acceptodds
Under review as a conference paper at ICLR 2027

Representational Similarity Across AI Weather Models

Abstract

AI weather models produce skilful forecasts competitive with traditional Numerical Weather Prediction, but it remains unclear whether these independently developed models learn similar internal representations. We investigate representational similarity across six AI weather models, which comprise three variants of WeatherNext 1 Graph (GraphCast), AIFS-CRPS, Aurora and ArchesWeather, using five similarity metrics which include CKA and orthogonal Procrustes. Spatial and temporal variability in the atmosphere is often dominated by climatological latitude and seasonal variations that obscure the effects of synoptic weather systems. To account for this, we construct a null calibration of the metrics that preserves the large-scale structure, making the metrics more responsive to similarity in the representation of weather systems. We find non-zero calibrated similarity between all model pairs, strongest and most consistent within the GraphCast family, while rankings among the other models depend on the metric and spatial data used. Global CKA relationships remain stable throughout the year, whereas regional comparisons show greater seasonal variability, particularly in the midlatitudes. We also assessed representational similarity across processor depth in individual models. GraphCast and AIFS-CRPS representations evolve relatively smoothly, whereas Aurora and ArchesWeather change more strongly due to processor-layer boundaries. Together, these results suggest partially shared representations of atmospheric states, whose organisation and construction differ across models, and demonstrate the importance of calibrated, multi-metric analysis.

Primary area:
interpretability and explainable AI
Keywords:
AI weather forecasting, representational similarity, interpretability, neural network representations, Centered Kernel Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UfSEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper ask whether distinct AI weather models produce similar internal representations. The authors use five different similarity scores, representing five different metric families, and null calibrate each one to remove background variation. They then compare 6 AI weather models under these similarity scores and find moderate to strong similarity between the different models. This effect is most pronounced within model families, e.g. between graphcast variants, and within specific latitude bands such as the midlatitudes, and is consistent across the five metrics.

Critical strengths
  1. This is an important and timely topic due to the recent proliferation of AI weather models. It could be useful to understand whether the models are learning similar (unresolved) processes or whether they are learning something very different. In which case, we may need to combine their strengths to further improve performance. Moreover, if representational similarity implies prediction correlation then high similarity might cause multi-model ensembles to severely underestimate predictive variance. If these models are used to estimate risk, then poor calibration would invalidate their risk estimates.
Critical weaknesses
  1. This work needs to more clearly motivate its problem and interpret its results in greater depth. In particular, Sections 1 or 2 need an clear explanation on why model similarity is important, what representation similarity captures that functional similarity does not, and why these metrics were chosen and what, specifically, they capture and how they differ. Moreover, it needs to establish if we should expect models to be similar or not and whether this indicates good or bad behavior across models. The figures in Sections 4 and 5 show that many AI weather models have similar internal representations, but the following analyses do not investigate why they are similar or what this implies. For example, could this mean all models share systematic biases? Could they all be under or over representing variability? Do they all lack certain parameterization or components such as ocean circulation dynamics or ice cap dynamics? More broadly, what do these similarities imply for forecasting performance, multi-model ensembles, and (potential?) limitations of current AI weather models? This topic warrants a much deeper investigation into the potential causes and implications of these similarities. The JJA dip in the midlatitudes in Figure 5 might be very telling here.

  2. This paper does not introduce new methodology. Prior work on climate model validation and comparison has developed new methods and shown why they are capturing a relevant quantity of interest and how these methods would interest the broader ML community. While this application could itself be a valuable contribution, the paper does not sufficiently demonstrate what the analysis reveals about current AI weather models. Its not clear how modelers should use these results to improve their models or whether this has any implications on forecasting performance.

Questions for authors
  1. Why does representation similarity matter and what insight does it provide about model behavior? Figure 2 validates representational similarity by showing its near 1-1 correspondence with functional similarity, so why not use one of the many functional similarity notions? Does representation similarity reveal something that functional does not?

  2. What is the motivation for each of these metrics? Figure 4 seems to show that CKA, RSA, and distance correlation are highly correlated, and similarly 2nd order cosine and orthogonal procrustes are also highly correlated. Why were these metrics specifically chosen and do they show differing behavior that is not present in this figure?

  3. Should we expect models to exhibit similarity because they are capturing the same physical processes or should we expect them to be dissimilar since they are independently generated? Which "way" are these tests being conducted? What might be driving these similarities?

  4. Are the p-values distributed uniformly under the null? The data are highly correlated so its not obvious that a permutation null will be uniform without further correction.

  5. Why should null calibration remove background variation, rather than just artificially lowering the statistic values? Couldn't background variations manifest in high similarity scores?

2 comments
AuthorsOpenReview ↗

(Part 1 of reply)

Thank you very much for taking the time to thoroughly go through our manuscript. Below we respond to the critical weaknesses (W) and questions (Q) raised.

W1 (motivation), (and Q1&3): Our study is motivated by the hypothesis that AI weather models converge towards a shared learned description of the atmosphere, as explored in Craig et al. 2026. Under this hypothesis, similarity would be expected between all model pairs. Functional similarity is not sufficient to test this, because the models produce skilful forecasts, especially at 6 h, they will likely have a close functional similarity independent of whether they share similar internal mechanisms. Figure 2 is included as an initial check that, within each model, functional changes correspond to changes in the representational similarity metrics, noting that this need not hold between different models. Interestingly, though monotonic, it is not exactly 1-1, and the relationship varies depending on the metric, indicating that different similarity metrics calculate similarity in slightly different ways. All tests are one-sided, whereby the null hypothesis is that similarity does not exceed that expected from the shared latitudinal and seasonal structure. We’ve updated the introduction to make our hypothesis clearer. Regarding what drives the similarity, we have additionally compared untrained versions of the models (Appendix A.12). Calibrated similarity between the untrained models is close to zero for nearly all pairs, indicating that the similarity between the trained models arises predominantly through training rather than from the architecture or input data alone, consistent with this hypothesis.

W1 (implications): Knowledge gained from calculating the similarity could have implications for understanding features of AI weather models as well as development. For instance, shared representations could imply that sources of weaknesses may be shared too. This is consistent with Selz & Craig 2026, who find that the failure of AI weather models to reproduce the rapid growth of small initial perturbations is largely independent of model architecture. As you have mentioned, shared representations could also lead to correlated errors and underdispersive multi-model ensembles. We have updated the discussion accordingly, together with concrete follow-up experiments which could include relating pairwise similarity to forecast-error correlation. A full investigation of the causes is unfortunately beyond the scope of this first study.

W2: While the similarity measures themselves are well-established, applying them to atmospheric data is not straightforward, as unrestricted permutation nulls are inappropriate due to destroying the latitudinal and seasonal structure. We therefore consider our key contribution to be the structure-preserving calibration, which retains this structure and provides an appropriate baseline from which to judge similarity. The importance of this calibration is shown in Figure 3, where one can see the raw CKA is far larger than the calibrated values. We have updated the contribution section of the introduction to make this clearer.

Q1: As noted in the reply to W1 regarding the motivation above, a high degree of functional similarity does not necessarily imply representational similarity, or that the models have learned a common internal understanding of the atmosphere. If this were true, as our results seem to indicate, it could form the basis of further research to identify what that understanding is, and how it relates to the current understanding, which is based mainly on the fluid equations. The discussion of this motivation in the future work section has been extended and made more specific.

Q2: A variety of metrics were chosen because it has been shown that there is no one single measure which best measures the representational similarity. Therefore, to test whether there was significant similarity, it is important to use a variety of measures. These specific metrics were chosen because they span the main families of Klabunde et al. 2025 and have a variety of different invariance properties (Table 2). CKA, RSA and distance correlation are RSM-based measures, second-order cosine is neighbourhood-based, and orthogonal Procrustes is alignment-based. Procrustes was in addition particularly chosen as it is less dominated by high-variance components than CKA (Cloos et al., 2024), consistent with its faster decrease in Figure 2. The high correlation within families in Figure 4 is therefore expected, along with differences appearing between families in the rankings (e.g. which model is most similar to GraphCast, Section 4.1). We have updated section 2.2 to make this clearer.

AuthorsOpenReview ↗

(Part 2 of reply)

Q3: This is the fundamental question. We would like to know whether the models “exhibit similarity because they are capturing the same physical processes or should we expect them to be dissimilar since they are independently generated”, but the answer is not straight-forward because we don’t know how physical processes would be represented in the models. At least, there is no evidence that they are solving the same equations or making the same approximations as conventional NWP models. This motivation has been emphasised in the introduction and future work sections.

Q4: The structure-preserving calibration introduced in this paper accounts for much of the spatial and temporal correlation in the data, but satisfies exchangeability only approximately, for example because of static variables such as the land-sea distribution. The p-values are therefore only approximately uniform under the null, so we treat them as a supporting diagnostic and base conclusions primarily on the calibrated scores. This is clarified in Appendix A.4.

Q5: Background variations, such as the equator-pole temperature gradient and the seasonal cycle, do produce high raw scores even without correspondence in the weather state, which is why calibration is important. For example, two locations at the same latitude and time of year have broadly similar temperature profiles regardless of the weather systems present. The calibration doesn’t just artificially lower the raw score - for each model pair and metric, a similarity threshold τ is computed from that pair’s own null realisations (which retain the background structure). Large background similarity therefore increases τ along with the raw score, leaving a smaller calibrated score that measures similarity beyond the shared latitudinal and seasonal structure (see Figure 13 as an example of raw and calibrated scores). As a result, the calibrated ordering can differ from the raw ordering, this is seen in figure 9 where AIFS-Aurora and AIFS-GC Small have the same raw CKA (0.54), but calibrated values of 0.10 and 0.29.

References: Craig et al. 2026, The physics of AI weather models, https://arxiv.org/abs/2605.23778 Klabunde et al. 2025, Similarity of Neural Network Models: A Survey of Functional and Representational Measures, https://dl.acm.org/doi/10.1145/3728458 Cloos et al. 2024, Differentiable Optimization of Similarity Scores Between Models and Brains, https://arxiv.org/abs/2407.07059 Selz & Craig 2026, Can AI‐Based Weather Prediction Models Simulate the Butterfly Effect? The Role of Architecture and Implementation, https://publikationen.bibliothek.kit.edu/1000194866

Reviewer_qSBfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates whether AI weather models learn similar internal representations of the atmosphere. It compares final-processor and intermediate-layer latents of six models on a common mesh using five similarity measures, with CKA as the primary one. To discount similarity arising from latitudinal and seasonal structure, scores are calibrated against a structure-preserving null built from random longitude rotations and circular temporal shifts. Analyses across forecast times, monthly windows, latitude bands, and processor depth reveal non-zero calibrated similarity for all model pairs, strongest within the GraphCast family, with architecture-dependent processor evolution.

Critical strengths
  • Timely cross-model interpretability question for AI weather models.

  • Domain-aware null calibration for spatiotemporal autocorrelation.

  • Extensive robustness checks across metrics, dates, and regions.

Critical weaknesses
  • Comparisons with raw input/output fields and randomly initialized networks would help strengthen the conclusions. And similar cross-family CKA findings were already reported by [1].

  • High latent similarity does not necessarily mean the models learn the same physical representations. A few high-variance large-scale patterns can dominate CKA. Also, since the affine translators are explicitly trained to align intermediate layers with the final layer, some improvement in CKA is expected.

  • I think the latent representations seem not to be compared under fully consistent settings. GraphCast is subsampled directly, whereas other models use interpolation without a clearly specified method. The models also differ in prediction targets and temporal resolution, which could affect the comparison. It would be better to ensure fairness further.

  • The CKA estimation may suffer from feature-count bias. Each minibatch contains only 1.1k–2.3k nodes, fewer than the 3,072 or 4,096 feature dimensions of some models. Using 140k samples in total does not necessarily resolve this issue, especially with a biased HSIC estimator.

Reference:

[1] The physics of AI weather models. arXiv preprint arXiv:2605.23778

Questions for authors

Please see the Weaknesses section.

1 comment

Thank you very much for taking the time to thoroughly go through our manuscript. Below we respond to the critical weaknesses (W) and questions (Q) raised.

W1: Inputs to all models are near-identical analyses of the same atmospheric state, and 6h forecasts differ very little between skilful models. As such, comparing raw input/output fields would not discriminate between models. Based on your recommendation, we have repeated the single-timestep comparison of Figure 4 using randomly initialised versions of all six models (Appendix A.12, Figure 25). It’s seen that the raw scores between the untrained models can be high, but after calibration they are mostly reduced to close to zero. The exceptions are between ArchesWeather-Aurora (possibly due to similar transformer architecture) and within the GraphCast family. The similarity reported in the paper therefore arises predominantly through training, rather than from the architectures or input data alone. Craig 2026 did indeed report similarity between AI weather models. This was a provisional study however, as they only looked at two model families’ final processor step, and used individual mesh node points and the CKA/CKNNA. Furthermore, they did not use a null calibration suited to spatially and temporally correlated data.

W2: We agree that CKA emphasises high-variance components. This is why we thought it important to include other similarity measures, in particular orthogonal Procrustes, which is more sensitive to low-variance directions (consistent with Figure 2 and Cloos et al. 2024). We do not claim the models learn the same physical representation, however having similar representations is an indication that they may encode the atmospheric state in similar ways. It is an expected result that the use of translators will increase the CKA value. The translated results are used primarily to assess which differences remain after the change in latent basis, and are therefore only summarised briefly in Section 5. This is updated in A.2.

W3: We added Appendix A.1.5 describing how the latents are mapped onto a common set of locations. Direct selection for GraphCast is exact, because the GraphCast-Small nodes are a subset of the Standard and Operational meshes. Aurora, ArchesWeather and AIFS-CRPS are linearly interpolated: bilinearly on the regular latent grids of Aurora and ArchesWeather, and with barycentric weights on a Delaunay triangulation for AIFS-CRPS. The latents have been checked to be spatially coherent on synoptic scales using the tool of Tempest et al. 2026, giving us confidence in our ability to use interpolation. All comparisons are matched by aligning the forecast time, so every model represents the same atmospheric state. There are differences in models remaining, such as timestep and prediction target, however we decided to still include a model even if it wasn’t homogenous with others, as we were interested whether there would still be similarity seen between its latents.

W4: Feature-count bias is important to check. We therefore carried out sub-sampling experiments (Figures 22 and 23), whereby we calculated raw and calibrated scores for subsets of as few as 500 nodes. It was found that calibrated CKA was stable from approximately 1,000 nodes, even where the latent dimension is larger. The smallest minibatches, those of the single hemispheric midlatitude bands (1,140 nodes, Appendix A.7), are above this limit, while the equatorial and combined midlatitude subsets contain 1,818 and 2,280 nodes respectively. Furthermore, the observed and null scores share the same minibatch size and feature dimension, so any inflation of the biased estimator also raises the similarity threshold, τ, and is therefore largely removed by the calibration. We have updated the sample-size statement in Section 4.2 accordingly.

References: Cloos et al. 2024, Differentiable Optimization of Similarity Scores Between Models and Brains, https://arxiv.org/abs/2407.07059 Tempest et al. 2026, Mechanistic Interpretability Tool for AI Weather Models, https://www.iccs-meeting.org/archive/iccs2026/papers/167880105.pdf Craig et al. 2026, The physics of AI weather models, https://arxiv.org/abs/2605.23778

Reviewer_4zxhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper evaluates representational similarity across leading weather models (three variants of GraphCast, AIFS-CRPS, Aurora and ArchesWeather) and their internal layers using a range of similarity metrics.

Critical strengths
  1. This paper uses various similarity metrics across leading models and between internal layers to evaluate the representational similarity. Also, the calibrated similarity are carefully designed to preserve spatial and temporal dependency.

  2. The observations and comparisons across the models are interesting, and the importance of calibrated, multi-metric analysis is demonstrated.

Critical weaknesses
  1. The extent of separation between strong shared seasonal or regional representation and model-specific representation of the weather state is not clear to me. Seasonal representation takes a major part in Aurora's representation (https://arxiv.org/abs/2606.26361). Also, regional representation is related to the same forcing of topography and location encoding and will of course contribute to the similarity of the latent space. It would be helpful to show how effectively is the distinction of these shared factors from the similarities in the representation of the weather state.

  2. Latent space similarity does not necessarily mean the ways to interpret the latent space are similar given substantial architectural differences among these models in encoders, backbones, and decoders. Therefore, the similarity of representation makes more sense with interaction of the weights in the model. Though this point is partially covered in the limitations, a deeper insight regarding this important aspect would strengthen the paper.

Questions for authors
  1. Could you please provide quantitative results of how effectively the null calibration accounts for shared static variables and geographical forcings, such as topography and location encoding, beyond preserving spatial and temporal dependencies?
1 comment

Thank you very much for taking the time to thoroughly go through our manuscript. Below we respond to the critical weaknesses (W) and questions (Q) raised.

W1, Q1: Each null realisation rotates one representation in longitude and, for monthly windows, shifts it in time by 5–29 days within the month. It therefore retains the latitudinal and seasonal structure of both representations while breaking their correspondence in the weather state. Similarity due to latitude or season, including a strong seasonal component as reported for Aurora, is thus contained in the threshold τ and removed.

Static forcings that vary with longitude, such as orography, land–sea contrast and location encodings however, are not aligned in the rotated null. Their contribution can be estimated in Figure 3, whereby in the left plot, five-day blocks are compared with an anchor block within each model. Blocks far from the anchor share all static forcings and encodings, but not the weather state, so any remaining calibrated similarity bounds the static contribution. It is seen that the calibrated CKA falls to approximately 0-0.1 (from 0.2-0.85 raw) for all models except ArchesWeather. Since a model encodes static fields identically to itself, we expect this to be an upper bound for the cross-model cases. We have clarified this in Sections 2.2.1 and 3.2.

W2: We agree that high similarity scores don’t necessarily imply high similarity amongst all aspects of the model, especially since the architectures between them can vary substantially. Analysing the weights in combination with the latents, for example by testing whether a representation mapped from one model can be decoded by another, will be an important next step and is now discussed in section 7. We believe, however, that the latents alone already indicate how each model represents the atmospheric state, and whether this is similar across models.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.