acceptodds
Withdrawn from ICLR 2027

Evaluating Predictor-Corrector Dependence in Hard-Constrained Climate Emulators

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Climate emulators often enforce physical constraints by correcting each state predicted by a neural network, but evaluating only the corrected forecast conflates predictor skill with the effect of the correction. We introduce an evaluation framework that separates these contributions by comparing corrected forecasts, pre-correction predictions along corrected trajectories, and autonomous rollouts with correction disabled. We further cross correction during training and inference to distinguish its immediate effect from how it shapes the learned predictor, and use matched nonphysical controls to test whether improvements are specific to the physical correction. In ACE2, correction substantially improves physical consistency. In a three-seed ACE2 pilot, models trained with correction produce less accurate uncorrected precipitation forecasts at short horizons than models trained without correction, but more accurate forecasts at 10 and 30 days. This contradicts our preregistered hypothesis that the deficit would grow with forecast horizon. In a separate nine-seed experiment, directly supervising the uncorrected output reverses its precipitation forecast deficit at 72 hours while maintaining the accuracy of corrected forecasts at 24 hours. In ConCerNet's heat-equation system, projection-trained models have higher uncorrected trajectory-based vector-field RMSE than models trained without projection, while restoring the physical projection recovers accuracy far more effectively than a matched nonphysical control. Together, these results show that hard corrections can reshape the predictors trained alongside them, creating dependencies that are not visible from corrected forecasts alone. Hard-constrained emulators should therefore be evaluated by separately measuring predictor accuracy, correction effects, and their interaction during training. Code available at https://anonymous.4open.science/r/ICLR-Corrector.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Climate emulation, physics-guided machine learning, hard constraints, model auditing, predictor-corrector decomposition, weather forecasting.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jk92OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors focus on the dependency between neural predictor and deterministic physical correctors in hard-constrained climate emulators. They stress that assessing corrected forecasts may hide the role and possible weaknesses of the predictor itself.

In this regard, the authors present an evaluation scheme based on 3 different schemes: corrected rollouts, predictions (before correction) but evaluated on corrected trajectories, and autonomous rollouts (correction turned off). In addition, they cross train/inference correction, where the weights of the predictor are held constant when the inference condition changes. Moreover, the authors add matched non-physical controls to separate physical correction effect from generic perturbation effect.

For ACE2, they run an assessment of forecast accuracy and physical consistency and dependency on correction, over variables and forecasting horizons. They find that correction effect is horizon-dependent. They test with a three-seed pilot which shows that models trained on correction give a worse uncorrected precipitation forecast in short horizon, but it gives better uncorrected forecast in longer horizon. On a different nine-seed experiment, they observe that supervision directly on raw precipitation prediction can help to offset the deficit of uncorrected 72h errors. But keeping the accuracy of corrected forecast on 24h.

Moreover, in other experiments on ACE2S-SHiELD+, and adapted ConCerNet's heat equation system, they try to verify the generality of the dependency between predictor and corrector. In general, they recommend to assess raw predictor accuracy, correction effect and training/inference effect separately.

Critical strengths
  1. Systematic evaluation of the interplay between predictor training and physical correction. Nice experimental setting to clearly differentiate between the direct effect of the physical corrector application and its impact on the trained predictor during training. The way to differentiate raw predictions on corrected trajectories and autonomous uncorrected rollouts is very good. Especially the latter also reveals differences in the distribution of recurrent inputs.

In particular, the application of matched controls, crossed training-inference experiments, and multi-horizon predictions enable uncovering effects that are otherwise hidden by ordinary evaluation with corrected output.

This same structure is of potential value when assessing other autoregressive models with deterministic correction operators.

  1. Transparent experiments and informative intervention results.

Positive and negative results (including negative results like failure of preregistered hypothesis that uncorrected precipitation deficit increases as a function of forecast horizon) are presented.

Intervention with 9-seed raw supervision generates relatively strong empirical evidence of the power of directly supervising pre-correction outputs to greatly benefit autonomous prediction of precipitation, with no harm to corrected accuracy for short horizons.

Authors carefully separate statistical significance from practical significance.

Authors explain weaknesses of small seed experiments and evaluation within the pre-training period.

These transparency factors lend confidence to the empirical findings.

Critical weaknesses

1. Insufficient conceptual novelty and mechanistic de-coupling of correction dependency.

The main finding - the fact that hard physical correctors may mask degradation of raw neural predictions - is a topic that has already been studied by Chapman et al. (2026) -- including the topic of unidentifiable directions in correction operators and direct supervision of raw outputs as a potential solution.

This paper builds upon the study with autonomous rollouts, crossed training-inference evaluation, and matched controls.

These are valuable experimental developments but the conceptual progress compared to the existing work is still relatively modest.

In a related issue, for the ConCerNet experiment, the physical projection clearly cancels out the mean of the predicted vector field.

If training only supervises projected outputs, it may not control for errors in the canceled direction.

Therefore, part of the large uncorrected error ratio we see, and much of the benefit from reintroducing physical projection, might be due to this algebraic feature of the projection operator.

The current approach of evaluating vector field RMSE in terms of trajectory error also mixes vector field prediction error and differences in states at which the vector fields are evaluated.

A way to further unpack prediction error into physically allowed and constraint-violating components, using vector field evaluations at fixed state with state RMSE separately, would be useful for determining whether projection training does more to harm the learned physical dynamics or simply to render unidentifiable the discarded components.

Without such an analysis, it's hard to fully mechanistically interpret the ConCerNet results.

2. Not much evidence for generality in the key climate results.

The most interesting result from ACE2 (flipping the uncorrected precipitation error gap on forecast horizon) is only on 3 adaptation seeds. While I think the authors call out this experiment correctly as descriptive, there's a lack of statistical power that limits how much we can say from this.

Also, the months being evaluated are in the pretraining range for the checkpoint for ACE2 we get. While they're excluded from local adaptation, we can't say whether the behavior we're seeing is generic to unseen time periods of climate.

And while the 9 seed matched controls are statistically consistent for each seed, only 8 out of 24 comparisons are statistically above the authors' practical importance thresholds that they've set, and all of those are for temperature. So the benefits of short horizon forecasting for other variables don't seem to be as large as the statistical comparisons suggest.

We do have the direct supervision experiment which adds support. But that experiment focuses on different forecast horizons for the uncorrected and corrected evaluation endpoints. And there are no equivalent longer-horizon results to show whether this intervention is dealing with the wider predictor-corrector dependency that was highlighted in the pilot.

Finally, the additional climate model evidence lies within the ACE2 family and the ConCerNet experiment relies on a simplified heat equation system with a known analytic invariant.

This raises questions about the generality of these results with respect to different physical correction mechanisms and with respect to different forecasting environments.

Questions for authors

Q1. Mechanistic understanding of ConCerNet Results Can the authors provide a breakdown of the uncorrected vector field error into parallel and orthogonal components w.r.t the physically admissible space?

Specifically, how much of the 26.87x error ratio is accounted for by the projection-away component?

Further, could the authors compare the vector fields of both predictors at the same input states and show the RMSE of state trajectories vs vector field (trajectory) RMSE?

These will help to separate out learned dynamics degradation and unconstrained projection nullspace components.

Q2. Robustness of horizon-sensitive ACE2 results Is there a possibility to expand the crossed train-inference experiment beyond the three adaptation seeds, especially for the 10d and 30d horizons where the uncorrected precip error gap flips?

Is it possible to assess the same effects on dates outside of pretraining for the model as released?

Generally speaking, how much of the flip is still observed across different independent climate periods and initializations for forecasts?

Q3. Range of intervention in raw supervision The presented intervention is strong in improving 72hr uncorrected precip forecasts while not hurting 24hr corrected forecasting accuracy.

What are the corrected 72hr results on the withheld evaluation set?

How does the intervention affect 10d and 30d forecasting, other atmospheric variables, as well as the size of physical corrections?

It would also help add value to clarify what extra behaviour is specifically introduced by the current intervention in light of other related raw supervision intervention solutions discussed by Chapman et al. (2026).

Reviewer_ZLNLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how hard physical corrections affect the behavior of neural predictors in climate emulators. It proposes an evaluation framework that separates the effects of correction during training and inference, and examines this using ACE2 and ConCerNet. The authors find that correction can improve physical consistency while having mixed effects on forecast accuracy, depending on the variable and prediction horizon. They also show that directly supervising uncorrected predictions can improve their accuracy.

Critical strengths

None

Critical weaknesses
  1. The contribution of the paper is small, and novelty is limited. Compared to the most relevant cited prior work, Chapman et al. (2026), this paper presents only a slightly different evaluation setting. The additional insights the authors are able to draw from this are limited (plus further hindered by poor writing), and the evaluation itself is not comprehensive (see below).
  2. The evaluation is not comprehensive, and the choice of setup is questionable.
    • The work relies on evaluation using four months, in 1940, which is part of the ACE2 pretraining dataset. It’s unclear from the writing why a recent year was not used to create a more realistic evaluation setup. This is also a small time period, especially since the evaluation has only one evaluation per month.
    • Results are focused only on precipitation, and at short time horizons, but ACE2 is a climate emulator for which physical corrections ensure stability over multi-year/decadal scales. Precipitation is also hard to predict at those scales.
    • The only evaluation metric is RMSE, which is not very informative in weather/climate. Further, given that the work is trying to study the impact of imposing hard conservation corrections, having some qualitative and quantitative analysis of physical consistency in the outputs is very important but missing.
    • Section 3.1 discusses “pre-correction error measured along corrected rollout” but I cannot find such an experiment in the paper.
Questions for authors
  • Can the authors add a more thorough related work to better contextualize the contributions, and also better clarify what the contribution is.
  • Can the authors include some other evaluation metrics like ACC including something for physical consistency of the emulation output.
Reviewer_zJyVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies spectral scale imbalance in generative modeling of turbulence. The authors argue that differences in Fourier-mode energy produce imbalanced flow-matching targets and signal-to-noise ratios, making small-scale structures difficult to learn. WSFM estimates a mean and standard deviation for each velocity component and Fourier mode from the training ensemble, performs flow matching in the standardized spectral coordinates, and reconstructs physical fields through the inverse transform. Using SpectralDiT, the paper evaluates prediction errors, spectra, and physical statistics on cylinder wakes and additional flow datasets. Whitening substantially improves over raw Fourier flow matching, although the proposed per-mode formulation does not consistently outperform simpler radial whitening.

Critical strengths
  1. The method is simple and motivated by a relevant physical issue: large scales dominate energy, while small scales are important for vorticity and dissipation. Fixed spectral standardization is a reasonable and accessible way to address this imbalance. It requires neither a separately trained encoder nor lossy compression and can be incorporated as a representation change in an existing generative model. The distinction between normalizing spectral amplitudes and learning non-Gaussian structure also provides useful intuition about the remaining modeling task.

  2. The improvement over raw Fourier coordinates is supported by useful controlled evidence. Table 2 compares representations with essentially the same backbone, parameter count, and training configuration, reducing model-size confounds. Relative to raw Fourier flow matching, WSFM reduces relative L2L_2 error from 0.292 to 0.215 and high-wavenumber spectral error by approximately 99%. Evaluation also includes structure functions, vorticity, enstrophy, and dissipation statistics, extending the analysis beyond pointwise reconstruction error. These results support spectral standardization as an effective practical choice.

Critical weaknesses
  1. The experiments establish the benefit of whitening over raw Fourier coordinates, but do not establish the practical value of the proposed per-mode formulation over radial whitening. In Table 2, radial whitening achieves lower relative L2L_2 error, 0.190 versus 0.215, and lower dissipation error, 0.153 versus 0.210. Appendix A.9 reports statistically significant paired differences, and the vorticity metrics also favor radial whitening. Conversely, the anisotropic-spectrum advantage of per-mode whitening is not statistically resolved. I appreciate the authors’ acknowledgment that the per-mode formulation adds exactness, but its practical necessity and significance remain unclear. Section 6 additionally discusses equivalence to a spectrally matched source combined with per-mode loss weighting; the incremental contribution beyond these components needs clearer delineation.

  2. The theoretical account conflates mode-wise variance equalization with stronger whitening and optimization-conditioning claims. Equation (4) normalizes marginal Fourier-mode variances but does not generally remove cross-mode correlations. Consequently, Proposition 3 requires additional covariance assumptions to conclude that a Gaussian field becomes CN(0,I)\mathrm{CN}(0,I). Likewise, a condition number of one for the diagonal scale matrix does not establish uniform conditioning of general regression or neural-network optimization. Equation (2) also conflates variance with the second moment when modal means are nonzero: under Equation (1), the target-velocity variance is Dk2+1D_k^2+1, rather than E∣qk∣2+1\mathbb E|q_k|^2+1. I appreciate the narrower empirical interpretation in Section 5.2, but the abstract, propositions, and derivations need consistent assumptions and scope.

Questions for authors
  1. Given that radial whitening achieves lower relative L2L_2, dissipation, and several vorticity errors, which results establish a practical advantage of the per-mode formulation? How should the statement that the methods do not separate on bulk metrics be reconciled with the significant paired differences reported in Appendix A.9?

  2. Please identify the precise matrix or operator described as uniformly conditioned and state the required assumptions. Could the authors also provide the raw-coordinate source and weighted objective corresponding to the L2L_2 flow-matching loss actually used? Appendix A.5 derives the L1L_1 case, whereas the L2L_2 correspondence involves weights proportional to Dk−2D_k^{-2}; the derivation should be aligned with the experimental objective.

  3. Appendix A.4 states that deterministic models receive 20 genuine historical frames while generative models use single-frame guidance, but subsequently describes every model as observing the same single frame. Please specify the historical frames, physical parameters, guidance, and frame-replacement rules available to each model on each dataset. For the unguided foil and combustion settings, how are generated samples associated with particular reference trajectories for pointwise evaluation?

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

Related papers

Discussion (0)

Sign in to comment.