog, conceptual, field or mesocosm experiment), since the text makes claims by type.
Explain or correct the values of 2 in the benchmark column, and recompute row 4. Label row 4.
Fix the Ma et al. link.
For each paper coded for agency (and for borderline cases such as Sockness), give the passage that justifies the code.
Recommendation
Major revision. The paper makes a worthwhile argument, but the evidence in Section 4.1 must be put on a firmer footing or presented as illustrative only, the recommendations should be made concrete, and the literature on climate model tuning and on trust in models should be engaged.
References cited in this review
Baumberger, C., Knutti, R., and Hirsch Hadorn, G.: Building confidence in climate model projections: an analysis of inferences from fit, WIREs Clim. Change, 8, e454, 2017.
Bauer, P., Stevens, B., and Hazeleger, W.: A digital twin of Earth for the green transition, Nat. Clim. Change, 11, 80–83, 2021.
Box, G. E. P.: Science and statistics, J. Am. Stat. Assoc., 71, 791–799, 1976.
Box, G. E. P.: Robustness in the strategy of scientific model building, in: Robustness in Statistics, edited by Launer, R. L. and Wilkinson, G. N., Academic Press, 201–236, 1979.
da Costa, N. C. A. and French, S.: Science and Partial Truth, Oxford University Press, 2003.
Edwards, P. N.: A Vast Machine: Computer Models, Climate Data, and the Politics of Global Warming, MIT Press, 2010.
Frigg, R. and Nguyen, J.: Modelling Nature: An Opinionated Introduction to Scientific Representation, Springer, 2020.
Held, I. M.: The gap between simulation and understanding in climate modeling, Bull. Am. Meteorol. Soc., 86, 1609–1614, 2005.
Hourdin, F., et al.: The art and science of climate model tuning, Bull. Am. Meteorol. Soc., 98, 589–602, 2017.
Hughes, R. I. G.: Models and representation, Philos. Sci., 64, S325–S336, 1997.
Jeevanjee, N., Hassanzadeh, P., Hill, S., and Sheshadri, A.: A perspective on climate model hierarchies, J. Adv. Model. Earth Syst., 9, 1760–1771, 2017.
Lahsen, M.: Seductive simulations? Uncertainty distribution around climate models, Soc. Stud. Sci., 35, 895–922, 2005.
MacKenzie, D.: Inventing Accuracy: A Historical Sociology of Nuclear Missile Guidance, MIT Press, 1990.
Mauritsen, T., et al.: Tuning the climate of a global model, J. Adv. Model. Earth Syst., 4, M00A01, 2012.
Parker, W. S.: Model evaluation: an adequacy-for-purpose view, Philos. Sci., 87, 457–477, 2020.
Saltelli, A., et al.: Five ways to ensure that models serve society: a manifesto, Nature, 582, 482–484, 2020.
Schmidt, G. A., et al.: Practice and philosophy of climate model tuning across six US modeling centers, Geosci. Model Dev., 10, 3207–3223, 2017.
Thompson, E. L. and Smith, L. A.: Escape from model-land, Economics, 13, 20190040, 2019.
Winsberg, E.: Science in the Age of Computer Simulation, University of Chicago Press, 2010.Citation: https://doi.org/10.5194/egusphere-2026-4944-RC1 -
EC2: 'Thank you for RC1', David Crookall, 08 Oct 2026
reply
Thank you for an excellent and thorough review. Much appreciated.
Citation: https://doi.org/10.5194/egusphere-2026-4944-EC2 -
AC1: 'Reply on EC2', Maarten Kleinhans, 08 Oct 2026
reply
Yes completely agree. Thank you very much for an inspiring and thorough review! I will respond in detail later.
Citation: https://doi.org/10.5194/egusphere-2026-4944-AC1
-
AC1: 'Reply on EC2', Maarten Kleinhans, 08 Oct 2026
reply
-
AC2: 'Reply on RC1', Maarten Kleinhans, 10 Oct 2026
reply
I thank the reviewer RC1 for a very helpful and thorough review. In this reply I only address the main comments and the four problems identified by RC1. The minor comments will be dealt with in the revision.
- Literature survey: RC1 proposes two options: present the survey as illustrative only, or turn it into a content analysis. While RC1 prefers the second option, I will likely go for the first. I agree with the reviewer that at present the literature survey is not thorough but at best illustrative. The difficulty I am facing with this paper is that I am seducing geoscientists to read something about philosophy of science (PoS) that I think is highly relevant for practice, while very little of all the useful work in PoS arrives in the geosciences. That means the paper needs to be very accessible and readable, and therefore concise. I am glad that the reviewer agrees this is worthwhile and appreciates the lively writing. I agree that most points of the reviewer will improve precision in such a way that it does not come at the cost of clarity and accessibility for this readership.
- Missing literature: I am aware of, and read many of, the bodies of literature in PoS and that related to climate modelling. The present selection of references was made for its accessibility with the intended audience of this paper in mind. I will carefully study the papers and books that RC1 suggests and include more of them. I think I prefer to write a bit more about the sociology of trust in models and the communication of model limitations to non-specialists. I submitted to this journal Geoscience Communication because I have the experience that a better understanding of the philosophical issues around model representation will help both the doing of geoscience and the communication in the context of societal issues.
- Recommendations: I thank RC1 for the excellent suggestion to make these concrete with a checklist applied to the four main cases.
- I agree with all conceptual points made by RC1 and will clarify these mostly as proposed. I will expand on the Box statement, while maintaining the rhetorical assertions. I must do so because Box is so frequently cited and because I experienced in many interactions that geoscientists actually believed in it, or did not continue to think about model representation, after citing Box.
- It is a good idea to make a table or figure to map how the four estuary cases map onto the DDI and DEKI elements and I will probably expand figure 2 to do this.
Citation: https://doi.org/10.5194/egusphere-2026-4944-AC2
-
EC2: 'Thank you for RC1', David Crookall, 08 Oct 2026
reply
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 301 | 106 | 57 | 464 | 66 | 47 | 43 |
- HTML: 301
- PDF: 106
- XML: 57
- Total: 464
- Supplement: 66
- BibTeX: 47
- EndNote: 43
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Publisher’s note: the content of this comment was removed on 30 September 2026 since the comment was posted by mistake.