Pith. sign in

REVIEW 3 major objections 5 minor 54 references

LLM-guided PDE discovery works when noise-robust derivatives and residual filters give the language model a clean fitness signal.

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

2026-07-14 10:54 UTC pith:NYWCVHTS

load-bearing objection Solid integration of LLM skeletons with spline/SWR evaluation that beats the usual baselines on noise and anti-memorization PDEs; “structural recovery” sometimes means a good effective surrogate, and the paper is mostly honest about that. the 3 major comments →

arxiv 2607.10546 v1 pith:NYWCVHTS submitted 2026-07-12 cs.LG

LLM-PDESR: Robust PDE Discovery via Subdomain Weighted Residuals and LLM-Guided Symbolic Hypothesis Generation

Jinyang Du , Hao Ma , Xiaohu Shi , Bo Yang , Yanchun Liang , Heow Pueh Lee , Chunguo Wu This is my paper
classification cs.LG
keywords PDE discoverysymbolic regressionlarge language modelssubdomain weighted residualsquintic B-splinesnoise robustnessPareto multi-objective optimizationscientific machine learning
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 the partial differential equations that govern a physical system from noisy measurements is hard: the combinatorial space of candidate formulas is huge, and standard finite-difference residuals amplify high-frequency noise so badly that search algorithms chase mathematical artifacts instead of true laws. This paper claims that pairing a large language model as a structural hypothesis generator with a continuous evaluation environment—C4 quintic spline derivatives plus subdomain weighted residuals that act as natural low-pass filters—restores a faithful fitness landscape. A Pareto loop then feeds the best accurate-yet-simple candidates back to the model so it can refine them. On 23 canonical one-dimensional PDEs the method recovers the correct skeletons far more reliably under noise than prior symbolic-regression and sparse-regression baselines; it also recovers exact or core terms on five novel equations built to block memorization, and it extracts a reusable structural skeleton for a one-dimensional atmospheric-wind surrogate from real ERA5 reanalysis data that transfers across geography and season once coefficients are re-fit.

Core claim

The central claim is that LLM-PDESR—an end-to-end loop that lets a language model propose executable equation skeletons while evaluating them with quintic-spline differentiation and subdomain weighted residuals, then selecting via a Pareto accuracy-complexity front—recovers true governing PDE structures from noisy data more reliably than existing methods and yields a transferable 1D atmospheric surrogate directly from reanalysis observations.

What carries the argument

Subdomain weighted residual (SWR) evaluation on C4 quintic B-splines: the residual of a candidate PDE is integrated against compact smooth test functions over local windows, producing a low-pass-filtered loss whose BFGS optimization supplies the LLM with non-deceptive fitness scores and a Pareto front that prunes bloat.

Load-bearing premise

That the spline-plus-SWR fitness landscape is faithful enough for the language-model proposals to match the true governing operators rather than merely effective residual-minimizing surrogates, especially for nested high-order terms and the multi-parameter atmospheric skeleton.

What would settle it

On the Forced Quintic Swift-Hohenberg or Traffic-Flow novel systems, or under 20 percent noise on the high-order benchmark suite, show that the recovered skeleton either systematically omits essential nonlinear or nested terms while still achieving low SWR loss, or fails to transfer when the structural skeleton is frozen and only coefficients are re-optimized on a held-out spatiotemporal domain.

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

If this is right

  • Noise-robust continuous evaluation lets language-model search scale to free-form PDEs without hand-crafted dictionaries.
  • Pareto feedback systematically suppresses equation bloat that plagues genetic and neural symbolic regression.
  • The same skeleton can be re-parameterized for new geographic or seasonal regimes once the structure is frozen.
  • Novel out-of-distribution equations become a practical test that distinguishes genuine discovery from corpus memorization.

Where Pith is reading between the lines

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

  • If the SWR low-pass property is the dominant reason for success, the same evaluation core could be swapped under non-LLM generators such as genetic programming or reinforcement-learning controllers.
  • The partial recoveries on topography-constrained chemotaxis and forced Swift-Hohenberg suggest the method may systematically prefer effective homogenized operators over exact micro-scale forcings.
  • Extending the quintic-spline/SWR pipeline to unstructured 2-D/3-D grids would be the natural next stress test for real multi-physics discovery.

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

3 major / 5 minor

Summary. The manuscript proposes LLM-PDESR, a closed-loop PDE discovery framework that couples LLM-generated symbolic skeletons with a noise-robust numerical evaluator: C^4 quintic B-spline spatial derivatives, subdomain weighted residual (SWR) fitness as a low-pass filter, BFGS coefficient identification, and a Pareto front with a 10 imes MSE_min parsimony gate for feedback and final selection. It reports near-perfect structural recovery on 23 canonical 1D PDEs under 0–20% additive noise, exact or partial recovery on five synthetically designed novel systems (including a coupled predator–prey PDE), and extraction of a 14-parameter 1D coupled atmospheric circulation skeleton (1D-CACE) from ERA5 winds with spatiotemporal OOD coefficient re-optimization. Ablations isolate spline, SWR, and Pareto contributions; Appendix B supplies error bounds and variance-reduction arguments for the evaluator.

Significance. If the claims hold, the work is a meaningful advance for scientific machine learning: it pairs free-form LLM hypothesis generation with a mathematically motivated continuous evaluator that addresses the well-known finite-difference noise amplification problem, and it supplies both a memorization-resistant novel-PDE suite and a real reanalysis case study. Strengths include multi-baseline comparisons (Fig. 2, Table 1), component ablations (Fig. 4), transparent failure on the traffic-shock system (App. E.5), algebraic-equivalence discussion for Michaelis–Menten recovery, and theoretical support for spline/SWR error and variance reduction (App. B). These elements make the contribution more than an LLM wrapper and of clear interest to the PDE-discovery and SciML communities.

major comments (3)
  1. [§4.3, Table 1, App. E.2–E.4] §4.3, Table 1, and App. E.2/E.4: The central claim of “true discovery” / “structural recovery” on the five novel PDEs is only partially secured by the SWR+Pareto landscape. Topography-Constrained Chemotaxis recovers an effective logistic + linear advection surrogate that entirely misses the nested (u(1−u)cos(x))_x term; Forced Quintic Swift–Hohenberg recovers u_xxxx but replaces −u³+u⁵ and the multiplicative Gaussian pump by a free cubic plus additive Gaussian. Both partial skeletons achieve low SWR-MSE and pass the 10×MSE_min gate, yet are not the true operators. The paper labels these “≈” and discusses them in the appendix, but the abstract and §1 still frame the suite as validating genuine symbolic discovery. Please either (i) tighten claims to distinguish exact skeleton recovery from low-MSE phenomenological surrogates, or (ii) add controlled experiments showing when/whether the true
  2. [§4.4, Eq. (5), Table 2] §4.4, Eq. (5), Table 2: OOD validation of 1D-CACE freezes the 14-term skeleton and re-optimizes all coefficients θ1…θ14 in each regime (R² ≈ 58–76%). The paper itself interprets xv/xu as first-order spatial closures rather than fundamental physics. Without controls—e.g., random or dictionary-generated multi-term skeletons of comparable complexity re-fit the same way, or coefficient transfer with only a subset free—it is hard to separate “invariant dynamical mechanisms” from flexibility of a rich ansatz. A short control table or null-model comparison is needed to support the OOD generalization claim as stated.
  3. [Algorithm 1, §3.3] Algorithm 1 (lines 12–13) and §3.3: Final selection uses the fixed heuristic MSE(F) ≤ 10 × MSE_min then min complexity. This gate is load-bearing for “avoidance of equation bloat,” yet no sensitivity to the factor 10 (or alternatives such as knee-point / AIC-style criteria) is reported. Given that partial novel recoveries and the 14-parameter 1D-CACE both survive this gate, a brief sensitivity or alternative-selection analysis would strengthen the claim that the recovered skeletons are preferred for structural reasons rather than gate choice.
minor comments (5)
  1. [Figure 2] Fig. 2 axis labels and some appendix figures appear as Unicode glyph dumps in the preprint PDF; ensure vector fonts render correctly in the camera-ready version.
  2. [Appendix D] Appendix D: DeepSeek-V3 is accessed via a closed API with a temperature schedule “every 30,000 tokens.” For reproducibility, state model snapshot/date if available, and report variance over multiple independent runs (seeds) on at least the novel suite.
  3. [§3.2, Algorithm 2] K = 10 subdomains and ω = 2 L_domain/K are fixed throughout; a one-paragraph sensitivity note (or reference to App. B scaling) would help readers port the method.
  4. [§2.2] Related work: briefly position against concurrent LLM-SR / LLM-guided discovery lines (e.g., LLM-SR, FunSearch-style program search) beyond EqGPT, so the Pareto-feedback + continuous evaluator distinction is sharper.
  5. [§3, Eq. (1)] Eq. (1) and multivariate prompts: notation for D > 1 is introduced early but most of the numerical pipeline is written for scalar u; a short remark that SWR is applied componentwise (or jointly) would clarify the predator–prey and 1D-CACE cases.

Circularity Check

0 steps flagged

No significant circularity: structural recovery is scored against independent synthetic ground truth and external ERA5 fields; OOD transfer freezes the discovered skeleton and only re-optimizes coefficients.

full rationale

LLM-PDESR is an empirical discovery pipeline, not a first-principles derivation that could collapse into its inputs. Candidate skeletons are proposed by an LLM, then scored by an independent numerical evaluator (C4 quintic B-splines + subdomain weighted residuals + BFGS; Eqs. 2–4, Alg. 2, App. B). Success is defined by exact or partial match to held-out ground-truth operators on 23 canonical and five author-constructed novel PDEs (Table 1, App. E), plus OOD R² after freezing structure and re-fitting only numeric coefficients on ERA5 (Table 2, Eq. 5). The 10×MSE_min Pareto gate (Alg. 1) is a selection heuristic, not a definition of the target operator. Partial recoveries (Topography Chemotaxis effective advection; Forced Swift-Hohenberg missing quintic) are reported as partial and analyzed as phenomenological surrogates, not re-labeled as exact recovery. Appendix B error bounds are classical approximation/spectral arguments, not self-citations of uniqueness theorems. No load-bearing step reduces a claimed prediction to a fitted input or to an author-only uniqueness claim by construction. LLM pretraining exposure to canonical PDEs is a contamination risk mitigated by the novel suite, not circularity of the derivation chain.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on standard approximation theory for quintic B-splines and weak-form residual filtering, plus several hand-chosen numerical and selection hyperparameters that define the fitness landscape the LLM sees. The 1D-CACE object is an invented reduced-order skeleton whose coefficients are free per regime. No new physical particles or forces are postulated; the load-bearing modeling choices are the evaluation operators and the parsimony gate.

free parameters (7)
  • Number of SWR subdomains K
    Fixed at K=10 in Algorithm 2; controls window width and noise averaging strength of the fitness signal.
  • SWR window width ω = 2 L_domain/K
    Hand-set relative to domain length; directly sets the low-pass cutoff of the residual filter (App. B.2).
  • Parsimony gate: MSE ≤ 10 × MSE_min
    Order-of-magnitude tolerance in Algorithm 1 that selects the final skeleton among Pareto elites; changes which equation is declared recovered.
  • LLM sampling temperature schedule and population size (4 candidates/iteration, top-3 feedback)
    Controls exploration/exploitation of the symbolic search; App. D.
  • BFGS initialization θ0 = 1 and 30s evaluation timeout
    Affects which skeletons converge to finite MSE versus being discarded as invalid.
  • 1D-CACE coefficients θ1…θ14
    Fourteen free parameters re-optimized per spatiotemporal regime after the structural skeleton is frozen; load-bearing for OOD R² claims.
  • Noise model γ ∈ {0.05,0.10,0.20} and additive Gaussian corruption
    Defines the robustness benchmark; not fitted but chosen as the experimental noise protocol.
axioms (5)
  • standard math Quintic B-spline smoothing of noisy spatial fields yields derivative error bounds superior in practice to finite differences for n≤4 (Theorem 1 / App. B.1).
    Uses classical Schoenberg approximation theory plus a regularized noise term; standard but the regularization details are implementation-dependent.
  • standard math Integrating residuals against compact sin^4 test functions attenuates white-noise variance by ~O(1/N_ω) relative to pointwise MSE (Theorem 2).
    Standard weak-form / WSINDy-style argument; underpins the claim that SWR fixes the fitness landscape.
  • domain assumption Occam’s razor via AST-node complexity on a Pareto front yields the true physical skeleton when MSE is within one order of magnitude of the best fit.
    Selection rule in §3.3 / Algorithm 1; not proved, only motivated.
  • domain assumption LLM pretraining plus constrained prompts can propose physically plausible free-form PDE skeletons without requiring a fixed dictionary.
    Core hypothesis-generation premise (§3.1); success depends on model and prompt constraints.
  • ad hoc to paper Freezing a multi-term 1D skeleton and re-fitting coefficients across regions tests invariant dynamics rather than mere flexibility of the ansatz.
    Used to interpret 1D-CACE OOD results (§4.4, App. F); a modeling judgment, not a theorem.
invented entities (2)
  • 1D-CACE structural skeleton (coupled ut, vt with 14 parametric terms including xv, xu closures) no independent evidence
    purpose: Interpretable reduced-order surrogate for meridionally averaged ERA5 10 m winds under severe 3D→1D truncation.
    Not a known textbook PDE; discovered and named in this paper. Independent evidence is limited to OOD R² after coefficient re-fit, not external first-principles derivation.
  • Five structurally novel benchmark PDEs (topography chemotaxis, Michaelis–Menten morphogenesis, forced quintic Swift–Hohenberg, traffic bottleneck, cross-chemotactic predator–prey) no independent evidence
    purpose: Anti-memorization test suite for free-form discovery methods.
    Synthesized for this paper; useful as benchmarks but not independently established physical laws.

pith-pipeline@v1.1.0-grok45 · 38034 in / 3999 out tokens · 52030 ms · 2026-07-14T10:54:48.191053+00:00 · methodology

0 comments
read the original abstract

Discovering governing partial differential equations (PDEs) from noisy observational data is a fundamental challenge in scientific machine learning. Traditional symbolic regression (SR) methods often struggle to identify accurate equations within vast combinatorial search spaces, largely due to their inability to incorporate essential domain-specific prior knowledge. Furthermore, reliance on pointwise evaluations and discrete finite differences inherently amplifies high-frequency noise, creating deceptive fitness landscapes that derail the optimization process. To resolve these bottlenecks, we propose LLM-PDESR, a framework that integrates the structural hypothesis generation of Large Language Models (LLMs) with a mathematically rigorous evaluation environment. By employing C^4-continuous quintic splines for robust differentiation and subdomain weighted residuals as natural low-pass filters, our approach effectively mitigates the fitness landscape distortion that plagues existing methods. A Pareto-driven feedback loop then enables the LLM to iteratively refine candidate equations, balancing predictive accuracy with structural parsimony. We evaluate LLM-PDESR on 23 canonical PDEs and five structurally novel equations (including a multivariate system) specifically designed to preclude dataset memorization and test true discovery capabilities. Demonstrating real-world applicability, the framework successfully extracts a consistent structural skeleton for an interpretable 1D dynamical surrogate (1D-CACE) directly from noisy ERA5 reanalysis data. Extensive experiments and out-of-distribution testing confirm that LLM-PDESR significantly outperforms state-of-the-art methodologies in structural recovery, noise resilience, and the avoidance of spurious complexity and equation bloat.

Figures

Figures reproduced from arXiv: 2607.10546 by Bo Yang, Chunguo Wu, Hao Ma, Heow Pueh Lee, Jinyang Du, Xiaohu Shi, Yanchun Liang.

Figure 1
Figure 1. Figure 1: The LLM-PDESR architecture. The automated pipeline integrates three core modules: (1) LLM-Driven Hypothesis Generation, (2) Robust Parameter Identification via splines, SWR, and BFGS optimization, and (3) Pareto-Driven Evaluation providing dynamic feedback. Algorithm 1 LLM-PDESR: The End-to-End Discovery Framework Require: Observations D, Initial prompt C, Max generations G Ensure: Optimal PDE skeleton N ∗… view at source ↗
Figure 2
Figure 2. Figure 2: Noise robustness on benchmark PDEs. Impact of varying observational noise on (a) structural success rate and (b) relative coefficient error. We first evaluate the Structural Success Rate (Figure 2a). LLM-PDESR achieves a 100% success rate at 0% noise and maintains consistently high robustness as noise increases (5%–20%). Most baselines (e.g., SGA-PDE, DISCOVER, WSINDy_PDE) suffer severe degradation, while … view at source ↗
Figure 3
Figure 3. Figure 3: Hemispheric OOD predictions of 1D-CACE. The model robustly reconstructs South Indian Ocean dynamics (austral winter) despite severe geographical shifts. Ours (Full) w/o Spline w/o SWR w/o Pareto 0 20 40 60 80 100 Recovery Rate (%) 100.0% 69.6% 60.9% 30.4% (a) Structural Success Rate ↑ Ours (Full) w/o Spline w/o SWR w/o Pareto 10−13 10−11 10−9 10−7 10−5 10−3 10−1 Mean Squared Error (MSE) (b) MSE of Recovere… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of the LLM-PDESR framework. Impact of removing core components on (a) structural success rate and (b) MSE distributions under noise-free conditions. integral residuals with pointwise MSE), and (3) w/o Pareto (removing the multi-objective selection, retaining only MSE as the fitness metric) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Topography-Constrained Chemotaxis. LLM-PDESR accurately approximates macro￾scopic wave propagation and logistic saturation, providing a robust phenomenological surrogate without requiring the explicit cos(x) micro-scale forcing. E.3 Morphogenesis with Michaelis-Menten kinetics Ground Truth Equation: ut = 0.1uxx − 1.0 u 0.5 + u + exp(−x 2 ) (27) Physical Motivation: This system captures the biochemical conc… view at source ↗
Figure 6
Figure 6. Figure 6: Morphogenesis. LLM-PDESR achieves perfect structural recovery and cross-section alignment. The discovered equation algebraically mirrors the true Michaelis-Menten fractional kinetics and accurately identifies the explicit spatial Gaussian injection. E.4 Forced quintic Swift-Hohenberg Ground Truth Equation: ut = 1.5 exp(−0.1x 2 )u − 2uxx − uxxxx − u 3 + u 5 (29) Physical Motivation & Dynamical Justification… view at source ↗
Figure 7
Figure 7. Figure 7: Forced Quintic Swift-Hohenberg. The discovered skeleton successfully captures the fundamental 4 th-order pattern-forming instability and correct wavelength selection, despite missing the explicit quintic term. E.5 Traffic flow with spatial bottleneck Ground Truth Equation: ut = −(u(1 − u)(1 − 0.5 tanh(x)))x + 0.05(ux/u)x (31) Physical Motivation: A macroscopic traffic model incorporating a physical road na… view at source ↗
Figure 8
Figure 8. Figure 8: Traffic Flow. A transparent example of algorithmic failure. The numerical stiffness caused by the tanh(x) bottleneck prevents the framework from learning the correct localized flux, resulting in discovered dynamics that fail to reproduce the steep traffic shockwave. the continuous system is sampled on a uniform spatiotemporal grid (x, t) ∈ [0, 10] × [0, 10] with a resolution of 128 × 1000. Baseline Discove… view at source ↗
Figure 9
Figure 9. Figure 9: Cross-Chemotactic Predator-Prey. LLM-PDESR perfectly isolates the cross-variable spatial coupling (vux)x. The microscopic alignment in (c) visually confirms that the discovered equation faithfully forces the predator (v) to actively track the spatial gradient of the prey (u), with zero observable deviation from the ground truth. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: In-Distribution Performance (North Pacific, Jan 1–10). The 1D-CACE skeleton reconstructs the chaotic wind tendency (R2 = 76.28%), successfully balancing macroscopic wave tracking with high-frequency fluctuations. F.2.2 Spatial transferability: West Pacific (Jan 1–10) To rigorously test spatial robustness, we transfer the frozen skeleton to the West Pacific (140◦–170◦E). Dominated by the Kuroshio current’s… view at source ↗
Figure 11
Figure 11. Figure 11: Spatial OOD Performance (West Pacific, Jan 1–10). The frozen skeleton adapts robustly to distinct oceanic forcings without numerical divergence (R2 = 60.97%), confirming the structural invariance of the discovered advection terms. F.2.3 Temporal phase shift: North Pacific (Feb 1–10) Finally, we evaluate the model under Temporal Drift by shifting the window forward by one month. Due to rapid internal atmos… view at source ↗
Figure 12
Figure 12. Figure 12: Temporal Drift Performance (North Pacific, Feb 1–10). Despite severe atmospheric phase shifts, the invariant skeleton reconstructs new weather tendencies (R2 = 58.64%), proving the extraction of genuine dynamical abstractions over statistical memorization. G Limitations While LLM-PDESR demonstrates robust symbolic discovery capabilities across diverse physical systems, several limitations warrant future i… 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

54 extracted references · 5 linked inside Pith

  1. [1]

    Data-driven discovery of partial differential equations.Science advances, 3(4):e1602614, 2017

    Samuel H Rudy, Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Data-driven discovery of partial differential equations.Science advances, 3(4):e1602614, 2017

    2017

  2. [2]

    Pde-net: Learning pdes from data

    Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. Pde-net: Learning pdes from data. InInternational conference on machine learning, pages 3208–3216. PMLR, 2018

    2018

  3. [3]

    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, 378:686–707, 2019

    2019

  4. [4]

    Deepmod: Deep learning for model discovery in noisy data.Journal of Computational Physics, 428:109985, 2021

    Gert-Jan Both, Subham Choudhury, Pierre Sens, and Remy Kusters. Deepmod: Deep learning for model discovery in noisy data.Journal of Computational Physics, 428:109985, 2021

    2021

  5. [5]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    2021

  6. [6]

    Recent advances in symbolic regression.ACM Computing Surveys, 57 (11):1–37, 2025

    Junlan Dong and Jinghui Zhong. Recent advances in symbolic regression.ACM Computing Surveys, 57 (11):1–37, 2025

    2025

  7. [7]

    Transformer-based planning for symbolic regression.Advances in Neural Information Processing Systems, 36:45907–45919, 2023

    Parshin Shojaee, Kazem Meidani, Amir Barati Farimani, and Chandan Reddy. Transformer-based planning for symbolic regression.Advances in Neural Information Processing Systems, 36:45907–45919, 2023

    2023

  8. [8]

    Graph-based symbolic regression with invariance and constraint encoding

    Ziyu Xiang, Kenna Ashen, Xiaofeng Qian, and Xiaoning Qian. Graph-based symbolic regression with invariance and constraint encoding. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  9. [9]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

    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 sciences, 113(15):3932–3937, 2016

    2016

  10. [10]

    Sparse regression for large data sets with outliers

    Lea Bottmer, Christophe Croux, and Ines Wilms. Sparse regression for large data sets with outliers. European Journal of Operational Research, 297(2):782–794, 2022

    2022

  11. [11]

    Multifactorial genetic programming for symbolic regression problems.IEEE transactions on systems, man, and cybernetics: systems, 50(11): 4492–4505, 2018

    Jinghui Zhong, Liang Feng, Wentong Cai, and Yew-Soon Ong. Multifactorial genetic programming for symbolic regression problems.IEEE transactions on systems, man, and cybernetics: systems, 50(11): 4492–4505, 2018

    2018

  12. [12]

    Deep recognition of partial differential equations based on reinforcement learning and genetic algorithm.The Journal of Supercomputing, 81(5): 740, 2025

    Jinyang Du, Renyun Liu, Du Cheng, Qingliang Li, and Fanhua Yu. Deep recognition of partial differential equations based on reinforcement learning and genetic algorithm.The Journal of Supercomputing, 81(5): 740, 2025

    2025

  13. [13]

    Explainable artificial intelligence by genetic programming: A survey.IEEE Transactions on Evolutionary Computation, 27(3):621–641, 2022

    Yi Mei, Qi Chen, Andrew Lensen, Bing Xue, and Mengjie Zhang. Explainable artificial intelligence by genetic programming: A survey.IEEE Transactions on Evolutionary Computation, 27(3):621–641, 2022. 10

    2022

  14. [14]

    Interpretable scientific discovery with symbolic regression: a review

    Nour Makke and Sanjay Chawla. Interpretable scientific discovery with symbolic regression: a review. Artificial Intelligence Review, 57(1):2, 2024

    2024

  15. [15]

    Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024

    2024

  16. [16]

    Rl-gpt: Integrating reinforcement learning and code-as-policy.Advances in Neural Information Processing Systems, 37:28430–28459, 2024

    Shaoteng Liu, Haoqi Yuan, Minda Hu, Yanwei Li, Yukang Chen, Shu Liu, Zongqing Lu, and Jiaya Jia. Rl-gpt: Integrating reinforcement learning and code-as-policy.Advances in Neural Information Processing Systems, 37:28430–28459, 2024

    2024

  17. [17]

    Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

    Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

    2023

  18. [18]

    The impact of large language models on scientific discovery: a preliminary study using gpt-4.arXiv preprint arXiv:2311.07361, 2023

    Microsoft Research AI4Science and Microsoft Azure Quantum. The impact of large language models on scientific discovery: a preliminary study using gpt-4.arXiv preprint arXiv:2311.07361, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Scimon: Scientific inspiration machines optimized for novelty

    Qingyun Wang, Doug Downey, Heng Ji, and Tom Hope. Scimon: Scientific inspiration machines optimized for novelty. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 279–299, 2024

    2024

  20. [20]

    Llm-sr: Scientific equation discovery via programming with large language models.arXiv preprint arXiv:2404.18400, 2024

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K Reddy. Llm-sr: Scientific equation discovery via programming with large language models.arXiv preprint arXiv:2404.18400, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Finite-difference time-domain methods.Nature Reviews Methods Primers, 3(1):75, 2023

    FL Teixeira, C Sarris, Y Zhang, D-Y Na, J-P Berenger, Y Su, M Okoniewski, WC Chew, V Backman, and Jamesina J Simpson. Finite-difference time-domain methods.Nature Reviews Methods Primers, 3(1):75, 2023

    2023

  22. [22]

    Finite difference methods for differential equations.Draft version for use in AMath, 585(6):112, 1998

    Randall J LeVeque. Finite difference methods for differential equations.Draft version for use in AMath, 585(6):112, 1998

    1998

  23. [23]

    On the use of nonstandard finite difference methods.Journal of Difference Equations and Applications, 11(8):735–758, 2005

    Kailash C Patidar. On the use of nonstandard finite difference methods.Journal of Difference Equations and Applications, 11(8):735–758, 2005

    2005

  24. [24]

    Survey on algebraic numerical differentiation: historical developments, parametrization, examples, and applications.International Journal of Systems Science, 53(9):1848–1887, 2022

    Amine Othmane, Lothar Kiltz, and Joachim Rudolph. Survey on algebraic numerical differentiation: historical developments, parametrization, examples, and applications.International Journal of Systems Science, 53(9):1848–1887, 2022

    2022

  25. [25]

    Kadierdan Kaheman, Steven L Brunton, and J Nathan Kutz. Automatic differentiation to simultaneously identify nonlinear dynamics and extract noise probability distributions from data.Machine Learning: Science and Technology, 3(1):015031, 2022

    2022

  26. [26]

    Mohammad Prawesh Alam, Arshad Khan, and Pradip Roul. High-resolution numerical method for the time-fractional fourth-order diffusion problems via improved quintic b-spline function.Journal of Applied Mathematics and Computing, 71(1):133–171, 2025

    2025

  27. [27]

    Quintic spline interpolation with minimal feed fluctuation.J

    Kaan Erkorkmaz and Yusuf Altintas. Quintic spline interpolation with minimal feed fluctuation.J. Manuf. Sci. Eng., 127(2):339–349, 2005

    2005

  28. [28]

    Asymptotic consistency of the wsindy algorithm in the limit of continuum data.IMA Journal of Numerical Analysis, 45(6):3264–3312, 2025

    Daniel A Messenger and David M Bortz. Asymptotic consistency of the wsindy algorithm in the limit of continuum data.IMA Journal of Numerical Analysis, 45(6):3264–3312, 2025

    2025

  29. [29]

    A pareto front grid guided multi-objective evolutionary algorithm.Applied Soft Computing, 136:110095, 2023

    Ying Xu, Huan Zhang, Lei Huang, Rong Qu, and Yusuke Nojima. A pareto front grid guided multi-objective evolutionary algorithm.Applied Soft Computing, 136:110095, 2023

    2023

  30. [30]

    Pareto set learning for multi-objective reinforcement learning

    Erlong Liu, Yu-Chang Wu, Xiaobin Huang, Chengrui Gao, Ren-Jian Wang, Ke Xue, and Chao Qian. Pareto set learning for multi-objective reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18789–18797, 2025

    2025

  31. [31]

    The era5 global reanalysis.Quarterly journal of the royal meteorological society, 146(730):1999–2049, 2020

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis.Quarterly journal of the royal meteorological society, 146(730):1999–2049, 2020

    1999

  32. [32]

    Symbolic genetic algorithm for discovering open-form partial differential equations (sga-pde).Physical Review Research, 4(2):023174, 2022

    Yuntian Chen, Yingtao Luo, Qiang Liu, Hao Xu, and Dongxiao Zhang. Symbolic genetic algorithm for discovering open-form partial differential equations (sga-pde).Physical Review Research, 4(2):023174, 2022. 11

    2022

  33. [33]

    Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Physical Review Research, 6 (1):013182, 2024

    Mengge Du, Yuntian Chen, and Dongxiao Zhang. Discover: Deep identification of symbolically concise open-form partial differential equations via enhanced reinforcement learning.Physical Review Research, 6 (1):013182, 2024

    2024

  34. [34]

    Weak sindy for partial differential equations.Journal of Computational Physics, 443:110525, 2021

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

    2021

  35. [35]

    Generative discovery of partial differential equations by learning from math handbooks.Nature Communications, 16(1):10255, 2025

    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.Nature Communications, 16(1):10255, 2025

    2025

  36. [36]

    Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    Pith/arXiv arXiv 2010

  37. [37]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021

    2021

  38. [38]

    Distilling free-form natural laws from experimental data.science, 324 (5923):81–85, 2009

    Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data.science, 324 (5923):81–85, 2009

    2009

  39. [39]

    Interpretable machine learning for science with pysr and symbolicregression

    Miles Cranmer. Interpretable machine learning for science with pysr and symbolicregression. jl.arXiv preprint arXiv:2305.01582, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Partial differential equations discovery with epde framework: Application for real and synthetic data.Journal of Computational Science, 53: 101345, 2021

    Mikhail Maslyaev, Alexander Hvatov, and Anna V Kalyuzhnaya. Partial differential equations discovery with epde framework: Application for real and synthetic data.Journal of Computational Science, 53: 101345, 2021

    2021

  41. [41]

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients.arXiv preprint arXiv:1912.04871, 2019

    Brenden K Petersen, Mikel Landajuela, T Nathan Mundhenk, Claudio P Santiago, Soo K Kim, and Joanne T Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients.arXiv preprint arXiv:1912.04871, 2019

    Pith/arXiv arXiv 1912

  42. [42]

    Morl4pdes: Data-driven discovery of pdes based on multi-objective optimization and reinforcement learning.Chaos, Solitons & Fractals, 180: 114536, 2024

    Xiaoxia Zhang, Junsheng Guan, Yanjun Liu, and Guoyin Wang. Morl4pdes: Data-driven discovery of pdes based on multi-objective optimization and reinforcement learning.Chaos, Solitons & Fractals, 180: 114536, 2024

    2024

  43. [43]

    The coefficient of determination r-squared is more informative than smape, mae, mape, mse and rmse in regression analysis evaluation.Peerj computer science, 7:e623, 2021

    Davide Chicco, Matthijs J Warrens, and Giuseppe Jurman. The coefficient of determination r-squared is more informative than smape, mae, mape, mse and rmse in regression analysis evaluation.Peerj computer science, 7:e623, 2021

    2021

  44. [44]

    Hybrid cnn-transformer network with a weighted mse loss for global sea surface wind speed retrieval from gnss-r data.IEEE Transactions on Geoscience and Remote Sensing, 2025

    Xin Qiao, Qingyun Yan, and Weimin Huang. Hybrid cnn-transformer network with a weighted mse loss for global sea surface wind speed retrieval from gnss-r data.IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  45. [45]

    SIAM, 2007

    Randall J LeVeque.Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM, 2007

    2007

  46. [46]

    Springer, 2006

    Jorge Nocedal and Stephen J Wright.Numerical optimization. Springer, 2006

    2006

  47. [47]

    A structured l-bfgs method and its application to inverse problems.Inverse Problems, 40(4):045022, 2024

    Florian Mannel, Hari Om Aggrawal, and Jan Modersitzki. A structured l-bfgs method and its application to inverse problems.Inverse Problems, 40(4):045022, 2024

    2024

  48. [48]

    Era5- land: A state-of-the-art global reanalysis dataset for land applications.Earth system science data, 13(9): 4349–4383, 2021

    Joaquín Muñoz-Sabater, Emanuel Dutra, Anna Agustí-Panareda, Clément Albergel, Gabriele Arduini, Gianpaolo Balsamo, Souhail Boussetta, Margarita Choulga, Shaun Harrigan, Hans Hersbach, et al. Era5- land: A state-of-the-art global reanalysis dataset for land applications.Earth system science data, 13(9): 4349–4383, 2021

    2021

  49. [49]

    An evaluation of era5 precipitation for climate monitoring.Quarterly Journal of the Royal Meteorological Society, 148(748):3152–3165, 2022

    David A Lavers, Adrian Simmons, Freja Vamborg, and Mark J Rodwell. An evaluation of era5 precipitation for climate monitoring.Quarterly Journal of the Royal Meteorological Society, 148(748):3152–3165, 2022

    2022

  50. [50]

    The era5 global reanalysis from 1940 to 2022

    Cornel Soci, Hans Hersbach, Adrian Simmons, Paul Poli, Bill Bell, Paul Berrisford, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Raluca Radu, et al. The era5 global reanalysis from 1940 to 2022. Quarterly Journal of the Royal Meteorological Society, 150(764):4014–4048, 2024. 12 A Prompt design and numerical evaluation pipeline This section detail...

    1940

  51. [51]

    For functions, use ONLYnp.sin,np.cos,np.tanh, andnp.exp

    Spatial Derivatives & Math:Use ONLY the provided d_dx(u, order) (max order=4). For functions, use ONLYnp.sin,np.cos,np.tanh, andnp.exp

  52. [52]

    Do not use hard-coded numbers

    Numeric Coefficients:You must use the params array for all physical constants (e.g., u_t_pred = params[0] * d_dx(u, 2) ). Do not use hard-coded numbers

  53. [53]

    Output Format:Assign the formula to u_t_pred and end with return u_t_pred

  54. [54]

    • Forced Quintic Swift-Hohenberg:

    Pareto Evolution:Focus on core physical mechanisms. Evolve step-by-step to the simplest structure. Bloated or physically incorrect terms will be removed by the Pareto Frontier. Optimization Strategy:{PARETO_FRONTIER_PLACEHOLDER} 13 The specific task contexts supplied for the remaining five novel systems are detailed below: • Morphogenesis with Michaelis-M...

    2080