acceptodds
Withdrawn from ICLR 2027

OPDE: An End-to-end Framework for Solving Inverse PDE Problems from Observations Alone

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Solving inverse PDE problems from observations remains a fundamental challenge in applied mathematics and physics-informed machine learning. Classical approaches, such as system identification and symbolic regression, provide interpretable and closed-form representations but are typically restricted to expert-level skills. Recent advances in machine learning have exhibited promise, but the performance of existing solvers relies heavily on abundant labeled data and domain expert priors, thus hindering practical applications in mathematics and physics due to a lack of accuracy and robustness. This work considers the challenging task of reconstructing the PDE system from observations alone without any domain expert priors. We propose an end-to-end framework, named OPDE, for tackling this challenge. The workflow of OPDE consists of four modules: Data Encoder that converts observations into latent representations, Function Encoder that encodes symbolic candidate functions to queries, Decoder with Cross-Attention for conditioning queries on observation representations, and Dual Head that jointly predicts structure and coefficient identifications. The whole model is supervised by a combination loss relative to multi-label classification and coefficient regression, with a regularizer that fuses temporal chunks and consistency losses for enforcing prediction consistency across chunks. We conduct experiments on a collection of PDE systems to evaluate the effectiveness and robustness of OPDE in comparison with six baselines, which contains million PDE systems and GB observable data. Compared to six baselines, our method achieves the best performance, where OPDE improves 91.9% F1 and 52.42% EM and reduces 94.8% L2RE rather than the fine-tuned LLMs, with only 5.1% of the LLM's total parameters. Even corrupting 5% and 10% Gaussian noises to observations, OPDE improves F1 by 21.7% and 18.1%, while reducing L2RE by 56.4% and 47.1%, respectively, compared to the SOTA baseline. This is a laudable result, not only confirming the reconstruction advances of OPDE but also opening the door of developing foundation models for inverse PDE problem solving.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Foundation model, Inverse PDE problems, Physics-informed machine learning, Structure identification, Coefficient identification, Dual head

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_JViTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • Proposes a supervised framework that jointly predicts PDE structures and coefficients from spatiotemporal observations.
  • Combines observation and candidate-function encoders through cross-attention, with separate classification/regression heads and temporal consistency regularization.
  • Evaluates on 1.65 million synthetic coupled PDE systems, targeting inference without problem-specific optimization.
Critical strengths
  • Joint structure and coefficient identification is a meaningful inverse problem, and amortized inference could substantially reduce repeated optimization costs.
  • Avoiding explicit numerical differentiation of observations is attractive for noisy inverse problems.
  • Enforcing consistency across temporal segments is well motivated: different segments should identify the same governing equation. Ablations support its usefulness.
Critical weaknesses
  • The claim of operating “without domain expert priors” is overstated. The predefined candidate library and large corpus of labeled equations provide substantial prior information.
  • Noise robustness should be distinguished from relative baseline superiority. Exact match drops from 52.42% without noise to 1.31% and 0.24% at 5% and 10% noise, respectively.
  • Evaluation primarily uses random splits from the same wave-system generator. Generalization to different PDE families and the dynamical fidelity of recovered equations remain insufficiently established.
Questions for authors
  • Under what conditions are the structure and coefficients identifiable from a single trajectory? Would observations under multiple initial conditions improve identification?
  • When the recovered PDE is solved under unseen initial conditions, what are its solution errors and long-term stability properties?
  • How does the method behave when the true equation contains terms outside the candidate library or unseen combinations of terms?
  • For missed small-coefficient terms, can the authors measure the resulting solution error rather than only structural disagreement?
Reviewer_PbMqOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes O2 PDE, an end-to-end framework for recovering PDE structures and coefficients directly from observations. The framework combines a data encoder, a function encoder, a cross-attention decoder, and dual prediction heads to jointly identify PDE structures and coefficients. It further incorporates temporal chunking and consistency regularization to improve prediction consistency. Experiments are conducted on a large-scale synthetic dataset containing 1.65 million PDE systems, with comparisons against six baselines under clean and noisy observations. The results demonstrate competitive performance in PDE structure and coefficient identification.

Critical strengths

The method consistently outperforms existing baselines in PDE structure and coefficient identification, demonstrating its effectiveness in PDE reconstruction. The framework jointly identifies PDE structures and coefficients directly from observations without system-specific optimization, offering a promising approach to scalable PDE discovery.

Critical weaknesses

Most baselines are relatively outdated. Although DISCOVER (2024) and EqGPT (2025) are discussed, they are not evaluated. Other relevant works, such as Mechanistic PDE Networks (ICML 2025) and Discovering Symbolic Differential Equations with Symmetry Invariants (TMLR 2026), are also omitted. More recent comparisons or clear justifications are needed. The key contributions are not clearly articulated in the Introduction. The Introduction mainly describes the framework and its empirical results, without clearly identifying the specific research gaps addressed by this work. Moreover, the technical novelty and distinctions from existing end-to-end PDE discovery methods are not sufficiently explained. The proposed framework mainly combines existing components, including CNNs, Transformers, cross-attention, and dual prediction heads, without introducing substantial architectural innovations. Moreover, the ablation results show only marginal improvements from the Function Encoder, raising concerns about its necessity and contribution.

Questions for authors

Q1: According to Table 5, the Function Encoder provides only marginal improvements, with F1 increasing from 0.9609 to 0.9618 and EM from 51.96% to 52.42%. Could these gains be attributed to increased parameter counts rather than the proposed design? Q2: The paper emphasizes PDE discovery without domain expert priors. However, the proposed method still relies on a predefined candidate function library that contains the ground-truth PDE terms. Does this contradict the claimed prior-free setting? Furthermore, how well can the model generalize to unseen PDE structures or terms outside the predefined library? Q3. The paper claims robustness to observational noise. However, Table 2 shows substantial performance degradation under 5% and 10% noise: F1 drops from 0.9618 to 0.6934 and 0.6075, EM decreases from 52.42% to 1.31% and 0.24%, and L2RE increases from 0.0754 to 0.6328 and 0.7760, respectively. How do the authors reconcile these significant degradations with the claimed robustness?

Reviewer_7eL9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the joint identification of PDE structures and coefficients from spatiotemporal observations. O²PDE combines an observation encoder, a candidate-function encoder, a cross-attention decoder, and classification and regression heads to formulate equation discovery as supervised learning. Temporal chunking and consistency regularization are introduced to improve prediction. The authors generate 1.65 million systems based on two-dimensional multivariable wave equations and compare their method against six baselines using a shared candidate-function library. They also report Gaussian-noise evaluations, model-scaling experiments, and component ablations. The results demonstrate substantial improvements on clean data from the constructed distribution, although recovering complete equation structures remains difficult under noise.

Critical strengths

1.Formulating multivariable equation discovery as joint prediction with direct inference has practical value. On clean observations, Table 2 reports an F1 score of 0.9618, an exact-match rate of 52.42%, and an L2RE of 0.0754, substantially outperforming the evaluated baselines. These results suggest that priors learned across many systems can effectively replace part of the optimization otherwise required for each individual system, making this a worthwhile research direction.

2.The dataset and component analyses provide a useful foundation for further research. The experiments cover different numbers of variables and use a shared candidate-function library for library-based methods. The ablations in Table 5 show meaningful contributions from temporal chunking and the dual-head design to complete structure recovery. The appendix also examines the difficulty of identifying terms with small coefficients, helping explain failure modes.

Critical weaknesses

1.The methodological contribution beyond a conventional supervised architecture is not sufficiently established. The core components consist primarily of CNNs, Transformers, cross-attention, and multitask prediction. In Table 5, replacing the function encoder with fixed learnable queries only reduces F1 from 0.9618 to 0.9609 and exact match from 52.42% to 51.96%. This provides limited evidence that this central design contributes substantially. Comparisons with methods that optimize separately for each system also conflate architectural benefits with the benefits of large-scale labeled pretraining. A LoRA-finetuned vision-language model operating on heatmaps is not an adequate substitute for a supervised baseline using the same training data, numerical inputs, and a comparable model capacity.

2.The generalization and robustness evidence does not support the broad positioning of recovering PDEs “from observations alone” without domain priors. Section 3.1 explicitly uses an empirically constructed candidate-function library, and the training and test sets in Appendix A follow the same modified MV-Wave generation procedure. Distinct random seeds and duplicate checks prevent identical instances from appearing across splits, but do not establish generalization to unseen equation structures or data-generating distributions. Furthermore, Table 2 shows that exact match drops from 52.42% on clean data to 1.31% and 0.24% under 5% and 10% noise, respectively; O²PDE also does not achieve the best exact-match result at either noise level. Better F1 and coefficient errors therefore do not imply reliable recovery of complete equations.

Questions for authors

1.Novelty validation: What performance is achieved by a standard Transformer with structure and coefficient prediction heads, trained on the same data with the same numerical inputs and a comparable parameter count? Is the small improvement from the function encoder consistent across independent training runs?

2.Generalization clarification: How many equation structures in the current test set are absent from the training set? Could a structure-disjoint test split be constructed within the existing generator to distinguish identifying new coefficient instances from recovering unseen structures?

3.Robustness clarification: Are the very low exact-match rates under noise mainly caused by missing small-coefficient terms, introducing spurious terms, or threshold miscalibration? If the discovered equations are used for forward simulation, do their prediction errors still support practical usefulness?

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

Related papers

Discussion (0)

Sign in to comment.