acceptodds
Under review as a conference paper at ICLR 2027

PDE-LENS: Scaling End-to-End PDE Discovery through Spectral Representation Learning

Abstract

End-to-end PDE discovery maps an observed solution field to its governing equation. We study the end-to-end models that learn this map and find a recurring failure: they capture the slowly varying background of the field but miss the sharp fronts, narrow peaks, and oscillations that separate one candidate equation from another. We call this high-frequency neglect and locate the failure in the field encoder. PDE-LENS addresses it with PRISM, an encoder that operates directly on scattered observations in the spectral domain, using non-equispaced Fourier analysis. To train at scale, we build a quality-controlled dataset of 1.19 million numerically solved PDE trajectories from a library of 59 atomic terms, with LLM-assisted solver generation handling stiff or ill-posed cases. On unseen four-term equations, PDE-LENS reaches a median NMSE more than sixfold lower than PDE-FIND, the strongest non-pretrained baseline. Across 99 unseen trajectories over ten equation structures, PRISM recovers 65 within NMSE , against 50 for NeSymReS and 34 for a ViT encoder. For conventional symbolic regression, our encoder also achieved high performance. On the Feynman benchmark it also raises conventional symbolic-regression accuracy and high-frequency fidelity over E2E-SR and NeSymReS. PRISM's 256-dimensional latent representations also predict four dynamical properties with the best on three of four. Equation discovery thus yields not only the governing equation but a reusable, structured PDE latent space.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE discovery, symbolic regression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_3y6POpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PDE-LENS, a framework for discovering governing PDEs from observed fields. The authors identify a high-frequency neglect issue in current neural encoders. To address this, they introduce PRISM, a spectral encoder that uses non-uniform fourier transforms and pointwise operations to process scattered observations. The encoder is combined with a transformer based decoder to generate symbolic PDE expressions, followed by numerical coefficient refinement. The authors also contruct a synthetic dataset containing about 1.19 million pde trajectories. Experiments on MDBench, 10 unseen four-term PDE trajectories show improvements over several baselines.

Critical strengths
  1. Learning to recover governing PDES is a long existing direction, and the proposed large-scale pretraining framework offers an alternative to traditional equation discovery methods (e.g., SINDy). The construction of a large synthetic PDE dataset is also useful

  2. The proposed spectral encoder shows promising improvements over evaluated neural baselines. The latent space analysis provides some additional evidence that the learned representations contain information about underlying dynamics.

Critical weaknesses
  1. The methodological novelty is relatively limited, and the contributions of the proposed components are not sufficiently established. The main spectral operations in PRISM are closely related to existing Fourier neural operators. The symbolic decoder and coefficient fine-tuning are standard practice too. The main methodological contribution is only to adapt current SR model for PDE discovery. However, the paper lacks important ablation studies, particularly for the spectral branch, pointwise branch, and Moment Bridge. So, it is difficult to decide which of these components are responsible for the improvement. The central claim about high-frequency neglect also needs more direct experimental support.
  2. The experimental evaluation is weak to establish the claimed generalization and broader usefulness of the framework. Although the paper emphasizes unseen PDEs, the test equations are largely generated from the same 59 atomic terms used during training, with similar numerical settings. So it is unclear how well the model generalizes to genuinely new PDE or operator families. The unseen four-term benchmark contains only 10 trajectories, and the conventional symbolic regression evaluation uses only 15 Feynman equations. In addition, the paper does not have systematic evaluations under noisy or sparse observations. Furthermore, the field reconstruction evaluation uses the observed states rather than autonomous rollouts, so it is unclear whether the recovered equations can reproduce the correct dynamics when solved independently.
Questions for authors
  1. How well does PDE-LENS generalize to PDEs beyond the 59 atomic terms seen in training? The unseen 4-term equations are still built from same function library. Can you evaluate held-out operator PDE structures?
  2. Can the authors provide more detailed ablation studies of PRISM? (what are individual contributions of spectral branch, pointwise branch, and moment bridge?)
  3. What evidence shows high-frequency neglect is the main reason for poor performance of existing encoders? The results in the paper show better high-frequency recovery, but do not prove the underlying mechanism. Can you evaluate PDE pairs with similar low-freq behavior but different high-freq structures, or study how the results change with number of fourier modes?
  4. Can authors provide a more comprehensive eval of robustness and scaling? the current four-term benchmark contains only 10 trajectories, which is not sufficient to show performance improvements.
  5. A broader discussion of related work on symbolic representations for PDE solving and multimodal PDE foundation models could help better position this paper. 1. Finite expression method for solving committor problems (https://epubs.siam.org/doi/abs/10.1137/23M1583612), 2 Multi scale finite expression method for PDE (https://epubs.siam.org/doi/abs/10.1137/25M1816218) 3. PDE foundation model for scientific text description (https://arxiv.org/pdf/2502.06026)
Reviewer_BRCpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies large-scale pretraining for end-to-end PDE discovery, framed as conditional sequence generation from observed solution fields to symbolic equations. The authors identify high-frequency neglect: existing field encoders fail to generalize their recognition of sharp, localized or oscillatory structures that discriminate between candidate equations. They propose PRISM, an encoder that applies learned mode-wise channel operators to a non-equispaced Fourier transform of scattered observations, combines it with a pointwise branch, and aggregates the result through a weighted Moment Bridge. To enable pretraining, they construct a quality-controlled corpus of 1.19M trajectories from a 59-term atomic library, using LLM-assisted solver generation (CodePDE) for stiff cases.

Critical strengths
  1. The encoder comparison is controlled and shows a consistent gain. Encoders are swapped inside the same end-to-end framework, with the same decoder, data and refinement. Under that setup, PRISM recovers 65 of 99 unseen trajectories at S_RHS ≥ 0.99, against 50 for NeSymReS and 34 for ViT. Paired Wilcoxon tests are significant (p = 3.4e-3 and 2.5e-9), and the gain shows up across six of ten equation structures rather than a single family. The HF metrics move the same way (HF relative error 0.270 vs 0.494 / 0.612). This is a convincing evidence that the encoder design matters for PDE discovery.
  2. The data pipeline is useful if released.
Critical weaknesses
  1. The main explanation is not isolated by ablations. The paper’s central claim is that the spectral design of PRISM fixes high-frequency neglect. But there is no ablation that removes the spectral branch, changes the mode lattice. Without ablations, the gains could come from other factors: inductive bias in general, the pooling design, or capacity differences.
  2. The evaluation is narrow and mostly in-distribution. Benchmarks are 1D, periodic and noise-free. Can the authors extend to 2D/3D and non-periodic boundary conditions? There is no noise experiment, even though noise robustness is a central practical concern in PDE discovery.
  3. What are the parameter counts and training compute of PRISM, NeSymReS and ViT in the PDE experiments?
  4. The comparison set leaves out several directly relevant baselines, including MDBench, EqGPT, DISCOVER, and plain FNO/DSE-FNO encoder.
Questions for authors

See weaknesses above.

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

This work proposes an end-to-end framework for discovering the PDEs from observed solution fields. The paper introduces the PRISM encoder to address the limitations of existing field encoders in capturing high frequency features in solution fields. This encoder combines spectral-domain processing, point-wise processing, and a moment-based aggregation mechanism to preserves such features more effectively. The framework is pre-trained on 1.19 million PDE trajectories and uses a Transformer decoder to generate symbolic equations, and then separately refines the numerical coefficients. The paper evaluates the proposed framework on MDBench, unseen four-term equations, and conventional symbolic regression tasks. The results demonstrate improved equation recovery accuracy and high-frequency fidelity compared with existing baselines. The authors also analyze the latent space of PRISM and show that it captures information related to meaningful dynamical properties of PDE systems.

Critical strengths
  1. The paper identifies high-frequency neglect as an important limitation in existing end-to-end PDE discovery frameworks. The study observes that sharp fronts, narrow peaks, and oscillations can contain important information for distinguishing governing equations, which provides a clear motivation for improving the field encoder.
  2. The proposed PRISM encoder combines spectral and pointwise processing with a moment-based aggregation mechanism, enabling more effective capture high-frequency and localized features.
  3. The paper constructs a large-scale pretraining dataset containing 1.19 million numerically solved PDE trajectories. The use of automated solver generation and numerical accuracy checks provide a valuable resource for large-scale PDE discovery.
Critical weaknesses
  1. The contributions of PRISM are not sufficiently isolated. Specifically, the paper does not clearly evaluate the individual effects of the spectral branch, pointwise processing, and Moment Bridge. Conducting additional ablation studies would clarify which components are primarily responsible for the observed improvements in PDE discovery and high-frequency fidelity.
  2. The paper identifies high-frequency neglect as a key limitation of existing end-to-end PDE discovery methods and demonstrates that PRISM improves high-frequency fidelity. However, the current experiments do not fully clarify whether the improvements in equation recovery are primarily attributable to better preservation of high-frequency information or other differences in the encoder representations.
  3. In the main experiment, the evaluation of equation recovery is based on normalized prediction errors and the reconstruction accuracy of the right-hand-side. However, a low reconstruction error does not necessarily indicate that the correct structure of the governing equation has been accurately identified. Evaluating the exact equation structure recovery and coefficient estimation accuracy would provide more evidence for the effectiveness of the proposed PDE discovery framework.
Questions for authors
  1. The framework is primarily trained and evaluated based on 1D PDEs. Can the authors discuss whether the proposed method can be extended to 2D or 3D problems?
  2. The experiments mainly use clean numerical solutions, while practical PDE discovery often involves measurement noise, sparse observations, and irregular sampling. Could the authors further evaluate the robustness of PRISM under these more realistic observational conditions?
  3. The PRISM encoder uses a 32*32 signed-mode lattice. Could the authors clarify the motivation for this choice? How sensitive are the equation discovery accuracy and high-frequency fidelity to the number of retained Fourier modes?
Reviewer_z8iJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an end-to-end approach to discovering symbolic PDEs from observed solution fields. Its PRISM encoder combines spectral processing and pointwise transformations to retain information useful for distinguishing candidate equations; a decoder proposes a symbolic expression, and its coefficients are refined against a derivative residual. The authors generate more than one million filtered trajectories for training and compare PRISM with several alternative encoders.

Critical strengths

The data-generation pipeline is substantial and includes explicit numerical screening. The authors report retaining approximately 1.19 million trajectories after checks for numerical stability and spectral behavior.

The evaluation compares encoders across a broader diagnostic set than a handful of standard PDE examples. On 99 trajectories, PRISM achieves an RHS score of at least 0.99 on 65 cases, compared with 50 for NeSymReS and 34 for the tested ViT encoder. The paper also reports spectral and representation analyses, and explicitly describes the limitations of its reconstruction procedure.

Critical weaknesses

The results do not yet establish recovery of the governing symbolic equation. The main metrics assess how well a fitted right-hand side matches derivatives or reconstructs observed trajectories. In the reconstruction experiment, the observed field is supplied at every step; this is not an autonomous rollout. Different expressions may agree along one trajectory without representing the same governing operator. The paper would be stronger with representative recovered expressions, structural or symbolic-equivalence metrics, coefficient errors, and tests on new initial conditions.

Generalization and the source of the gains are insufficiently isolated. The paper mentions a strict test partition but does not clearly specify whether equation structures and coefficient configurations are disjoint from training. This is a reporting gap, not evidence of leakage. In addition, PRISM changes several components together, while the comparisons do not isolate the spectral layers, moment aggregation, or training compute. It is therefore difficult to determine whether high-frequency preservation itself causes the improvement.

Questions for authors

What are the exact train/test split units and counts? Are test equation structures, coefficients, and initial conditions independently held out, and how are equivalent expressions handled? Can the authors report symbolic or structural recovery rates and representative recovered equations, together with autonomous rollouts or tests on new initial conditions? How much does each component of PRISM contribute under matched parameter and compute budgets? Was the optional consistency loss used in each experiment? Can the full term library, sampling distributions, solver-validation checks, and training settings be released or documented? How many retained trajectories depend on the generated-solver fallback? How robust is the method to noise, sparse observations, different resolutions, and boundary conditions outside the training distribution?

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted their rebuttals. As the author-reviewer discussion period is coming to a close soon, please read their responses and engage in the discussion. Thank you for your time and effort.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.