acceptodds
Under review as a conference paper at ICLR 2027

CARE-PDE: Robust PDE Discovery under Noisy Data–Equation Correspondence

Abstract

Data-driven partial differential equation (PDE) discovery aims to recover explicit governing equations from spatiotemporal observations. Existing methods are primarily limited to clean data or data corrupted by simple, single-source noise. In practice, however, multiple noise sources can produce heterogeneous contamination, causing some observations to deviate substantially from the underlying dynamics and weakening their correspondence with the governing equation. We formulate this problem as Noisy Data–Equation Correspondence (NDEC) and propose the Correspondence-Aware Robust Equation Discovery for PDEs (CARE-PDE) framework. Specifically, we develop Reliability-Aware Field Reconstruction (RAFR) by combining a novel Cauchy loss with adaptive reliability estimation to reduce the influence of heavily corrupted observations. Building on these estimations, Reliability-Guided Local Legendre Projection (RGLP) further exploits the estimated reliability to construct reliable local supports and recover stable derivatives through orthogonal projection. Local Integral Constraint Aggregation (LICA) is then presented to aggregate the reconstructed strong-form relations into local integral constraints, reducing sensitivity to pointwise estimation errors. Finally, Stability-Selected Sparse Regression (SSR) identifies consistently supported terms to recover sparse governing equations. Experiments on 12 PDE systems across different contamination scenarios show that CARE-PDE recovers the correct equation support in all 36 tasks and achieves lower true-term coefficient errors than the ten state-of-the-art baselines, demonstrating its effectiveness and robustness under the tested NDEC conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Robust PDE discovery, Noisy Data--Equation Correspondence, Reliability estimation

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_N1CnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

CARE-PDE discovers governing PDEs from observations with sample-level heterogeneous noise. It combines Cauchy-loss field reconstruction with GMM reliability estimation, reliability-guided local Legendre projection for derivatives, local integral constraints, and stability-selected sparse regression. On 12 synthetic PDE systems it recovers the exact support in all 36 main tasks and reports lower true-term coefficient errors than ten baselines, with additional results on harder noise settings, derivative quality, parameter sensitivity and ablations.

Critical strengths
  1. The method addresses a real difficulty in noisy equation discovery and propagates observation reliability coherently through reconstruction, derivative estimation and sparse identification. Regularizing the field before differentiation is well motivated by the instability of high-order derivatives.
  2. The evaluation spans 12 systems, including high-order and coupled dynamics, with one shared configuration. The derivative-quality diagnostics and ablations support the combination of projection and integral constraints; removing LICA reduces exact recovery from 36/36 to 28/36 (Table 3).
Critical weaknesses
  1. The evaluation protocol leaves model selection unclear. Only the training split is corrupted while the 20% validation split stays clean (Sec. 4.1, Eqs. 17–18), and Sec. 3.5 keeps the H checkpoints with the lowest reconstruction error without stating on which data this error is computed; if it is the clean split, selection uses clean targets that the NDEC setting does not provide. Independent-run variability and baseline tuning are not reported, and some entries are inconsistent: DeepMoD on Chaffee–Infante at (0%, 50%, 25%) has Etrue = 100% but Efull = 0 (Table 10), and Burgers2D at β = 50%, α = 0.25 lists P = 1.00, R = 0.50, F1 = 1.00 (Table 16).
  2. The experiments use only randomly assigned two-level Gaussian noise (Eqs. 17–18), which covers the motivating contamination scenarios only partially. Removing AGRE still recovers 35/36 tasks, and removing RGLP replaces projection with automatic differentiation altogether (Table 3), so no ablation isolates reliability-adaptive windows from ordinary fixed-window polynomial smoothing. The harder-setting failures, e.g. Chaffee–Infante at (5%, 25%, 50%) with F1 = 0.67 (Table 10) and λ–ω at (0%, 50%, 50%) with F1 = 0.68 (Table 12), are not discussed.
Questions for authors
  1. (Clarification) On which data is the reconstruction error used to retain the H checkpoints computed, and do clean validation targets influence any selection step? How were the shared hyperparameters and the baseline settings chosen? Please report variability over independent runs and correct the entries in Tables 10 and 16.
  2. (Clarification) How do reliability-adaptive windows compare with fixed-window projection under otherwise identical settings? What causes the harder-setting failures above, and is there evidence of robustness beyond two-level Gaussian noise? Is the adaptive Cauchy scale detached during optimization, as the gradient analysis in Appendix C assumes?
1 comment
Reviewer_N1CnOpenReview ↗

Response to Authors

Thank you for the detailed responses and additional experiments, which have addressed all of my concerns regarding the evaluation protocol, multi-seed variability, ablations and implementation details. I am raising Soundness from 2 to 3 and my overall rating from 2 to 3.

Reviewer_nXBVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper combines robust field reconstruction with observation reliability estimation and local polynomial projection to recover governing PDEs from heterogeneously corrupted observations and uses integral constraints with stability-based sparse regression to select equation terms and estimate their coefficients.

Critical strengths
  1. The pipeline is well motivated. Errors in the reconstructed field can become much larger after differentiation and then affect equation identification. The method addresses these stages through robust reconstruction and local projection before forming integral constraints. Using the estimated observation reliability in both local window selection and regression also gives the components a clear connection.

  2. The experimental study is extensive. It covers different equation types and contamination settings with a shared configuration. The appendix includes component ablations and their combinations. It also examines reliability separation and derivative quality. Reporting both equation structure and coefficient errors helps assess whether the recovered equations are correct beyond a small fitting residual.

Critical weaknesses
  1. I am not clear how the adaptive residual scale is handled during optimization. The scale is calculated from the current reconstruction residuals, but the robustness analysis differentiates the loss as if this scale were fixed. The paper does not state whether gradients are stopped through the scale estimate. This matters because when the scale is determined by the residual median, multiplying all residuals by the same factor leaves the written normalized loss unchanged. Differentiating through the scale can therefore give different updates from those used in the robustness analysis. The scale update and gradient convention need to be specified so that the analysis matches the actual training procedure.

  2. The checkpoint-selection protocol is not clear enough. The described data split adds contamination only to the training observations, which appears to leave the validation observations clean. The method then retains checkpoints with the lowest reconstruction errors but does not specify which observations and loss are used to calculate these errors. The final equation-selection score is unsupervised, but this does not explain how the earlier checkpoint shortlist is formed. If clean validation observations guide that shortlist, they could contribute to the reported robustness. I cannot conclude that this happened from the paper. However, the current description does not establish whether the complete procedure works without clean observations for model selection.

Questions for authors
  1. Is the adaptive residual scale treated as a constant during each gradient update? Please explain when it is recomputed and whether gradients pass through the median-based estimate. The robustness analysis should explicitly match this choice. If gradients pass through the scale, please explain the effect of the common-rescaling invariance and check whether it can encourage larger reconstruction errors while reducing the normalized objective.

  2. What observations and reconstruction loss are used to select the retained checkpoints? Please clarify whether clean validation data are used for checkpoint selection or any other tuning step. If they are used, please compare with selection based only on noisy observations and report the effect on equation structure and coefficient accuracy. If they are not used, an explicit description of the selection rule would resolve this concern.

If these concerns are answered well with clear explanations and supporting evidence, I would be willing to change my recommendation to acceptance.

Reviewer_L1CHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies PDE discovery under heterogeneous observation noise, describing the setting as Noisy Data–Equation Correspondence. CARE-PDE combines a Cauchy loss with residual mixture modeling to estimate observation reliability, then uses this reliability in local support selection, Legendre projection, integral-constraint construction, and stability-selected sparse regression. The authors evaluate the method on 12 PDE systems and recover the correct equation support in all 36 system–contamination settings in the main experiments. The appendix further provides evaluations under additional noise settings, full coefficient errors, reliability-estimation analyses, local-feature diagnostics, hyperparameter sensitivity studies, and combinatorial ablations.

Critical strengths

1.The propagation of reliability throughout the discovery pipeline is coherent and well motivated. The method goes beyond applying a robust loss during reconstruction: it also uses reliability to determine local windows, weight constraints, and identify equation structures. This directly addresses the amplification of corrupted observations through derivative estimation and sparse regression. In Table 3, removing RAFR, RGLP, or LICA reduces the number of exactly recovered tasks to 30/36, 33/36, and 28/36, respectively, indicating that these stages materially affect the final results.

2.The experimental coverage and mechanism-oriented analyses are substantial. The 12 systems include higher-order derivatives, two-dimensional spatial domains, and coupled variables, with comparisons against ten relevant methods. The appendix reports full coefficient errors that penalize spurious terms, ablates all (2^5) component configurations, and examines derivative quality on the KS equation. These analyses provide stronger support for the method’s effectiveness than final performance metrics alone.

Critical weaknesses

1.The novelty should be delineated more precisely, as some novelty claims are overstated. The current NDEC setting concerns heterogeneous observation corruption under a shared governing law. Cauchy robust losses, residual mixture modeling, local polynomial approximation, and stability selection are established ideas. The most convincing contribution is therefore the propagation of reliability and the coordinated use of these tools for PDE discovery. The existing ablations show that the components are useful, but do not fully establish whether the proposed RGLP–LICA pathway offers a distinctive advantage over combining the same robust reconstruction front end with an established weak-form identification method.

2.Unresolved details of the experimental protocol affect practical applicability and reproducibility. Section 4.1 explicitly describes adding noise only to the training set after splitting the observations into training and validation sets. It is therefore important to clarify whether the validation set remains clean and how it is used for checkpoint selection and baseline tuning. Although the appendix includes settings where the lower-noise group is also corrupted, the noise remains primarily an independent Gaussian mixture at randomly selected locations. This does not fully represent spatially clustered sensor failures or systematic bias. The main tables also do not report variability across observation sampling, noise realizations, and training randomness, so the 36/36 recovery result should be interpreted as applying to the reported configurations.

Questions for authors

1.Experimental-protocol clarification: Is the validation set noisy? Which dataset is used to compute the reconstruction errors for retaining the five checkpoints? Do all baselines use the same information for model selection? Were the shared values of (c), (\alpha), and the window settings selected using ground-truth equation-recovery performance on the systems included in the final evaluation?

2.Stability validation: Could the authors repeat one main contamination setting on the more challenging KS and λ–ω systems using several independent observation samples, noise realizations, and training runs, and report exact-recovery rates and coefficient-error variability?

3.Contribution attribution: With the same robustly reconstructed field, how much does RGLP–LICA improve over directly constructing weak-form constraints? This comparison would help establish the necessity of the proposed methodological design.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.