acceptodds
Under review as a conference paper at ICLR 2027

PS-PINN: Recovering Governing Equations and Readable Solutions from the Same Noisy Data

Abstract

Recovering governing equations and symbolic solutions from the same noisy observations offers interpretable descriptions of both physical dynamics and their observed realizations. Existing approaches typically recover either an explicit PDE with a neural surrogate or a readable solution under a known equation. We present \ourmodel, a data-driven framework for PDE discovery and symbolic solution discovery. The reconstructed solution combines a sparsity-regularized analytic core with a neural corrector, . A symbolic-first curriculum fits the core before learning neural corrections, giving explicit structure priority during training. The framework delivers both outputs in identifiable settings. On Wave with a dictionary encoding the true wavenumber, the corrector becomes negligible and the solution is effectively closed form; on Burgers, the core captures the smooth bulk while the corrector concentrates near the steep front. Coupling an Allen–Cahn operator estimated within a prescribed structure reduces reconstruction RMSE by a factor of relative to uncoupled training. Matched-budget controls further show that Adam-based joint training yields symbolic-core RMSE values 2.3 to 5.7 times those of symbolic-first training on Allen–Cahn and Burgers, often with little change in prediction error. These results show how symbolic-first solution learning complements equation discovery, retaining an explicit governing law and an inspectable solution structure.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
scientific machine learning, physics-informed neural networks, symbolic regression, PDE discovery, neuro-symbolic learning, interpretable modeling

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_KbAjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the PS-PINN (Physics–Symbolic Physics-Informed Neural Network) framework to recover the governing PDE equation (analytical model) as well as the solution from a noisy data set. The solution is represented in terms of two components as shown in equation (1): the first term is a linear sparse combination of terms selected from a dictionary while the second term, called neural corrector, is used to capture the residual left unmodeled by the first term. Figure 1 provide a general overview of the proposed approach. Overall, the paper is written in a very superficial way. No focus at all. The Abstract and Introduction section (like the majority of sections) are written in a style that makes almost impossible to understand what is the problem definition, the strategy adopted to solve the problem, and the merits of the proposed PS-PINN. The paper fails to bring any major technical contributions. In terms of technical novelty, this paper is very weak--no solid result. This is a clear reject case.

Critical strengths

None. Sorry this is a very superficial paper that fails to bring any notable contribution.

Critical weaknesses

The presentation style is very superficial. The paper is not focused at all. Very superficial writing.

No substantial technical contributions. The authors should be more careful in presenting well known "textbook"-level material as new contributions. What is the purpose of presenting Appendix O, given the fact that all the results in this Appendix are well known on the literature? Please check a textbook in Bayesian statistics that addresses Lasso estimation and you will see these results. This gives the impression of AI generated results generated blindly without checking the literature.

Everything in this paper seems to be argued and presented in a very ad-hoc and superficial manner. The authors are asked to provide more rigorous explanations and formal analytical proofs for their claims. Next I will illustrate only a few of the many statements that are not at all justified and are very weird. These statements should not be made without a rigorous justification.

On p. 5, you state "The symbolic core represents global structure using the prescribed basis functions: periodicity, separability, exponential decay, polynomial and trigonometric shape, and the smooth bulk of the field." What you mean by "the smooth bulk of the field"? How can you ensure the generalizability or scalability of the dictionary in general settings, given the diversity of PDEs dynamics?

On p. 5, a few lines down "The neural corrector is trained to approximate residual features that are not adequately represented by the dictionary, including steep fronts and unresolved details; the curriculum encourages this allocation but does not guarantee spatial localization." None of these claims is justified. It is hard to understand how the "steep fronts and unresolved details" can be captured; not clear why "the curriculum encourages this allocation". The curriculum may not have any knowledge of the regions of high dynamics.

Four lines below eq (6), "Whether the physics term is active, and which Fc it uses, is stated per experiment: the recovered rule of Sec. 4.1, a rule with prescribed terms and estimated coefficients, or the true rule, used to study the solution branch apart from discovery error." ; it is not clear how feasible/reliable are these claims.

Questions for authors

The problem addressed requires substantial technical contributions to make its conclusions generalizable to a larger class of PDEs. A formal analytical framework is needed to justify all the claims made and to offer the necessary mathematical rigor to ensure the validity of claimed results to a larger class of PDEs. The paper needs to be rewritten in much more focused style, easy to understand and with all the claims justified.

Reviewer_ESAeOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PS-PINN, a framework for recovering governing PDEs and symbolic neural solution representations from noisy observations. The method uses standard sparse regression for PDE discovery and represents solutions a s a sparse combination basis functions from a predefined library with a neural residual modeled by a MLP. The authors introduce a symbolic-first training schema to avoid neural component absorbing the symbolic structure. Authors performed experiments on five synthetic PDEs, showing that this method can retain better symbolic core than joint training, while achieving comparable recovery accuracy.

Critical strengths

The paper addresses an interesting problem in scientific ML, namely recovering both governing equations and interpretable solutions from noisy data. The analysis of the ambiguity between symbolic learning and neural components is useful. The experiments also include several comparisons, especially the training-order ablations between symbolic and neural residual. This supports the claim that prediction accuracy alone does not reflect the quality of the symbolic component.

Critical weaknesses
  1. Limited methodological novelty. The PDE discovery component relies on existing sparse regression methods (e.g, SINDy), and the solution representation combines pre-determined library functions with a MLP neural residual. In addition, symbolic-first training strategy is quite similar to staged optimization iin the neuro-symbolic learning literature, such as PINSN. Therefore, the overall methodological contribution is relatively incremental.

  2. Strong dependence on pre-defined symbolic dictionaries. The results in paper heavly rely on prior knowledge about the solutions. For example, the Wave dictionaries already contain the true (or nearly correct frequencies), and the 2d wave dictionary includes the true spatial and temporal frequencies. Appendix L also shows that larger dictionaries can produce nearly dense expressions.

  3. Limited evidence for meaningful physical interpretability. Although the recovered symbolic components are explicit expressions, I am not convinced whether they capture meaninful physical structures in more challenging PDEs. For burgers, the symbolic only RMSE is about 0.279, and 0.0076 for hybrid model, while the neural-only model achieves comparable accuracy. For Allen-Cahn, 18 out of 19 symbolic candidates remain active. These results raise a question: how much useful physical information the symbolic part actually provides?

  4. The advantage of the complete framework is not clear. Appendix K shows post hoc decomposition achieves statistically similar symbolic-core accuracy on Burgers and allen-Cahn. This weakens the claimed advantage of symbolic first training over simple baselines. In addition, the strong Wave extrapolation results also benefit from predefined frequencies close to the ground truth.

Questions for authors
  1. How does PS-PINN compare directly with PINSN and other related neuro-symbolic methods, given comparable solution dictionaries and training budgets?
  2. How powerful/strong is PS-PINN? Can you show successful symbolic recovery when the correct solution structure is not included in the predefined dictionary?
  3. Beyond core-only RMSE, what else shows the recovered symbolic expressions capture physically meaningful/stable structures?
  4. Given the post-hoc decomposition produces comparable symbolic cores on Burgers and Allen-Cahn, what is the main practical advantage of symbolic first curriculum in such more challenging settings?
  5. Symbolic learning is a important area in scientific machine learning, especially for interpretable solutions for PDEs. Stronger discussion of related work could further strength the impact of the paper (Finite Expression method for solving committor problems, https://epubs.siam.org/doi/abs/10.1137/23M1583612), and multi scale Finite expression method (https://epubs.siam.org/doi/abs/10.1137/25M1816218)
Reviewer_kKgJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce PS-PINN, which estimates a governing PDE from noisy observations and reconstructs the observed solution. They do this using an explicit symbolic core plus a neural corrector. The central argument is that prediction accuracy alone cannot determine how to fit each component when trained jointly. Instead, PS-PINN first trains a symbolic “core” then fits the neural corrector. They evaluate on 5 synthetic PDEs and conduct diagnostics analyzing the learned symbolic core.

Critical strengths
  1. Clear motivation and situation in relevant works. The introduction and methodology clearly distinguish governing-equation discovery from recovering a readable solution. Proposition 1 provides a concise explanation of why the objective does not favor preserving the symbolic core, and clearly sets up the problem setting.
  2. Informative analysis. Figure 2 provides useful evidence that the symbolic component captures meaningful structure within the supplied dictionary, while the neural component addresses features the dictionary cannot represent. I appreciate the reporting of core-only error, corrector share, branch overlap, and localization, rather than evaluating solely through reconstruction.
  3. Training ablations. The stage-by-stage results, Table 5, help explain what each stage of training contributes. Comparisons with joint training, full-budget two-stage training, and proximal coefficient helps strengthen the analysis.
Critical weaknesses
  1. Solution priors. The dictionaries are hand picked per system, which makes it difficult to disentangle the choice of dictionary and the method’s advancements. While the work mainly uses symbolic regression libraries out of the box, analysis of additional dictionary choices would be useful. For example, the Burgers and noisy Allen–Cahn cores retain most candidate terms, limiting practical sparsity.
  2. The stability and benefit of joint refinement is unclear. Stage IV reintroduces the allocation ambiguity motivating the curriculum. Table 5 already shows degradation of the Allen–Cahn core during refinement, while the full-budget two-stage alternative retains a better core. The paper acknowledges this trade-off, but does not establish how a practitioner should select the refinement budget and why stage 4 is necessary part of the curriculum.
  3. Noise robustness. All noise experiments use additive Gaussian white noise. It is not clear if the results would transfer under spatially or temporally correlated noise, or a different noise model. Gaussian smoothing is restricted to equation discovery, but its effect on coefficient bias also deserves clearer emphasis.
  4. Missing bas lines. The paper acknowledges that a PINSN variant already uses symbolic-first, residual-second training, but provides no direct quantitative comparison with PINSN.
Questions for authors
  1. What happens if Stage IV is continued substantially longer? Does symbolic-core accuracy progressively deteriorate? Please distinguish the behavior with the retained L1L_1 penalty on Wave/Allen–Cahn from Burgers, where Stage IV omits this penalty.
  2. Could the authors specify the role of (b_0) more explicitly: which configurations include it, when it is updated or frozen, and how it enters the physics loss and decomposition metrics? Appendix O excludes the trainable intercept from its theorem, so a clearer connection would help.
  3. What primarily explains PS-PINN’s advantage over the oracle PINN in Table 4: the supplied solution dictionary, optimization schedule, or differences in the training objective? The Burgers neural-only ablation nearly matches the hybrid.
  4. How should a practitioner choose the solution dictionary and decide whether to enforce the recovered PDE when the true solution and coefficients are unavailable? How much system specific knowledge is used in designing the experimental protocols?
  5. Could the authors evaluate correlated noise or a second mouse model?
Reviewer_D2JTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PS-PINN (Physics–Symbolic Physics-Informed Neural Network), a pipeline that, from one set of noisy spatio-temporal observations D={(xi,ti,yi)}D = \{(x_i,t_i,y_i)\} (Eq. 2), returns two artifacts: (i) an estimated governing PDE F^\hat F obtained by Gaussian smoothing, finite differences, and sparse regression (STLSQ / PySR) over a candidate term library (Sec. 4.1, Eq. 4; Fig. 1b); and (ii) a hybrid solution representation u^(x,t)=Φ(x,t)α+rθ(x,t)\hat u(x,t)=\Phi(x,t)\alpha + r_\theta(x,t) (Eqs. 1 and 5), where Φ\Phi is a prescribed analytic dictionary, α\alpha is a sparse coefficient vector, and rθr_\theta is a neural corrector. Because the joint objective of Eq. (6) does not identify how accuracy should be allocated between Φα\Phi\alpha and rθr_\theta (Proposition 1), the authors introduce a symbolic-first curriculum (Table 1; Sec. 4.4; Algorithm 1): Stage I fits α\alpha with rθ≡0r_\theta\equiv 0, Stage II trains rθr_\theta on the residual, and optional Stages III–IV refine alternately/jointly. Coupling of F^\hat F into the physics residual Lphys\mathcal L_{\mathrm{phys}} is stated per experiment, not decided automatically (Secs. 4.3, 2 “Discovered-physics reuse”; Limitations, §6). Allocation is monitored with ρNN\rho_{NN} (Eq. 3), overlap γ\gamma, shock localization ρshock\rho_{\mathrm{shock}}, and core-only error. Experiments cover five synthetic PDEs (1D Wave, Burgers, Allen–Cahn, 1D Heat, 2D Wave) under mainly Gaussian noise, with curriculum vs. joint controls (Tables 5, 12), dual-artifact recovery (Table 3), main reconstruction (Table 4), component ablations (Sec. 5.3; Table 16), and Wave domain extrapolation (Table 6).

Critical strengths
  1. Clear diagnosis of hybrid misallocation. Proposition 1 and the surrounding analysis in Sec. 4.4 correctly observe that data/physics terms of Eq. (6) depend only on the sum Φα+rθ\Phi\alpha+r_\theta, while ℓ1\ell_1 penalizes α\alpha alone, so emptying the formula into the network can strictly lower the objective. The curriculum vs. matched joint controls (Table 5B; Table 12) show that similar hybrid RMSE can coexist with much worse core-only RMSE (e.g., Allen–Cahn core 0.4770.477 vs. 0.0860.086; Burgers 0.6900.690 vs. 0.2790.279), which is a useful caution for hybrid scientific models.

  2. Explicit reporting of the split. Metrics ρNN\rho_{NN}, core-only RMSE, and visualizations in Figure 2 / Table 13 make the division of labor inspectable. On minimal Wave, ρNN≈3×10−6\rho_{NN}\approx 3\times 10^{-6} and the deliverable is effectively closed-form within the supplied dictionary; on Burgers, ρNN=0.21\rho_{NN}=0.21 with core-only RMSE ≈0.278\approx 0.278 against hybrid ≈0.0076\approx 0.0076 (Sec. 5.2), which honestly shows when “readability” is only a backbone.

  3. Honest scoping in places. The authors state that “this paper’s contribution lies in what is delivered alongside the equation rather than in the discovery algorithm” (Sec. 4.1), that dictionaries are “physics-informed and hand-specified” (Sec. 4.2; Appendix B), that coupling is an experimental choice (Sec. 4.3; §6 Limitations), and that Wave extrapolation uses a periodicity prior near the true wavenumber (Table 6 caption; Appendix B/M). These admissions help the reader calibrate claims.

Critical weaknesses
  1. Novelty and significance relative to prior hybrid methods are unclear. The main components are established: sparse PDE discovery with STLSQ/PySR, a prescribed analytic dictionary plus neural residual, optional physics losses, and staged or alternating optimization. The main contribution therefore appears to be an integration and evaluation protocol rather than a new discovery or learning principle. The paper should explain more sharply what new knowledge this protocol adds beyond PINN-SR, PINSN, PDE-LEARN, R-DISCOVER, and related hybrid responsibility-allocation methods. The curriculum result is useful, but by itself may be insufficient for the claimed ICLR significance.

  2. Readable-solution claims depend strongly on hand-specified priors. The solution dictionary Φ\Phi is fixed separately for each benchmark and is not learned beyond term selection. The Wave dictionaries use frequencies near the true value, the minimal Wave setting is explicitly oracle-dependent, Heat uses near-truth separable modes, and Allen–Cahn uses tanh-front families. The experiments therefore show recovery under a supplied solution family, not data-alone discovery of an open-form readable solution. Table 6 is a valid extrapolation experiment, but mainly tests extrapolation of a selected, prior-informed structure. This substantially limits the scope of the generalization claims.

  3. The discover-then-enforce loop is protocol-dependent. Section 4.3 leaves both activation of Lphys\mathcal L_{\mathrm{phys}} and operator choice to each experiment. High-noise settings are uncoupled when the recovered operator is biased or non-identifiable, while several representation studies use analytic reference operators to isolate solution training from discovery error. The paper therefore demonstrates selective coupling under a prescribed protocol, but does not provide an automatic criterion for deciding when a recovered equation is reliable enough to enforce. This weakens the claim of a robust closed loop from noisy observations to a discovered and enforced law.

  4. Neural necessity and broad transfer are not established. Table 16 shows that a correction branch helps when the dictionary is incomplete, but compares only symbolic-only, neural-only, and hybrid variants. It does not compare the neural corrector with a non-neural residual such as a grid, spline, or local-basis model. On Burgers, the hybrid RMSE is about 0.00760.0076, while the symbolic-core RMSE is about 0.2790.279; this supports an analytic backbone plus correction, but only partially supports a readable solution. The evaluation is also synthetic and mostly in-distribution: the Wave extrapolation uses a strong periodicity prior, and the Burgers transfer result is single-seed. Misspecified libraries, unfamiliar equation families, and real measurements are not tested.

Questions for authors
  1. What is the precise novelty over combining an off-the-shelf PDE-discovery backend with PINSN/PINN-SR-style staged hybrid training? Which result would remain if the hand-designed solution dictionaries were removed?

  2. Can you compare the neural corrector against a non-neural residual under the same Stage-I core, budget, and physics operator?

  3. What automatic rule would determine whether to enforce F^\hat F? Please also clarify how the derivative-estimator ensemble in Appendix C changes the recovered equation and downstream reconstruction metrics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.