acceptodds
Under review as a conference paper at ICLR 2027

ReAl: Refine neural operators via aligning dynamical flow

Abstract

Neural partial differential equation (PDE) surrogates trained with mean squared error (MSE) can capture dominant dynamics but often lose high-frequency information and fine-scale structures. Generative models (e.g., Flow Matching) can capture the distribution of fine-scale details, yet stochastic errors may accumulate during autoregressive rollouts and compromise long-term stability. To address these limitations, we introduce Refine and Align ReAl, a model-agnostic plug-in that augments pretrained PDE surrogates to improve the reconstruction of high-frequency structures associated with complex physical phenomena. In the refinement stage, a conditional flow-matching model learns the residual distribution around the deterministic base prediction, recovering missing fine-scale structures while retaining the dominant dynamics. In the alignment stage, to mitigate the generative discrepancies amplified in the autoregressive rollout, particularly in chaotic systems, we generate multiple trajectories from the same input and reinforce the refiner with trajectory-level rewards, thereby achieving long-horizon dynamical alignment.Experiments across diverse surrogate architectures and PDE tasks demonstrate that ReAl consistently improves high-frequency fidelity, forecasting accuracy, and long-horizon stability.

Primary area:
learning on time series and dynamical systems
Keywords:
Reinforcement learning, generative modeling, PDE forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PWDyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The proposed paper studies some issues of current neural surrogates regarding high frequency modeling and long range divergence. These issues have been well identified by the community and remain a strong limitation of the majority of the current models. ReAl proposes to solve this issue with a refiner trained in 2 steps, each of them addressing one these challenges. First a flow matching based training on the residual for a better modeling of high frequencies, and second a reinforcement learning training that aims at stabilizing long rollouts. The paper comes with some experiments to illustrate the claim and some appendices providing details on the metrics and baselines as well as some visualizations.

Critical strengths
  • I found the paper clearly written and easy to follow. Each component is introduced with the corresponding level of details for me. A minor modification for me would be to add in the appendices some details about on the RL side (algorithm, pseudo-code for instance)
  • The problem studied is well identified in the literature and an additional refiner could be pluged onto already existing models without architectural modifications
  • The datasets studied are challenging and very recently proposed to the community
Critical weaknesses
  • I think the experimental part could be improved to make the paper more convincing, especially regarding baseline comparison. In my opinion, a strong part of the literature is missing in the experiments as the method is only compared wrt to models without the additional refiner. I fell like it is a bit unfair to compare these methods as they have not been designed for this problem, when huge literature on the two subjects exists. Some papers are mentioned in the related work section. I think a strong experiment is missing here as we do not have any comparison with other methods addressing the same problems. Why not comparing to PDE-Refiner or Flow Refiner that you mentioned ? Additionally, I believe more recent baselines could be used. Why considering Transolver and not Transolver ++ [1] or ECHO [2] recently proposed ? While I acknowledge that very recent models can be difficult to spot and reproduce, a two year gap is quite long in the literature. One could think about Thermalizer [3], TNS [4], Dyslim [5] as more recent relevant work.

  • I also think that the related work section could be improved. In order to clearly position the paper, maybe that a sub-section on long-range stabilization method and another on high-frequency modeling would help position the paper wrt to existing method that actually target the same problem.

-The note at the bottom of page 3 is not demonstrated nor illustrated in the paper. A small ablation to study the impact on the prediction or history window would illustrate this point.

  • No limitation section is provided.

[1] Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries. Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, Mingsheng Long. 2025

[2] Efficient Generative Transformer Operators For Million-Point PDEs. Armand Kassaï Koupaï, Lise Le Boudec, Patrick Gallinari. 2025.

[3] Thermalizer: Stable autoregressive neural emulation of spatiotemporal chaos. Chris Pedersen, Laure Zanna, Joan Bruna, 2025.

[4] Towards Infinitely Long Neural Simulations: Self-Refining Neural Surrogate Models for Dynamical Systems. Qi Liu, Laure Zanna, Joan Bruna, 2026.

[5] DySLIM: Dynamics Stable Learning by Invariant Measure for Chaotic Systems. Yair Schiff, Zhong Yi Wan, Jeffrey B. Parker, Stephan Hoyer, Volodymyr Kuleshov, Fei Sha, Leonardo Zepeda-Núñez. 2024

Questions for authors
  • Have you studied what kind of errors are corrected by the refiner ? Is it numerical noise ? measurement errors ? or is it actual dynamical high frequencies ? The strong decrease in energy on figure 1a makes me wondering about the nature of these frequencies in the dynamics.
  • Can you provide the training cost and inference time & memory consumption ? As the method incurs an additional 2 staged training, this metrics is very important to evaluate its applicability. Also, how does it increase inference time ?
  • Are all baselines on the same experimental setting ? (10 input frames to predict the 10 following ?) Can you comment on the claim of "Long-range" ? In the experiments, are only 5 calls to the model required to predict the 50 time steps ? In this case and with the very little number of calls to the model, can we really talk about auto-regressive error accumulation ?
  • Isn't the trajectory-level reward used in the alignement phase training ? ( (lines 258-269 and equation 15)
  • With the two staged training, have you observed some forgetting in the refiner ? For instance are the high frequencies better reconstructed before the alignement phase ?
  • Could you precise if the alignement stage is trained only on the last rollout step (lines 260 & 310) or on every auto regressive step ?
  • Regarding W1, can you compare to dedicated methods for long range stabilization and high frequency reconstruction ? While Table 1 is convincing regarding the broad applicability of the method, I believe that it is more important to show that it is competitive wrt other existing method. Why one should use ReAl wrt any other designed model or correction method for these challenges ?
  • Have you evaluated the method on an OOD setting ? What I mean here would be to train the method on a given horizon, and evaluate the behavior of the refiner beyond this training horizon.
Reviewer_xgv7OpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

ReAl augments frozen PDE predictors with a conditional flow-matching residual generator and then post-trains this refiner using rewards from multi-step forecasting trajectories. The alignment objective combines positive and negative flow-matching terms with reference and supervised losses. Experiments on RealPDEBench report improvements across several backbones and forecasting horizons.

Critical strengths
  1. The paper addresses error accumulation in long-term physical prediction, an important practical limitation of learned simulators. Improvements across several backbones and tasks suggest that the overall refinement approach has value beyond a single benchmark configuration.
  2. Broad evidence of practical usefulness across PDE predictors. Table 1 evaluates four backbones, four physical tasks, and both Sim2Real and Real2Real settings, with improvements in nearly all reported configurations.
Critical weaknesses

It remains unclear whether the observed improvements stem specifically from the reward-based alignment mechanism. Compared to the pretrained baseline, ReAl introduces an additional 10,000 training iterations alongside several auxiliary objectives (reward-dependent, reference, and supervised terms; §4, Eq. 15). However, Figure 4 and Table 3 lack a compute-matched control baseline that incorporates these identical auxiliary losses while omitting informative reward ranking. Consequently, the performance gains cannot yet be definitively attributed to the paper's core alignment approach.

Questions for authors
  1. Under the same additional training budget and auxiliary losses, does reward-aware alignment outperform continued supervised refinement and a reward-shuffled control?
  2. Please specify the computational graph for Eqs. (13)–(15): the source of each residual target, which predictions are recomputed by the current model, and where gradients are stopped. In particular, how does the supervised term update the current refiner?
Reviewer_7MXWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a refine and align methodology to improve high frequency content of autoregressive PDE emulators. The first step uses a conditional flow matching based refinement framework, and then an RL based alignment framework to reward trajectories with better high frequency content. The paper shows results on the latest realpdebench dataset and across 4 backbone architectures. They find that their method works well in capturing high-frequency content of autoregressive PDE rollouts, improving against the baseline backbones in terms of pointwise errors and frequency content errors.

Critical strengths
  1. I have not seen a paper on PDEs before combining ideas from flow matching and reinforcement learning to improve the frequency content of long-rollouts in PDE surrogates, which is a legitimate problem with PDE surrogates. So the combination of ideas presented here appears to be novel.

  2. The dataset used for analysis, RealPDEBench is a very good and state-of-the-art dataset to base these calculations on. Because it has both simulated and real-world examples.

Critical weaknesses
  1. Missing simpler alternatives: It is not fully clear to me whether an RL approach is necessary here. As examples, how does this method compare against other approaches, such as rollout training. I think having some other method baseline like rollout training is important to understand how effective the RL strategy actually is.
  2. Some absent foundation model baselines: The authors show improvement in DPOT, but does this improvement translate to other more state-of-the-art foundation models, like Walrus? Given that the paper says the method is model-agnostic, I think having another recent foundation model would really strengthen the case.
Questions for authors
  1. As pointed out in the weakness, can the authors provide a comparison against rollout training baseline? and whether frequency improvement with that is comparable to your method.. in general, any way to show better the real need for RL will be very appreciated.

  2. Could you add a comparison taking another strong foundation model, like walrus as the backbone?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.