acceptodds
Under review as a conference paper at ICLR 2027

Data-Anchored Response Spaces for Closed-Loop Neural Components in Non-Differentiable Numerical Simulators

Abstract

Neural components embedded in atmospheric simulators shape the evolving states that become their future inputs, making accurate offline prediction insufficient for reliable coupled simulation. Model-based policy optimization can account for these downstream effects, but atmospheric applications lack explicit action labels in reference trajectories and are vulnerable to policies exploiting errors in learned dynamics surrogates. We introduce data-anchored response spaces to enable model-based policy optimization in this setting. Using a known dynamical backbone, we recover effective responses from reference state transitions and use them to initialize a response codebook comprising a compact set of prototype temperature and moisture tendencies. The policy predicts nonnegative mixing weights that sum to one, producing continuous convex combinations of these prototypes rather than selecting discrete actions. This representation anchors policy outputs in responses derived from data while retaining continuous control. We alternate policy optimization through differentiable surrogate rollouts with coupled-simulator evaluation and surrogate refinement on aggregated trajectories, screening candidate updates before acceptance. In a 72-hour global atmospheric simulation, the evaluated configuration reduces time-averaged specific-humidity and temperature RMSE by 30.4% and 38.5%, respectively, relative to the offline-trained policy. Corresponding mean absolute biases decrease by 92.9% and 28.4%. Its mean absolute humidity and temperature biases are also 92.1% and 28.7% lower than those of native physics, although RMSE remains higher. These results demonstrate the potential of combining data-anchored response representations with simulator-informed adaptation in atmospheric systems, where learned outputs continually reshape their own deployment conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
model-based policy optimization, scientific machine learning, atmospheric simulation, surrogate modeling, closed-loop learning, action representation

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_45shOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces Data-Anchored Response Spaces (DARS), a method to improve the stability of neural parameterization of neural surrogates applied to the scenario of atmospheric forecasting. The key contribution is the idea of using training prototypes to constrain the output of the neural network to be more physically realistic. The paper also proposes a method for closed loop adaptation of the neural network using the non-differentiable simulator. The results compare DARS to an unrestricted network, a componentwise limiter, and a simulator only forecast and find that DARS remains stable for 72 hour rollouts.

Critical strengths
  • The idea of using multivariate response profiles rather than controlling each component independently is clever and seems like a good way to improve stability.
  • The main result in Table 1 is interesting as it shows that DARS is most stable in comparison to the compared baselines across 3 seeds.
Critical weaknesses
  • The baselines are not strong enough in comparison to DARS as it is only compared to a unconstrained NN and a simple limiter. The authors should compare to a standard kernel regression model as well as a convex mixture router. In addition, ablations of the convex restriction, prototype selection, and learned routing methods would strengthen the paper.
  • The writing of the paper was unclear and difficult to understand. It took significant effort to understand the methods and results. Mathematical notation and terms are written before defining them, for example k=0k = 0 in the Introduction section.
  • The transfer experiments on the 22 initializations include multiple extra steps detailed in Appendix Q, so it is unclear if the DARS method is responsible for the positive results or if the extra steps cause the improvment.
Questions for authors

Have you compared DARS to a convex-combination model using the same response dictionary or other standard learned methods that use the same inputs?

Reviewer_pN9bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a neural network-based subgrid model for the GRIST atmospheric simulator. The basic idea of this model is to predict the temperature and moisture tendencies by essentially taking weighted averages of complete response profiles assembled from training data. So a neural network learns a representation that gets used to choose the weights in that weighted average. They report improved stability when this model is coupled to the original simulator.

Critical strengths

(1) I appreciate that they test this in the actual simulator. Furthermore, I thought the comparison before adaptation was interesting: while the unrestricted models terminate after around 20 hours, DARS completes the full 72-hour simulation across three training seeds due to its improved stability. So while it's only one atmospheric initial condition, I think it's still useful evidence that this choice of output representation can actually make a large practical difference.

(2) At an abstract level, what they're doing is restricting the available subgrid responses by forcing the model to average different responses from the training data. So I like this idea. I'm sure it's been done in other contexts. It kind of reminds me of Plenoxels, in the radiance field context. So it gives us a fairly interpretable way to limit extrapolation. I also think the combination of the different ideas in this paper, even though the ideas themselves might not be novel, seems to be novel, and I think it's interesting.

Critical weaknesses

(1) So right now my main concern with the paper is that I do not think these experiments show what this particular combination of techniques contributes over existing simpler methods. For example, there are works that already use random forests to predict joint atmospheric tendency profiles (Yuval and O'Gorman, 2020). They also discuss averaging training outputs to remain within the convex hull and discuss that as a possible reason for robustness. Furthermore, learning a distance metric for kernel-weighted regression has a long history, for example Weinberger and Tesauro (2007).

So the authors do acknowledge related approaches, but I think an acknowledgement is not enough. I actually want to see a comparison to understand what's the delta on previous work. I mean, I want to know what's new in this and what actually matters. Because right now I just see this fairly complex machinery. I don't know which part matters.

(2) The other main issue I have is that the headline accuracy improvements come from a different setup from the main fixed-dictionary adaptation experiment. The reported 31.4% humidity and 61.2% temperature RMSE reductions across 22 initial conditions are relative to the offline parent, but Appendix Q describes additional changes to the response basis, calibration, and prescribed temperature tendencies. The final temperature-basis update is directly supervised and does not use the dynamics surrogate.

Deployment also uses a prescribed sequence of models over the first 12 hours, followed by a transition and a fixed continuation model. I should say that the authors do distinguish these experiments. My concern is that the larger improvements do not show the reader how much comes from the central proposal of adapting the selection rule while keeping the response dictionary fixed.

The router-only experiment shows improvements, but for the example seed 1117 they are about 3.0% for humidity and 2.8% for temperature over the full trajectory. I think that is useful, though I think it still leaves the explanation for the much larger transfer gains unresolved. These are different experiments, so their difference cannot be attributed to any one added component. I would also like to see the extra changes separated within the transfer setup, so I can maybe understand which parts of the method are actually responsible for those gains.

Questions for authors

(1) Do you have results comparing DARS with a network that directly predicts softmax weights over the same response dictionary, controlling for output ranges and the shrinkage toward the training mean? If not, this is the comparison I would find most useful for assessing things.

(2) Are there matched results for the 22-initial-condition experiment that separate routing adaptation from the additional response-basis updates and temperature corrections?

Reviewer_QfNBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Data-Anchored Response Spaces (DARS) to improve the stability of neural components coupled with non-differentiable numerical simulators. DARS constructs fixed multivariate response prototypes from training transitions and employs Gaussian routing with surrogate-assisted adaptation. Experiments show that DARS completes 72-hour simulations across three seeds, outperforming unrestricted and componentwise-limited readouts. A separate evaluation on 22 initial conditions reports substantial RMSE reductions relative to the offline model. However, the experimental evidence does not sufficiently isolate the contributions of the proposed components.

Critical strengths
  1. The paper addresses the discrepancy between offline prediction accuracy and closed-loop stability, which is highly relevant to scientific machine learning.

  2. DARS demonstrates improved long-horizon simulation stability across three seeds and consistent RMSE improvements over its offline parent across 22 evaluation initializations. Evaluation using the original numerical simulator strengthens the practical relevance of the study.

Critical weaknesses
  1. The proposed framework largely combines existing ideas from kernel regression, RBF networks, prototype mixtures, and surrogate-assisted adaptation. The convex-hull constraint is a direct consequence of normalized nonnegative weights. The paper does not sufficiently establish what fundamentally distinguishes DARS from existing structured prediction methods.

  2. The stability comparison includes only unrestricted and componentwise-limited readouts, without stronger structured baselines. Moreover, epsilon normalization can shrink outputs toward the normalization mean. The experiments do not rule out this implicit suppression as an alternative explanation for the stability improvements.

  3. The 22-initialization experiment incorporates response-basis adaptation, additional temperature-tendency corrections, and a scheduled deployment strategy. Without corresponding ablations, the reported RMSE improvements cannot be attributed specifically to the proposed response-space representation.

Questions for authors
  1. What is the essential methodological distinction between DARS and existing kernel/RBF mixture models? Can stronger structured baselines be provided?

  2. Can the authors demonstrate that the stability improvements arise from preserving joint response structure rather than epsilon-induced output shrinkage?

  3. How much of the reported transfer improvement is attributable to response-basis adaptation, temperature corrections, and the deployment schedule individually?

  4. Can the authors report numerical replay errors to validate the reconstructed effective responses?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.