Pith. sign in

REVIEW 2 major objections 10 minor 117 references

Post-hoc evaluation of discovered PDEs is multifaceted, and existing metrics only partly cover conflicting goals, so they can overstate a new physical law.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 13:32 UTC pith:226SYEFT

load-bearing objection Useful first taxonomy of PDE post-hoc metrics with honest KS pathology demos; the decoupling claim is a bit cleaner than the metrics actually are, but the survey still deserves engagement. the 2 major comments →

arxiv 2607.23753 v1 pith:226SYEFT submitted 2026-07-26 cs.LG

On the post-hoc Evaluation of PDE Discovery: A Multifaceted Challenge of Scientific Advancement

Baptiste Mathevon , Farah Cherfaoui , Amaury Habrard , Marc Sebban This is my paper
classification cs.LG
keywords PDE discoveryphysics-informed machine learningpost-hoc evaluationevaluation metrics taxonomysparsityphysical consistencyout-of-distribution generalizationKuramoto-Sivashinsky
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Discovering a partial differential equation from data is only half the scientific job; deciding whether the equation is a real governing law is harder. The paper argues that this post-hoc check must jointly weigh predictive accuracy, physical consistency, interpretability or sparsity, out-of-distribution generalization, and—when available—recovery of a known ground-truth form. Those aims often conflict, and the many metrics scattered across machine learning, numerical analysis, information theory, and symbolic regression each address only part of the problem. The authors therefore assemble what they present as the first taxonomy of such metrics, spell out each metric’s strengths and failure modes on a chaotic Kuramoto–Sivashinsky running example, and offer practical recommendations aimed at more standardized, less over-interpreted evaluation. Readers who design discovery algorithms or who apply them to real systems get a map of what current scores actually certify—and what they do not.

Core claim

The central claim is that reliable post-hoc evaluation of a discovered PDE cannot rest on any single family of scores: accuracy, sparsity, physical consistency, long-horizon or out-of-distribution behavior, and optional ground-truth recovery must be considered together, and today’s metrics leave large gaps that invite overstated claims of new physical theory. The authors support this by organizing the literature into a taxonomy keyed to five evaluation questions, documenting pathologies on deliberately close and distant Kuramoto–Sivashinsky variants, and deriving concrete guidelines plus open research directions.

What carries the argument

A taxonomy of PDE evaluation metrics organized around five questions (solution accuracy, physical consistency, sparsity/interpretability, out-of-distribution generalization, and ground-truth recovery), summarized in a capacity table and stress-tested on a Kuramoto–Sivashinsky running example that separates a near-correct equation from a spurious one.

Load-bearing premise

That metrics can be treated as algorithm-agnostic measures of the equation itself, and that lessons from a constructed Kuramoto–Sivashinsky pair plus literature synthesis extend to a general evaluation method when the true law is unknown.

What would settle it

Apply the recommended multi-metric protocol (term/coefficient recovery when known, solution and rollout errors, sparsity–accuracy trade-offs, physical-property checks, and out-of-distribution initial conditions and long horizons) to several held-out systems with known ground truth; if near-correct and clearly wrong equations still receive overlapping or reversed rankings, or if practitioners following the guidelines still publish overstated “new law” claims, the taxonomy’s claim to standardize reliable evaluation fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Algorithm papers should report more than coefficient or term recovery: at minimum solution/rollout errors and out-of-distribution checks when claiming a recovered law.
  • Users validating a new scientific equation should treat small residuals or high sparsity as insufficient without physical-consistency and long-horizon tests.
  • Standard practice should separate “recover a known PDE” benchmarks from true discovery settings where ground truth is unavailable.
  • Future metrics should target functional structure (e.g., Sobolev-type or transport-based comparisons), numerical solvability, and fractional PDEs, which classical scores mishandle.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Reviewers of PDE-discovery work could require an explicit mapping of reported scores onto the five questions, making overclaim easier to spot.
  • The same taxonomy could grade human-proposed or simplified equations, not only machine-learned ones, turning evaluation into a shared scientific checklist.
  • Emphasis on rollout and changing initial conditions suggests that surrogate generators (neural operators) will become part of evaluation pipelines when real out-of-distribution data are scarce.
  • If solvability and fractional-order metrics become standard, discovery algorithms may be forced to prefer equations that are both sparse and cheap to integrate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 10 minor

Summary. This is a survey of post-hoc evaluation metrics for PDE discovery. The authors organize the problem around five questions (Q1 solution accuracy, Q2 physical consistency, Q3 sparsity/interpretability, Q4 out-of-distribution generalization, Q5 ground-truth recovery), and compile a taxonomy of metrics drawn from sparse regression, information theory, symbolic regression, and numerical analysis: support/coefficient recovery scores (TPR, precision/recall, coefficient errors, sensitivity-weighted and Lyapunov-weighted variants, NDCG, Tanimoto), prediction errors (MSE/MAE on u and u_t, spectral errors, rollout error), sparsity and accuracy–sparsity trade-offs (rewards, AIC/BIC/PIC, MDL variants), physical-consistency checks, and OOD protocols. Each family gets an explicit pros/cons analysis, and a running Kuramoto–Sivashinsky example (PDE_A close to GT, PDE_B with spurious terms) illustrates pathologies: small short-horizon errors and small residuals can mislead, rollout errors expose long-horizon divergence, and AIC/BIC are swamped by likelihood at large n. Table 1 maps metrics to Q1–Q5; §5 gives guidelines (recover vs. discover, interpolation vs. extrapolation, prior availability) and perspectives (Sobolev-seminorm metrics, Wasserstein/gradient-flow evaluation, numerical solvability, fractional PDEs). A software repository (PDE-Evaluation) is announced.

Significance. If the "first taxonomy" claim holds (it is appropriately hedged), this fills a real gap: PDE-discovery papers evaluate on a case-by-case basis, and over-interpretation of recovery scores is common. Strengths worth naming: the survey is not a bare enumeration — each metric carries an explicit pros/cons analysis; the KS running example is a genuinely useful didactic device that reproducibly demonstrates specific failure modes (e.g., Table 6's AIC/BIC likelihood-swamping at n≈10^6; Table 7's rollout-gap collapse at T=300; the zero-residual pitfall in Fig. 4); the promised PDE-Evaluation repository would make the taxonomy directly usable; and the guidelines in §5 (interpolate AND extrapolate; condition on prior availability) are actionable. The paper's conclusions are modest and mostly supported by its own experiments. Its contribution is organizational and cautionary rather than a new method, which is appropriate for a survey.

major comments (2)
  1. [§1, Table 1, §4.2, Table 7] §1 ('By decoupling the identified PDE from the discovery process, we shift the focus... to the scientific discovery itself') together with Table 1 frames the metrics as intrinsic properties of the equation object. But many surveyed criteria are not functions of the equation alone: S_terms (Eq. 25) depends explicitly on |Θ| (the authors note this in §4.2); C(α̂ᵀΘ) (Eq. 26) is representation-dependent; all prediction/rollout/OOD criteria (Eqs. 18–24, 37) presuppose a solver, discretization, IC/BC, and validation data — the running example fixes ETD1, δt=0.05, Nx=1024, L=22, so those numbers partly encode solver and regime, not symbolic error; and Table 7 shows Conv_x/Conv_t = 0 for both PDE_A and PDE_B, i.e., they cannot separate equations the paper argues are dramatically different (the text concedes this). The paper acknowledges each dependency piecemeal, but the central framing and the
  2. [§3.2, Eqs. (14)–(15), Table 3] The definition of NDCG in Eqs. (14)–(15) deviates from the standard DCG normalization in the cited reference [37] (Järvelin & Kekäläinen): (i) the rank value itself is used as the gain, whereas standard DCG uses graded relevance (e.g., 2^{rel}−1); with rank-as-gain, a less important term contributes more gain, which inverts the intended weighting unless the ordering convention is carefully defined; (ii) the denominator is DCG(rank(α)) rather than the ideal DCG, so the ratio is not bounded by 1 — a wrong permutation placing large rank values early yields NDCG > 1, contradicting the surrounding text ('The closer NDCG is to 1, the better'). Since this is a survey whose value rests on accurate metric definitions, please either restate the metric as actually used in [36] (with a citation-precise formula), or justify this variant and its range, and re-verify the Table 3 values (0.9862, 0.8855)
minor comments (10)
  1. [Table 3 paragraph (Running Example, §3.2)] The justification for omitting w^Lyap ε²_coef is overstated: estimating the largest Lyapunov exponent of the KS equation at L=22 is routine (Benettin-type algorithms) and not 'exponentially expensive'; the NP-hardness claim cites [43], which concerns meta-complexity, not Lyapunov exponents. Please soften or properly support this claim.
  2. [§4.3, Eq. (28)] Reward1 is stated as ∈ R_{>0}, but the R²-like factor can be negative when the residual error exceeds the variance of u_t; the codomain should be R.
  3. [§4.3, Eq. (27)] Score (Eq. 27) is written with nMAE(û_A, û_B), suggesting an error between the two estimated solutions rather than each against validation data u; presumably a log-ratio of errors against u is intended (as in [54]). Please clarify.
  4. [§3.2, Eq. (13)] Eq. (13): the summation index runs to |θ| rather than |Θ|, and unlike Eq. (12) there is no normalization by Σw_i — please state whether this is intentional.
  5. [§4.3 Running Example, Table 6] The text states the experiment 'comprises 1,064,000 data points', while §2 gives n_V = Nx × Nt = 1,024,000; please reconcile. Relatedly, Table 6 reports identical AIC_c and BIC values (−1.1E+07, −5.4E+06) for both PDEs; since the penalty difference is invisible at this precision, either report more significant digits or note explicitly that the displayed degeneracy is the point being made.
  6. [§4.1 Running Example, Table 4] §4.1 Running Example: the text gives PDE_A's rollout error as '0.08170' but Table 4 reports 0.8170 (the stated 20× ratio confirms the latter). Also, fMSE 22011.1 vs 1081.7 is a factor ≈20, not 'one order of magnitude'.
  7. [Abstract, §1] The 'to our knowledge, first taxonomy' claim should be explicitly differentiated from MDBENCH [12] and Ducos et al. [36] ('Evaluating PDE discovery methods...'), which also survey/benchmark evaluation methodology; both are cited but their overlap with the present contribution is not discussed.
  8. [§1 (contributions)] The PDE-Evaluation repository is announced but no URL is given; please add it, and indicate which of the Table 1 metrics are implemented.
  9. [Throughout] Typos/grammar: 'parcimony' (§1) → parsimony; 'Kuramoto–Sivanshinsky' (§5); 'Kuramoto-Sivashinsk' (Fig. 7 caption); 'comnbine' (§5, discussion of [104]); 'the the first two heatmaps' and 'the the largest errors' (§4.1); 'These metric can therefore also be considered' (§3.1); 'Wasserstein GANNs' (§5).
  10. [§3.1, Table 2] Table 2's LLM-as-judge results are a single-prompt anecdote; since the text already flags LLM evaluation as unreliable, consider noting model/version/prompt in the repository for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: taxonomy and didactic KS demo, not a self-justifying derivation of a physical law.

full rationale

This paper is a survey/taxonomy of existing post-hoc PDE evaluation metrics drawn from ML, numerical analysis, information theory, and symbolic regression. It does not fit a governing law to data and then declare that law discovered, nor does it present fitted parameters as independent predictions. The five key questions (Q1–Q5), Table 1 coverage map, and §3–§4 metric catalogue are organizational syntheses of external literature objects (TPR/F1-style support recovery, coefficient errors, MSE/rollout, AIC/BIC/MDL, physical-property checks, OOD protocols). The Kuramoto–Sivashinsky running example deliberately constructs PDE_A (near-GT) and PDE_B (farther, with spurious terms) under a fixed dictionary and ETD1 numerics to illustrate metric behavior and pathologies; GT is used transparently as a didactic reference, not as a hidden training target for a claimed new theory. Mild self-positioning (“to our knowledge, the first taxonomy”) and author recommendations in §5 do not force the taxonomy by definition or reduce central claims to self-citation chains. No self-definitional loop, fitted-input-as-prediction, uniqueness-from-authors, or ansatz-smuggling step is present. Circularity burden is nil.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

Load-bearing background is standard sparse-PDE-discovery framing plus the authors' methodological stance that evaluation should be post-hoc and multi-criteria. No new physical entities. Free choices are experimental (KS setup, hand-built competitor PDEs, hyperparameter values copied from cited metric papers) and affect illustrations more than the existence of the taxonomy.

free parameters (4)
  • Hand-chosen PDE_A / PDE_B coefficients and spurious terms = e.g. α_A≈(-0.99,-0.98,-0.985); α_B includes 0.0004u and -0.0008 u_x u_xx
    Competitor equations are constructed so A is near GT and B is farther; metric rankings partly reflect this design choice rather than blind algorithm outputs.
  • KS numerical discretization (δt, Nx, L, T, ETD1) = T=50, δt=0.05, Nx=1024, L=22, ETD1
    All Sec. 4 scores depend on this fixed solver setup; different schemes could change absolute errors and rollout curves.
  • Trade-off hyperparameters in Reward/MDL metrics = c0=0.2; ξ1=0.01; ξ2=0.0001; various MDL λ/ε
    Table 6 uses c0, ξ1, ξ2, λ, εd, cmax from original papers; rankings can shift with these knobs.
  • Dictionary Θ size and contents = Θ=(u, uu_x, u_xx, u_xxxx, u_x u_xx)
    S_terms and support metrics depend on the five-term library chosen for the example.
axioms (5)
  • domain assumption Most physical laws are sparse combinations of a few terms, so parsimony is a meaningful evaluation axis.
    Stated throughout §1–§2 and §4.2; motivates sparsity and accuracy–sparsity metrics.
  • ad hoc to paper Post-hoc metrics should evaluate the equation object independently of the discovery algorithm (noise/scarcity robustness of learners excluded).
    Explicit methodological decoupling in §1; shapes the whole taxonomy scope.
  • domain assumption When GT is available, recovery metrics (support/coefficients) are appropriate for algorithm proof-of-concept; when GT is unknown, data-based prediction, physics, and OOD checks are required for scientific discovery claims.
    Core of §3 vs §4 and §5 'Discover or Recover'.
  • standard math Standard definitions of MSE/MAE, information criteria, TPR/precision/recall, Wasserstein/OT background, and classical PDE well-posedness/numerics hold as used.
    Imported formulae in §§3–4 and perspective §5.
  • domain assumption Small residuals alone do not certify the true governing law if GT is unused or unavailable.
    §4.4 and Fig. 4 residual pitfall; central to caution against over-interpretation.
invented entities (2)
  • Five-question evaluation frame (Q1–Q5) and associated metric taxonomy independent evidence
    purpose: Organize scattered metrics and map each to accuracy, physics, sparsity, OOD, or GT recovery.
    Conceptual scaffold of the paper (Fig. 1, Table 1); not a physical entity, but the main invented organizing structure.
  • PDE-Evaluation software repository no independent evidence
    purpose: Let users compute surveyed metrics on their own PDEs.
    Announced in §1; external artifact claim without commit hash in text.

pith-pipeline@v1.2.0-grok45-kimik3 · 41445 in / 3658 out tokens · 74311 ms · 2026-07-30T13:32:37.949647+00:00 · methodology

0 comments
read the original abstract

Partial differential equation (PDE) discovery aims to identify from data the governing law of a physical system. Constituting a cornerstone of scientific advancement, it has become during the past decade a major line of research in the rapidly evolving field of Physics-informed Machine Learning (PiML). Among the remaining open problems to address in this domain, the post-hoc evaluation of discovered PDEs raises the particular difficulty of being multifaceted. Indeed, it requires jointly considering predictive accuracy, physical consistency, interpretability, and out-of-distribution generalization capacity. Given that some of these properties are conflicting, it is worth noting that the wide range of existing evaluation metrics only partially address the overall problem, potentially leading to overly interpreted conclusions about the validity of a presumed new physical theory. From an abundant literature spanning machine learning, numerical analysis, information theory or symbolic regression, we propose, to our knowledge, the first taxonomy of PDE evaluation metrics, and discuss their advantages and limitations in depth. Based on the observation that evaluation is often achieved on a case-by-case basis and that a universally accepted methodology remains elusive, we further provide recommendations with the aim of promoting standardized and reliable practices, before sketching promising future lines of research in this field. We argue that this paper is intended both for ML experts who design new PDE discovery algorithms and for users of these methods aiming, in real applications, to discover and validate well-founded scientific laws.

Figures

Figures reproduced from arXiv: 2607.23753 by Amaury Habrard, Baptiste Mathevon, Farah Cherfaoui, Marc Sebban.

Figure 1
Figure 1. Figure 1: Overview of PDE discovery and metrics (in blue) devoted to address the multifaceted [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulated solutions of the Kuramoto-Sivashinsky equation [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pointwise prediction absolute error between ˆu [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Possible pitfall of over-interpreting small residuals: 3D solutions of the Kuramoto [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Subset of Scanning Electron Microscopy (SEM) images of surfaces irradiated by a femto [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulation data obtained from the GT KS equation [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Long-horizon simulations (T ∈ [0, 300]) of the Kuramoto-Sivashinsk GT P DEGT (left), P DEA (center) and P DEB (right). The rollout error ϵ t rollout(ˆu, u) over time (dashed-lines) is also reported. overstatements regarding the validity of a discovered PDE, we formulate in this section practical guidelines based on the findings presented in this survey. We finish the paper by proposing some promising resea… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

117 extracted references · 11 linked inside Pith

  1. [1]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 2019

    2019

  2. [2]

    Stuart, and Anima Anandkumar

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhat- tacharya, Andrew M. Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InICLR 2021, Austria, May 3-7, 2021. OpenReview.net, 2021

    2021

  3. [3]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6), 5 2021

    2021

  4. [4]

    Brunton, Joshua L

    Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the National Academy of Science, 113(15):3932–3937, April 2016

    2016

  5. [5]

    PDE-net: Learning PDEs from data

    Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. PDE-net: Learning PDEs from data. In Jennifer Dy and Andreas Krause, editors,Proceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pages 3208–3216. PMLR, 10–15 Jul 2018

    2018

  6. [6]

    A robust sindy approach by combining neural networks and an integral form.arXiv, abs/2309.07193, 2023

    Ali Forootani, Pawan Goyal, and Peter Benner. A robust sindy approach by combining neural networks and an integral form.arXiv, abs/2309.07193, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Pde-learn: Using deep learning to discover partial differential equations from noisy, limited data.Neural Networks, 174:106242, 2024

    Robert Stephany and Christopher Earls. Pde-learn: Using deep learning to discover partial differential equations from noisy, limited data.Neural Networks, 174:106242, 2024

    2024

  8. [8]

    Unrolled-sindy: A stable explicit method for non linear pde discovery from sparsely sampled data

    Fayad Ali Banna, Antoine Caradot, Eduardo Brandao, Jean-Philippe Colombier, R´ emi Emonet, and Marc Sebban. Unrolled-sindy: A stable explicit method for non linear pde discovery from sparsely sampled data. InECML 2026, 2026

    2026

  9. [9]

    PhysPDE: Rethinking PDE discovery and a physical HYpothesis selection benchmark

    Mingquan Feng, Yixin Huang, Yizhou Liu, Bofang Jiang, and Junchi Yan. PhysPDE: Rethinking PDE discovery and a physical HYpothesis selection benchmark. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  10. [10]

    Nathan Kutz, and Steven L

    Kadierdan Kaheman, J. Nathan Kutz, and Steven L. Brunton. Sindy-pi: a robust algorithm for parallel implicit sparse identification of nonlinear dynamics.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 476(2242), October 2020

    2020

  11. [11]

    Discovery of partial differential equations from highly noisy and sparse data with physics-informed information criterion.Research, 6, January 2023

    Hao Xu, Junsheng Zeng, and Dongxiao Zhang. Discovery of partial differential equations from highly noisy and sparse data with physics-informed information criterion.Research, 6, January 2023. 34

    2023

  12. [12]

    Mdbench: Bench- marking data-driven methods for model discovery

    Amirmohammad Ziaei Bideh, Aleksandra Georgievska, and Jonathan Gryak. Mdbench: Bench- marking data-driven methods for model discovery. InProc. Conf. AAAI Artif. Intell., 2026

    2026

  13. [13]

    Ap- proximation Error of Sobolev Regular Functions with tanh Neural Networks: Theoretical Impact on PINNs

    Benjamin Girault, R´ emi Emonet, Amaury Habrard, Jordan Patracone, and Marc Sebban. Ap- proximation Error of Sobolev Regular Functions with tanh Neural Networks: Theoretical Impact on PINNs. ECML 2024, March 2024

    2024

  14. [14]

    Error estimates for physics-informed neural networks approximating the navier–stokes equations.IMA Journal of Numerical Analysis, 44(1):83–119, 02 2024

    Tim De Ryck, Ameya D Jagtap, and Siddhartha Mishra. Error estimates for physics-informed neural networks approximating the navier–stokes equations.IMA Journal of Numerical Analysis, 44(1):83–119, 02 2024

    2024

  15. [15]

    On the convergence of PINNs.Bernoulli, 31(3):2127 – 2151, 2025

    Nathan Doum` eche, G´ erard Biau, and Claire Boyer. On the convergence of PINNs.Bernoulli, 31(3):2127 – 2151, 2025

    2025

  16. [16]

    Miettinen.Nonlinear multiobjective optimization

    K. Miettinen.Nonlinear multiobjective optimization. Kluwer, Boston, USA, 1999

    1999

  17. [17]

    Ro- drigues

    Beatriz Regina Brum, Luiza Lober, Isolde Terezinha Santos Previdelli, and Francisco A. Ro- drigues. Discovering equations from data: symbolic regression in dynamical systems.Journal of Physics: Complexity, 7, 2025

    2025

  18. [18]

    Data-driven discovery of governing differential equations across physical systems.arXiv, abs/2606.09638, 2026

    Siyu Lou, Hao Xu, Wenguan Wang, Lu Lu, Hao Sun, Yang Liu, Linfeng Zhang, Dongxiao Zhang, and Yuntian Chen. Data-driven discovery of governing differential equations across physical systems.arXiv, abs/2606.09638, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Lagergren, John T

    John H. Lagergren, John T. Nardini, G. Michael Lavigne, Erica M. Rutter, and Kevin B. Flores. Learning partial differential equations for biological transport models from noisy spatio-temporal data.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 476(2234):20190800, February 2020

    2020

  20. [20]

    Mechanistic PDE networks for dis- covery of governing equations

    Adeel Pervez, Efstratios Gavves, and Francesco Locatello. Mechanistic PDE networks for dis- covery of governing equations. InForty-second International Conference on Machine Learning, 2025

    2025

  21. [21]

    Data-driven discovery of PDEs via the adjoint method.Transactions on Machine Learning Research, 2025

    Mohsen Sadr, Tony Tohme, and Kamal Youcef-Toumi. Data-driven discovery of PDEs via the adjoint method.Transactions on Machine Learning Research, 2025

    2025

  22. [22]

    On the stability of feature selection algorithms.Journal of Machine Learning Research, 18(174):1–54, 2018

    Sarah Nogueira, Konstantinos Sechidis, and Gavin Brown. On the stability of feature selection algorithms.Journal of Machine Learning Research, 18(174):1–54, 2018

    2018

  23. [23]

    Learning sparse nonlinear dynamics via mixed-integer optimization.Nonlinear Dynamics, 7(11), 2022

    Dimitris Bertsimas and Wes Gurnee. Learning sparse nonlinear dynamics via mixed-integer optimization.Nonlinear Dynamics, 7(11), 2022

    2022

  24. [24]

    Tanimoto random features for scalable molecular machine learning

    Austin Tripp, Sergio Bacallado, Sukriti Singh, and Jos´ e Miguel Hern´ andez-Lobato. Tanimoto random features for scalable molecular machine learning. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  25. [25]

    Dynamic bayesian optimization framework for instruction tuning in partial differential equation discovery

    Junqi Qu, Yan Zhang, Shangqian Gao, and Shibo Li. Dynamic bayesian optimization framework for instruction tuning in partial differential equation discovery. InSubmitted to ACL Rolling Review - January 2026, 2026. under review

    2026

  26. [26]

    Data-driven discovery of dynamical systems in pharmacology using large language models

    Samuel Holt, Zhaozhi Qian, Tennison Liu, James Weatherall, and Mihaela van der Schaar. Data-driven discovery of dynamical systems in pharmacology using large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 96325–96366. Curran...

    2024

  27. [27]

    Foundation models for equation discovery in high energy physics

    Manuel Morales-Alvarado. Foundation models for equation discovery in high energy physics. arXiv, 2510.03397, 10 2025

    arXiv 2025

  28. [28]

    Parshin Shojaee, Ngoc-Hieu Nguyen, Kazem Meidani, Amir Barati Farimani, Khoa D Doan, and Chandan K. Reddy. LLM-SRBench: A new benchmark for scientific equation discovery with large language models. InForty-second International Conference on Machine Learning, 2025

    2025

  29. [29]

    Saveliev, Samuel Holt, Nabeel Seedat, David L

    Evgeny S. Saveliev, Samuel Holt, Nabeel Seedat, David L. Bentley, Jim Weatherall, and Mihaela van der Schaar. Influence-guided symbolic regression: Scientific discovery via llm-driven equation search with granular feedback.arXiv, abs/2605.29184, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    Messenger and David M

    Daniel A. Messenger and David M. Bortz. Weak sindy for partial differential equations.Journal of Computational Physics, 443:110525, October 2021

    2021

  31. [31]

    Fasel, J

    U. Fasel, J. N. Kutz, B. W. Brunton, and S. L. Brunton. Ensemble-sindy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 478(2260), April 2022

    2022

  32. [32]

    Callaghan, and Dongxiao Zhang

    Hao Xu, Yuntian Chen, Rui Cao, Tianning Tang, Mengge Du, Jian Li, Adrian H. Callaghan, and Dongxiao Zhang. Generative discovery of partial differential equations by learning from math handbooks.Nat Commun., 16(1), 2025

    2025

  33. [33]

    Discovering symbolic partial differ- ential equation by abductive learning

    En-Hao Gao, Cunjing Ge, Yuan Jiang, and Zhi-Hua Zhou. Discovering symbolic partial differ- ential equation by abductive learning. InAdvances in Neural Information Processing Systems 39 (NeurIPS’25), 2025

    2025

  34. [34]

    Learning gradient flow: Using equation discovery to accelerate engineering optimization.Computer Methods in Applied Mechanics and Engineering, 460:119099, 2026

    Grant Norman, Conor Rowan, Kurt Maute, and Alireza Doostan. Learning gradient flow: Using equation discovery to accelerate engineering optimization.Computer Methods in Applied Mechanics and Engineering, 460:119099, 2026

    2026

  35. [35]

    What are lyapunov exponents, and why are they interesting?Bulletinof the American Mathematical Society, 54(1):79–105, 2017

    Amie Wilkinson. What are lyapunov exponents, and why are they interesting?Bulletinof the American Mathematical Society, 54(1):79–105, 2017

    2017

  36. [36]

    Evaluating PDE discovery methods for multiscale modeling of biological signals

    Andr´ ea Ducos, Audrey Denizot, Thomas Guyet, and Hugues Berry. Evaluating PDE discovery methods for multiscale modeling of biological signals. InSpringer LNBI, volume 15959, Lyon, France, September 2025. Fran¸ cois Fages and Sabine P´ er` es

    2025

  37. [37]

    J¨ arvelin and J

    K. J¨ arvelin and J. Kek¨ al¨ ainen. Cumulated gain-based evaluation of IR techniques.ACM Trans- actions on Information Systems (TOIS), 20(4):422–446, 2002

    2002

  38. [38]

    On uncertainty-penalized bayesian information criterion.arXiv, abs/2404.16881, 2024

    Pongpisit Thanasutives and Ken ichi Fukui. On uncertainty-penalized bayesian information criterion.arXiv, abs/2404.16881, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Robert Stephany and Christopher J. Earls. PDE-READ: human-readable partial differential equation discovery using deep learning.Neural Networks, 154:360–382, 2022

    2022

  40. [40]

    Robust pde discovery under sparse and highly noisy conditions via attention neural networks.ArXiv, abs/2506.17908, 2025

    Shilin Zhang, Yunqing Huang, Nianyu Yi, and shihan Zhang. Robust pde discovery under sparse and highly noisy conditions via attention neural networks.ArXiv, abs/2506.17908, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Chem-r: Learning to reason as a chemist.arXiv, abs/2510.16880, 2025

    Weida Wang, Benteng Chen, Di Zhang, Wanhao Liu, Shuchen Pu, Ben Gao, Jin Zeng, Xiaoyong Wei, Tianshu Yu, Shuzhou Sun, Tianfan Fu, Wanli Ouyang, Lei Bai, Jiatong Li, Zifu Wang, Yuqiang Li, and Shufei Zhang. Chem-r: Learning to reason as a chemist.arXiv, abs/2510.16880, 2025

    arXiv 2025

  42. [42]

    Discovery of interpretable t c descriptors in conventional superconductors guided by symbolic regression.Physical Review Materials, 10, 05 2026

    Fang Lim, Jinbo Pan, and Shixuan Du. Discovery of interpretable t c descriptors in conventional superconductors guided by symbolic regression.Physical Review Materials, 10, 05 2026. 36

    2026

  43. [43]

    Np-hardness of approximating meta-complexity: A cryptographic approach.SIAM Journal on Computing, 54(4):819–886, 2025

    Yizhi Huang, Rahul Ilango, and Hanlin Ren. Np-hardness of approximating meta-complexity: A cryptographic approach.SIAM Journal on Computing, 54(4):819–886, 2025

    2025

  44. [44]

    Discovering symbolic differential equations with symmetry invariants.Transactions on Machine Learning Research, 2026

    Jianke Yang, Manu Bhat, Bryan Hu, Yadi Cao, Nima Dehmamy, Robin Walters, and Rose Yu. Discovering symbolic differential equations with symmetry invariants.Transactions on Machine Learning Research, 2026. Featured Certification, J2C Certification

    2026

  45. [45]

    Pdebench: An extensive benchmark for scientific machine learning

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pfl¨ uger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 1596–1611. Curr...

    2022

  46. [46]

    Physics-informed temporal align- ment for auto-regressive PDE foundation models

    Congcong Zhu, Xiaoyan Xu, Jiayue Han, and Jingrun Chen. Physics-informed temporal align- ment for auto-regressive PDE foundation models. InForty-second International Conference on Machine Learning, 2025

    2025

  47. [47]

    Physics-informed learning of governing equations from scarce data.Nature Communications, 12(1), October 2021

    Zhao Chen, Yang Liu, and Hao Sun. Physics-informed learning of governing equations from scarce data.Nature Communications, 12(1), October 2021

    2021

  48. [48]

    Vadgama, and Rose Yu

    Jianwei Yang, Ohm Rishabh Venkatachalam, Mohammad Kianezhad, Sharvaree P. Vadgama, and Rose Yu. Think like a scientist: Physics-guided llm agent for equation discovery.ArXiv, abs/2602.12259, 2026

    arXiv 2026

  49. [49]

    Veeling, Paris Perdikaris, Richard E Turner, and Johannes Brandstet- ter

    Phillip Lippe, Bastiaan S. Veeling, Paris Perdikaris, Richard E Turner, and Johannes Brandstet- ter. PDE-refiner: Achieving accurate long rollouts with neural PDE solvers. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  50. [50]

    Diffusionrollout: Uncertainty- aware rollout planning in long-horizon PDE solving.Transactions on Machine Learning Re- search, 2026

    Seungwoo Yoo, Juil Koo, Daehyeon Choi, and Minhyuk Sung. Diffusionrollout: Uncertainty- aware rollout planning in long-horizon PDE solving.Transactions on Machine Learning Re- search, 2026

    2026

  51. [51]

    Stable long-horizon pde forecasting via latent structured spectral propagators.arXiv, abs/2605.10154, 2026

    Xiaoxiao Lu, Ye Yuan, and Jiahao Shi. Stable long-horizon pde forecasting via latent structured spectral propagators.arXiv, abs/2605.10154, 2026

    Pith/arXiv arXiv 2026

  52. [52]

    Feynman.The Character of Physical Law

    Richard P. Feynman.The Character of Physical Law. MIT Press, Cambridge, MA, 1965

    1965

  53. [53]

    README: Rapid equation discovery with multimodel encoders

    Gregory Kang Ruey Lau, Yue Ran Kang, Zi-Yu Khoo, Apivich Hemachandra, Ruth Wan Theng Chew, and Bryan Kian Hsiang Low. README: Rapid equation discovery with multimodel encoders. InNeurIPS 2025 AI for Science Workshop, 2025

    2025

  54. [54]

    Discovering symbolic models from deep learning with inductive bi- ases

    Miles Cranmer, Alvaro Sanchez Gonzalez, Peter Battaglia, Rui Xu, Kyle Cranmer, David Spergel, and Shirley Ho. Discovering symbolic models from deep learning with inductive bi- ases. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 17429–17442. Curran Associates, ...

    2020

  55. [55]

    Zhen Zhang, Zongren Zou, Ellen Kuhl, and George Em Karniadakis. Discovering a reac- tion–diffusion model for alzheimer’s disease by combining pinns with symbolic regression.Com- puter Methods in Applied Mechanics and Engineering, 419:116647, 2024

    2024

  56. [56]

    Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Phys

    Mengge Du, Yuntian Chen, and Dongxiao Zhang. Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Phys. Rev. Res., 6:013182, Feb 2024. 37

    2024

  57. [57]

    Akaike.Information Theory and an Extension of the Maximum Likelihood Principle, pages 267–281

    H. Akaike.Information Theory and an Extension of the Maximum Likelihood Principle, pages 267–281. Akademiai Kiad´ o, Budapest, 1973

    1973

  58. [58]

    Estimating the dimension of a model.The Annals of Statistics, 6(2):461–464, March 1978

    Gideon Schwarz. Estimating the dimension of a model.The Annals of Statistics, 6(2):461–464, March 1978

    1978

  59. [59]

    N. M. Mangan, J. N. Kutz, S. L. Brunton, and J. L. Proctor. Model selection for dynamical systems via sparse regression and information criteria.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 473(2204):20170009, August 2017

    2017

  60. [60]

    Kolmogorov

    A.N. Kolmogorov. On tables of random numbers.Theoretical Computer Science, 207(2):387–395, 1998

    1998

  61. [61]

    Bartlett, Harry Desmond, and Pedro G

    Deaglan J. Bartlett, Harry Desmond, and Pedro G. Ferreira. Exhaustive symbolic regression. IEEE Transactions on Evolutionary Computation, 28(4):950–964, 2024

    2024

  62. [62]

    Vit´ anyi

    Paul M.B. Vit´ anyi. How incomputable is kolmogorov complexity?Entropy, 22(4):408, April 2020

    2020

  63. [63]

    Symbolic regression via MDLformer-guided search: from minimizing prediction error to minimizing description length

    Zihan Yu, Jingtao Ding, Yong Li, and Depeng Jin. Symbolic regression via MDLformer-guided search: from minimizing prediction error to minimizing description length. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  64. [64]

    Ai feynman: A physics-inspired method for symbolic regression.Science Advances, 6(16), 2020

    Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression.Science Advances, 6(16), 2020

    2020

  65. [65]

    Symmetry-constrained language-guided program synthesis for discovering governing equations from noisy and partial observations.ArXiv, abs/2601.00088, 03 2026

    Mirza Baig and Syeda Gillani. Symmetry-constrained language-guided program synthesis for discovering governing equations from noisy and partial observations.ArXiv, abs/2601.00088, 03 2026

    arXiv 2026

  66. [66]

    On the ability of deep networks to learn symmetries from data: A neural kernel theory.J

    Andrea Perin and St´ ephane Deny. On the ability of deep networks to learn symmetries from data: A neural kernel theory.J. Mach. Learn. Res., 26:145:1–145:70, 2024

    2024

  67. [67]

    Otto, Nicholas Zolman, J

    Samuel E. Otto, Nicholas Zolman, J. Nathan Kutz, and Steven L. Brunton. A unified framework to enforce, discover, and promote symmetry in machine learning.Journal of Machine Learning Research, 26(248):1–83, 2025

    2025

  68. [68]

    Lie point symmetry and physics-informed networks

    Tara Akhound-Sadegh, Laurence Perreault-Levasseur, Johannes Brandstetter, Max Welling, and Siamak Ravanbakhsh. Lie point symmetry and physics-informed networks. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  69. [69]

    Greenberg

    Yunfei Huang and David S. Greenberg. Geometric and physical constraints synergistically en- hance neural PDE surrogates. InForty-second International Conference on Machine Learning, 2025

    2025

  70. [70]

    Lie point symmetry data aug- mentation for neural PDE solvers

    Johannes Brandstetter, Max Welling, and Daniel E Worrall. Lie point symmetry data aug- mentation for neural PDE solvers. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors,Proceedings of the 39th International Con- ference on Machine Learning, volume 162 ofProceedings of Machine Learning Research, pages...

    2022

  71. [71]

    Physics and lie symmetry informed Gaussian pro- cesses

    David Dalton, Dirk Husmeier, and Hao Gao. Physics and lie symmetry informed Gaussian pro- cesses. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Proceedings of the 41st International Con- ference on Machine Learning, volume 235 ofProceedings of Machine Learning Research...

    2024

  72. [72]

    Maddix, Shima Alizadeh, Gaurav Gupta, and Michael W

    Derek Hansen, Danielle C. Maddix, Shima Alizadeh, Gaurav Gupta, and Michael W. Mahoney. Learning physical models that can respect conservation laws.Physica D: Nonlinear Phenomena, 457:133952, January 2024

    2024

  73. [73]

    Nguyen, and Luca Daniel

    Anthony Baez, Wang Zhang, Ziwen Ma, Subhro Das, Lam M. Nguyen, and Luca Daniel. Guar- anteeing conservation laws with projection in physics-informed neural networks. InNeurIPS 2024 Workshop on Data-driven and Differentiable Simulations, Surrogates, and Solvers, 2024

    2024

  74. [74]

    Dimensional homo- geneity constrained gene expression programming for discovering governing equations.arXiv, abs/2211.09679, 2024

    Wenjun Ma, Jun Zhang, Kaikai Feng, Haoyun Xing, and Dongsheng Wen. Dimensional homo- geneity constrained gene expression programming for discovering governing equations.arXiv, abs/2211.09679, 2024

    Pith/arXiv arXiv 2024

  75. [75]

    Dimensionally-consistent equation discovery through probabilistic attribute grammars.Information Sciences, 632:742–756, 2023

    Jure Brence, Saˇ so Dˇ zeroski, and Ljupˇ co Todorovski. Dimensionally-consistent equation discovery through probabilistic attribute grammars.Information Sciences, 632:742–756, 2023

    2023

  76. [76]

    Nexus: A revolutionary framework for autonomous equation discovery

    Mehardeep Singh. Nexus: A revolutionary framework for autonomous equation discovery. Au- gust 2025

    2025

  77. [77]

    Holden and N

    H. Holden and N. Risebro.Front Tracking for Hyperbolic Conservation Laws. Springer-Verlag, New York City, NY, USA, 2002

    2002

  78. [78]

    Hairer, C

    E. Hairer, C. Lubich, and G. Wanner.Geometric numerical integration, Structure-preserving algorithms for ordinary differential equations, volume 31. Springer-Verlag Berlin, 2006

    2006

  79. [79]

    Photonic Learning in Ultrafast Laser-Induced Complexity.Ultrafast Sci- ence, March 2026

    Fayad Ali Banna, Eduardo Brand˜ ao, Anthony Nakhoul, R´ emi Emonet, Marc Sebban, and Jean- Philippe Colombier. Photonic Learning in Ultrafast Laser-Induced Complexity.Ultrafast Sci- ence, March 2026

    2026

  80. [80]

    Discovery of nonlinear dynamics with auto- mated basis function generation.arXiv, abs/2605.09696, 2026

    Mohammad Amin Basiri and Charles Nicholson. Discovery of nonlinear dynamics with auto- mated basis function generation.arXiv, abs/2605.09696, 2026

    Pith/arXiv arXiv 2026

Showing first 80 references.