acceptodds
Under review as a conference paper at ICLR 2027

Learning to Repair Physical Violations in Frozen Weather Foundation Models

Abstract

Weather foundation models can achieve strong predictive accuracy while producing states that violate physically meaningful relationships. We study whether such violations can be repaired after forecasting, without modifying the forecasting model. ConserveFM learns an additive correction conditioned on a frozen forecast, the preceding atmospheric state, and forecast lead, and optimizes predictive fidelity together with moisture, hydrostatic, advection, and spectral consistency. We evaluate post-hoc repair on ERA5 under 60 matched stress conditions spanning three forecast leads, five structured corruption families, and four severities. An accuracy-oriented variant achieves the best stress NRMSE of and ACC of . A physics-oriented variant reduces the aggregate physical-violation ratio from to , while increasing stress NRMSE to . Direct state prediction occupies an intermediate Pareto-optimal regime. These results show that predictive accuracy and physical consistency are distinct objectives: stronger physical repair can be obtained without retraining the backbone, but not without a measurable predictive cost.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
weather foundation models, post-hoc repair, physics-informed learning, physical consistency, multi-objective learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mrddOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

A method which corrects ML weather models through a learned additive correction which is optimized for forecast accuracy and physical/spectral fidelity.

Critical strengths
  • they investigate if post-hoc corrections can improve AI weather models
Critical weaknesses
  • very little novelty and very little evaluations
  • the quality of the presentation is low
  • evaluation feels rushed - there is very little gain from applying the method from looking at Table 1 and this is just from a single run so it isnt clear what the uncertainty is
  • Figure 1 is convoluted and not easy to read.
  • the introduction does not cite most of the main breakthrough literature in the domain of atmospheric science and feels biased. Methods to cite would be FourCastNet, Pangu, Graphcast in chronological order.
  • Other important work related work to enforcing constraints and post-hoc correction aren't cited. i.e. see https://arxiv.org/abs/2310.02074, https://arxiv.org/pdf/2306.09739,
Questions for authors

why omit a large portion of the literature why the convoluted training procedure

Reviewer_4SahOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to present a post-hoc method for repairing physical violations in frozen weather foundation models. The proposed method, ConserveFM tries to learn an additive correction to the forecasting made by a foundation model, using an objective which is supposed to contain several physical terms. Experiments are conducted to evaluate the proposed method, using the ERA5 dataset and the ClimaX foundation model.

Critical strengths

The problem addressed is important.

Critical weaknesses
  1. Fundamental methodological concern: The high-level motivation of the paper is meaningful, but the proposed method, as currently presented, does not provide sufficient technical justification for its central claim of “physical repair.” The correction model itself appears to be a standard post-hoc residual correction, and the physical interpretation therefore seems to depend critically on the proposed physics-related objective terms (presented in Section 3.2). However, no precise mathematical definitions are provided for these terms. As a result, it is difficult to determine what physical relationships are actually being enforced, how the corresponding residuals are calculated, or whether the proposed repair is genuinely physics-informed rather than a learned forecast correction.
  2. Technical and presentation quality: In addition to the above mentioned fundamental problem with the method, key information of the experiments are not provided. In particular, no references for sufficient information are provided for the data (ERA5) or the weather foundation model (ClimaX) used, and more importantly, it is not clear how normalized RMSE measures physical consistency in this setting.
Questions for authors
  1. Could you provide further information about the physical terms in the so called "physics-aware objective" in Section 3.2?
  2. How NRMSE is used to measure physical consistency?
Reviewer_GotwOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces ConserveFM, a post-hoc module that adds a learned correction to forecasts from a frozen ClimaX backbone. The objective combines reconstruction with moisture, hydrostatic, advection, and spectral diagnostics. Accuracy-oriented and physics-oriented variants are compared on ERA5 under synthetic corruptions, showing a trade-off between predictive error and the reported violation ratios.

Critical strengths
  1. Treating physical correction as a modular post-hoc problem is practically appealing when a large forecasting model cannot be retrained. The residual path also gives a natural way to measure how much the output is changed.
  2. The paper is candid that predictive accuracy and its physical diagnostics identify different operating points. Reporting that the accuracy-oriented model worsens the advection score and that the original joint objective is unstable is valuable.
Critical weaknesses
  1. The four physical losses, five corruption operators, repair architecture, and most training hyperparameters are not defined precisely enough to reproduce the method. The text describes them only at a high level, and the advection quantity is explicitly an endpoint surrogate rather than a dynamical constraint. Reducing these scores therefore does not by itself show that the repaired atmospheric state is physically valid.
  2. ConserveFM-Phys is trained on the same corruption families used in the stress benchmark, the normalized headline table uses one final seed, and there is no matched predictive row for raw ClimaX or an external projection/post-processing baseline. The study also stops at independent 6–72 hour outputs rather than feeding repaired states into a rollout. These choices leave open whether the model repairs real forecast failures, generalizes beyond known synthetic perturbations, or improves downstream stability.
  3. Figure 1 is not publication-ready. Text is visibly overprinted or obscured in the frozen-forecaster and ConserveFM-Phys blocks, and several group labels collide with the diagram. It appears to have been automatically generated without a final manual check, making the central pipeline harder to verify.
Questions for authors
  1. Can the authors provide complete equations and implementation details for every physical loss and corruption operator, including the repair architecture and all selected weights?
  2. What are the raw ClimaX errors and physical diagnostics on the exact same clean and corrupted samples, and how does ConserveFM compare with a tractable analytical projection baseline?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.