acceptodds
Withdrawn from ICLR 2027

Learning Neural PDE Solvers from Physics Terms That Matter

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural PDE solvers are commonly trained on high-fidelity simulations of the full governing equations, which is costly and limits generalization to unseen physical regimes. Cheap simulations of simplified PDEs can help, but prior methods often fix a single basic form and assume the same physical term remains dominant. This is unreliable: the relative importance of advection, diffusion, reaction, and other operators can change with system parameters and evolving states.We present a physics-guided framework that estimates each PDE operator's contribution and uses Adaptive Dominant-Operator Routing to select the basic form of the current regime, then generates inexpensive auxiliary simulations from it. When no operator is clearly dominant, Confidence-Aware Multi-Fidelity Training down-weights cheap data according to routing confidence. Experiments on 1D advection-diffusion, 2D reaction-diffusion, and 2D Navier-Stokes cover distinct dominance regimes. The method jointly decides which simplified physics to simulate and how much to trust it, targeting better data efficiency, long-term prediction, and out-of-distribution generalization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Dominant Operator, Physics-Guided 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_ESzbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The title and abstract displayed on the webpage do not match the content of the attached PDF. The following review comments are based on the PDF attachment.

Jointly training neural operator learning models on the original PDEs and their simplified basic forms has previously been shown to effectively improve model performance. This manuscript argues that using fixed simplifications of the original PDEs has inherent limitations and further proposes an adaptive operator selection strategy and a discrepancy-based loss weighting method to improve training efficiency and model performance. The manuscript evaluates the proposed approach on Burgers, FitzHugh–Nagumo, and Navier–Stokes equations.

However, the manuscript suffers from a highly disorganized presentation and unclear explanations, making it difficult to follow. Furthermore, the method novelty is extremely limited. Finally, the experimental results fail to demonstrate consistent performance improvements across the three benchmarks. The manuscript also lacks sufficient evaluation on more challenging 3D PDE benchmarks, where reducing data generation costs and improving training efficiency would be particularly meaningful.

For these reasons, the reviewer does not recommend acceptance of this manuscript at ICLR.

Critical strengths

The title and abstract displayed on the webpage do not match the content of the attached PDF. The following review comments are based on the PDF attachment.

  1. This manuscript focuses on an important research problem. Jointly training neural operator models on the original PDEs and their simplified basic forms has shown considerable promise for solving complex 3D PDEs, where numerical simulations are particularly computationally expensive.
Critical weaknesses

The title and abstract displayed on the webpage do not match the content of the attached PDF. The following review comments are based on the PDF attachment.

  1. The manuscript suffers from a highly disorganized presentation and insufficient explanations of the method and results. The reviewer believes that the manuscript requires a substantial restructuring to improve its clarity, coherence, and readability.
  2. The effectiveness of the proposed method has not been sufficiently demonstrated. First, the proposed method fails to achieve consistent performance improvements across the three benchmarks. Second, the ablation results in Table 4 of the appendix indicate that the proposed dynamic weighting method does not consistently improve prediction accuracy and may even degrade performance. Furthermore, the manuscript lacks comparisons with standard neural operator training without auxiliary supervision, making it difficult to determine whether the proposed framework offers meaningful benefits over the standard neural operator training pipeline. More importantly, the experimental evaluation is restricted to relatively simple 1D and 2D PDE benchmarks, which is insufficient to convincingly show the effectiveness of the proposed framework.
Questions for authors

Please refer to the concerns raised in the Weaknesses.

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

This work is motivated by an important problem that is inherent to neural surrogates: they require large amounts of simulated trajectories to train it, which is very expensive. To alleviate it, the authors propose to also train the model on basic forms of the full PDE equation, to obtain more training points. Because the chosen simplification may be a poor proxy for parts of the training data, the authors propose an adaptive operator selection approach by generating 3 steps from each basic form and comparing it to the full form, allowing to give more weight to more important basic forms.

Critical strengths
  • I think the problem is important and exploring basic forms of the full PDE as a way to overcome the gap of data generation makes sense. Making the choice of auxiliary physics responsive to the regime is also a natural extension of existing works.

  • the authors report also unfavorable results, which is appreciable.

Critical weaknesses
  • The paper is very hard to read and understand fully.

  • The title reported on Openreview and in the paper is very different.

  • The experimental setup no longer tests the benefit that motivates the approach. The proposed method is an expensive offline process. Basic simulations are attractive because they are cheaper: they replace part of the full-PDE simulations at a comparable or reduced total simulation cost, and the baseline is a model trained on the original PDE alone with the same number of gradient steps. Here, every Basic target is generated from an anchor of an existing full trajectory, which keeps its own full target.

  • Missing baselines. There is no model trained with the full loss only. Without it, the paper cannot show that Basic supervision helps at all in this setup, adaptive or fixed.

  • Contradictory results across protocols. The main protocol favors adaptive selection on FHN and Navier–Stokes; the sparse-anchor protocol favors the fixed form on both.

  • Limited scope. Grids are small, the evaluation horizon is ten steps, and only seen parameter values are tested.

Questions for authors

What are the results of a full-only model configuration under both protocols?

How do the results in Table 1 change over several seeds?

What is the distribution of selected operators across Re and Da?

Through what mechanism can a one-step Basic target, aligned with an available full target, improve the full head?

Reviewer_EPG5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper explores how simplified PDEs can be used to train neural PDE solvers. Building on prior work that uses fixed Basic PDEs and fixed auxiliary-loss weights, the authors propose an adaptive approach that selects the most appropriate Basic operator at each training anchor. They select the operator by comparing simulations of the admissible Basic PDEs with the corresponding Full PDE trajectories over three saved time intervals. They also adapt the auxiliary-loss weight based on the measured discrepancy. The authors evaluate their approach on three PDEs (Burgers, FitzHugh–Nagumo, and NS) and compare it with fixed operator selection. The results are mixed. In the main all-time, single-step experiments, the adaptive method improves performance on FitzHugh-Nagumo and NS, while the fixed method performs better on Burgers.

Critical strengths
  • The authors evaluate their approach across different PDE settings. The paper is also transparent about its limitations and reports negative results rather than presenting only favorable outcomes.
Critical weaknesses
  • The practical advantage of the proposed method remains unclear to me. Although Appendix C discusses the computational cost of replaying Basic PDE simulations, I could not find actual runtime numbers or an end-to-end computational comparison that includes calibration, and model training. This makes it difficult to determine whether the method provides a meaningful computational efficiency advantage. The proposed approach requires simulating all admissible Basic PDEs from eligible training anchors and comparing their trajectories with future Full PDE states. Although the reported fixed and adaptive experiments share candidate-replay costs, the paper does not establish the total cost relative to a simpler fixed-form implementation. Furthermore, the accuracy improvements are not consistent across different PDEs and experimental protocols, making it difficult to assess whether the benefits justify the additional offline processing. There are also limitations in the experiments. First, all neural models are evaluated using only one training seed, so the statistical reliability of the improvements is unclear. Second, some models may not have fully converged. For example, the adaptive FitzHugh–Nagumo model achieves its best validation result at epoch 48 out of 50, while the adaptive Navier–Stokes model achieves its best result at epoch 38 out of 40. It is unclear whether the reported performance differences would remain the same with longer training. Finally, the authors do not include a Full-only training baseline. It is therefore possible that both fixed and adaptive approaches perform worse than an FNO trained using only Full PDE supervision.

  • The presentation of the paper needs substantial improvement. Important figures, including the method overview and main experimental results in the appendices, which makes the paper difficult to follow. Meanwhile, some implementation-specific details appear in the main text, although they would be more appropriate for the appendix. For example, Section 3.4 includes the following explanation: "For implementation traceability, the cache field selected_operator stores the minimum-J_k(3) route used for adaptive Basic supervision; formula_selected_operator stores a historical diagnostic and does not control this route." Such implementation details interrupt the explanation of the methodology. The paper would benefit from substantial reorganization to improve clarity and readability. Also, submitted abstract and title are different from submitted pdf.

Questions for authors
  • Why is a Full-only training baseline not included? Could the authors report the performance of an FNO trained without Basic auxiliary supervision in the main tables, using the same architecture and training protocol?

  • Could the authors extend the number of training epochs and report whether the models have converged? In particular, it would be useful to determine whether the relative performance of the fixed and adaptive methods changes with longer training.

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

Related papers

Discussion (0)

Sign in to comment.