arxiv: 2605.10687 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: no theorem link

The finite expression method for turbulent dynamics with high-order moment recovery

Xingjian Xu , Di Qi , Chunmei Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-12 05:33 UTC · model grok-4.3

classification 💻 cs.LG

keywords symbolic regressionturbulent dynamicsgenerative modelshigh-order momentsdata-driven modelingstochastic systemsfinite expression methodmoment recovery

0 comments

The pith

A two-stage data-driven framework recovers closed-form dynamics and accurately predicts statistical moments up to order five in turbulent systems.

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

The paper introduces a method to model turbulent dynamical systems that first uses symbolic regression to find exact mathematical expressions for the deterministic parts of the motion, including nonlinear interactions and external forces. It then applies generative models to account for the remaining random effects in the residuals. This combination is shown to recover the key terms correctly and to match statistical moments through the fifth order on test cases of stochastic triad models. If the approach holds, it offers a way to build interpretable models that capture the higher-order statistics essential for understanding turbulence without relying on predefined equation libraries.

Core claim

The framework combines the Finite Expression Method in Stage I to discover closed-form expressions of the deterministic dynamics, recovering nonlinear interaction terms and external forcing without predefined libraries, with generative models in Stage II to learn the residual stochastic components. This yields consistent estimators that quantify error in terms of data size and discretization, and numerical tests confirm accurate recovery of terms and prediction of moments up to order five across regimes in stochastic triad models.

What carries the argument

The Finite Expression Method for discovering closed-form deterministic dynamics without predefined libraries, combined with generative models for residual stochastic components.

If this is right

The method recovers nonlinear interaction terms and forcing expressions from observed data.
It predicts statistical moments up to the fifth order with accuracy verified in numerical experiments.
The symbolic estimator is consistent, with estimation error bounded by data size and numerical discretization.
Integration of interpretable symbolic discovery and data-driven stochastic modeling applies to complex turbulent systems.

Where Pith is reading between the lines

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

This hybrid approach could be extended to other systems with coupled statistics, such as fluid flows or atmospheric dynamics, where direct simulation of high moments is costly.
By keeping the deterministic part in closed form, the models remain interpretable for analysis while using generative components only for corrections.
Testing on higher-dimensional or real observational data would reveal whether the two-stage separation scales beyond the triad model examples.

Load-bearing premise

That the Stage I symbolic approximation leaves residuals whose stochastic structure can be faithfully captured by generative models without introducing bias or overfitting that would invalidate the higher-moment predictions.

What would settle it

Direct comparison of the framework's recovered expressions against the known true dynamics of a stochastic triad model, or mismatch between its predicted fifth-order moments and those computed from independent long-run simulations of the same system.

read the original abstract

Turbulent dynamical systems are characterized by nonlinear interactions and stochastic effects that generate coupled statistical quantities, such as non-zero higher-order moments, which are difficult to capture from data with accuracy. We propose a two-stage data-driven modeling framework that combines symbolic regression with generative models to jointly identify the governing dynamics and predict their key statistical quantities. In Stage I of the framework, the Finite Expression Method (FEX) is adopted to discover closed-form expressions of the deterministic dynamics, recovering nonlinear interaction terms and external forcing without predefined libraries. In Stage II, generative models are introduced to learn the residual stochastic components as a refined correction to the model error from the Stage I approximation, enabling accurate characterization of higher-order statistics. Theoretical analysis establishes the consistency of the symbolic estimator and quantifies the estimation error in terms of data size and numerical discretization. The model performance is verified through detailed numerical experiments on the stochastic triad models across multiple regimes, demonstrating that the framework successfully recovers interaction terms and forcing expressions, and accurately predicts statistical moments up to order five. These results highlight the potential of integrating interpretable symbolic discovery with data-driven stochastic modeling for complex turbulent systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper pairs library-free FEX symbolic recovery of deterministic turbulent dynamics with generative modeling of residuals to match moments up to order five, backed by consistency theory and triad-model tests.

read the letter

The paper's core idea is a two-stage pipeline for turbulent systems. Stage I applies the Finite Expression Method to extract closed-form deterministic dynamics and forcing terms without a fixed library. Stage II fits generative models to the remaining stochastic residuals so that higher-order statistical moments can be recovered accurately. The abstract reports that this recovers interaction terms on stochastic triad models and matches moments through order five, with a consistency proof for the symbolic estimator that ties error to data size and discretization steps.

Referee Report

2 major / 2 minor

Summary. The paper proposes a two-stage data-driven framework for turbulent dynamical systems. Stage I applies the Finite Expression Method (FEX) to symbolically recover closed-form deterministic dynamics, including nonlinear interactions and forcing, without a predefined library. Stage II uses generative models to learn the stochastic residuals from Stage I approximation error. Theoretical analysis claims consistency of the symbolic estimator with error bounds depending on data size and discretization. Experiments on stochastic triad models across regimes report successful term recovery and accurate prediction of statistical moments up to order five.

Significance. If the central claims hold, the integration of interpretable symbolic discovery with generative residual modeling offers a route to data-driven closure for systems whose statistics are dominated by higher-order moments. The parameter-free nature of FEX and the provision of consistency guarantees are positive features that distinguish the work from purely black-box approaches.

major comments (2)

[Abstract and §3] Abstract and §3 (theoretical analysis): the consistency result for the symbolic estimator is stated, but no derivation or bound is given for how the approximation error of the Stage II generative model propagates into the recovered cumulants of order 3–5. This propagation analysis is load-bearing for the claim that moments up to order five are accurately predicted without bias from residual modeling.
[§5] §5 (numerical experiments on triad models): the reported moment matches to order five are shown only for the joint framework; no ablation isolating the contribution of the generative residual model versus the FEX approximation alone is provided, leaving open whether the higher-moment accuracy is due to faithful residual statistics or to compensatory overfitting.

minor comments (2)

[§4] Notation for the residual process and the generative model architecture should be introduced with explicit definitions before the experiments; several symbols appear first in the figures without prior definition.
[§2] The description of the FEX optimization objective in Stage I would benefit from an explicit statement of the loss function and the regularization terms used to control expression complexity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment point by point below, indicating the revisions we will make to the manuscript.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (theoretical analysis): the consistency result for the symbolic estimator is stated, but no derivation or bound is given for how the approximation error of the Stage II generative model propagates into the recovered cumulants of order 3–5. This propagation analysis is load-bearing for the claim that moments up to order five are accurately predicted without bias from residual modeling.

Authors: We agree that a formal propagation analysis for the Stage II approximation error into higher-order cumulants is not provided in the current manuscript. The theoretical results in §3 establish consistency and error bounds specifically for the FEX symbolic estimator of the deterministic dynamics (in terms of data size and discretization), as stated in the abstract. The Stage II generative model is introduced to capture residual stochasticity and match higher-order statistics empirically, but we acknowledge the referee's point that the lack of a bound on error propagation is a gap for the full claim. In the revised version we will add a dedicated subsection (likely in §3 or §4) providing a propagation estimate. This will include a heuristic bound assuming the generative model converges in distribution to the true residual (e.g., via Wasserstein distance or moment-matching error), showing how this controls bias in cumulants up to order 5. revision: yes
Referee: [§5] §5 (numerical experiments on triad models): the reported moment matches to order five are shown only for the joint framework; no ablation isolating the contribution of the generative residual model versus the FEX approximation alone is provided, leaving open whether the higher-moment accuracy is due to faithful residual statistics or to compensatory overfitting.

Authors: The referee correctly notes that §5 reports results only for the complete two-stage framework. While the experiments demonstrate successful term recovery and accurate moment prediction up to order five across regimes, we did not include an explicit ablation isolating Stage I (FEX alone) versus the full model. This leaves open the possibility of compensatory effects. We will revise §5 to add an ablation study: for each triad regime we will report (i) moments obtained from the FEX deterministic approximation alone (with residuals set to zero or simple noise) and (ii) moments from the full FEX + generative model. This will quantify the incremental contribution of the generative residual stage and help rule out overfitting as the source of higher-moment accuracy. revision: yes

Circularity Check

0 steps flagged

No circularity: independent theoretical consistency result supports the two-stage framework

full rationale

The paper's derivation chain consists of Stage I symbolic recovery via FEX (with an explicit consistency theorem quantifying error via data size and discretization) followed by Stage II generative modeling of residuals, then empirical verification on triad models for moments up to order 5. The theoretical estimator consistency is stated as an independent analysis rather than a tautology or self-citation reduction, and the moment predictions are presented as out-of-sample verification rather than a fitted quantity renamed as a prediction. No load-bearing step reduces by construction to its own inputs; the framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No explicit free parameters, axioms, or invented entities are identifiable from the abstract. The method inherits standard assumptions of symbolic regression (e.g., existence of a sparse closed-form expression) and generative modeling (e.g., that residuals admit a learnable distribution), but these are not enumerated or justified in the provided text.

pith-pipeline@v0.9.0 · 5496 in / 1162 out tokens · 41393 ms · 2026-05-12T05:33:53.422226+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages · 5 internal anchors

[1]

Journal of Machine Learning Research , volume=

Finite expression method for solving high-dimensional partial differential equations , author=. Journal of Machine Learning Research , volume=

work page
[2]

Journal of machine learning for modeling and computing , volume=

Learning epidemiological dynamics via the finite expression method , author=. Journal of machine learning for modeling and computing , volume=. 2025 , publisher=

work page 2025
[3]

SIAM Journal on Scientific Computing , volume=

A finite expression method for solving high-dimensional committor problems , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

work page 2025
[4]

Identifying Stochastic Dynamics Via Finite Expression Methods , author=

work page
[5]

SIAM Journal on Scientific Computing , volume=

A training-free conditional diffusion model for learning stochastic dynamical systems , author=. SIAM Journal on Scientific Computing , volume=. 2025 , publisher=

work page 2025
[6]

Journal of the atmospheric sciences , volume=

An interpretation of atmospheric low-order models , author=. Journal of the atmospheric sciences , volume=

work page
[7]

Physica D: Nonlinear Phenomena , volume=

A priori tests of a stochastic mode reduction strategy , author=. Physica D: Nonlinear Phenomena , volume=. 2002 , publisher=

work page 2002
[8]

2016 , publisher=

Introduction to turbulent dynamical systems in complex systems , author=. 2016 , publisher=

work page 2016
[9]

Climate change 2007: The physical science basis

Climate models and their evaluation , author=. Climate change 2007: The physical science basis. Contribution of Working Group I to the Fourth Assessment Report of the IPCC (FAR) , pages=. 2007 , publisher=

work page 2007
[10]

SGDR: Stochastic Gradient Descent with Warm Restarts

Sgdr: Stochastic gradient descent with warm restarts , author=. arXiv preprint arXiv:1608.03983 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[11]

SymTorch: Symbolic Distillation of Neural Networks

SymTorch: A Framework for Symbolic Distillation of Deep Neural Networks , author=. arXiv preprint arXiv:2602.21307 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl

Interpretable machine learning for science with PySR and SymbolicRegression. jl , author=. arXiv preprint arXiv:2305.01582 , year=

work page internal anchor Pith review arXiv
[13]

2025 , publisher=

Review of PySR: high-performance symbolic regression in Python and Julia , author=. 2025 , publisher=

work page 2025
[14]

Virgolin and S

Symbolic regression is NP-hard , author=. arXiv preprint arXiv:2207.01018 , year=

work page arXiv
[15]

Proceedings of the national academy of sciences , volume=

Discovering governing equations from data by sparse identification of nonlinear dynamical systems , author=. Proceedings of the national academy of sciences , volume=. 2016 , publisher=

work page 2016
[16]

Journal of Computational Physics , volume=

Weak SINDy for partial differential equations , author=. Journal of Computational Physics , volume=. 2021 , publisher=

work page 2021
[17]

The Journal of chemical physics , volume=

Sparse learning of stochastic dynamical equations , author=. The Journal of chemical physics , volume=. 2018 , publisher=

work page 2018
[18]

Science advances , volume=

AI Feynman: A physics-inspired method for symbolic regression , author=. Science advances , volume=. 2020 , publisher=

work page 2020
[19]

SIAM Journal on Applied Dynamical Systems , volume=

On higher order drift and diffusion estimates for stochastic SINDy , author=. SIAM Journal on Applied Dynamical Systems , volume=. 2024 , publisher=

work page 2024
[20]

Neural Computation , volume=

Uncovering dynamical equations of stochastic decision models using data-driven SINDy algorithm , author=. Neural Computation , volume=. 2025 , publisher=

work page 2025
[21]

Journal of Machine Learning for Modeling and Computing , volume=

Modeling unknown stochastic dynamical system via autoencoder , author=. Journal of Machine Learning for Modeling and Computing , volume=. 2024 , publisher=

work page 2024
[22]

Communications of the ACM , volume=

Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

work page 2020
[23]

Computer Science Review , volume=

A comprehensive survey and analysis of generative models in machine learning , author=. Computer Science Review , volume=. 2020 , publisher=

work page 2020
[24]

arXiv preprint arXiv:2601.19740 , year=

Error estimates of a training-free diffusion model for high-dimensional sampling , author=. arXiv preprint arXiv:2601.19740 , year=

work page arXiv
[25]

Advances in neural information processing systems , volume=

On deep generative models for approximation and estimation of distributions on manifolds , author=. Advances in neural information processing systems , volume=

work page
[26]

arXiv preprint arXiv:2602.06021 , year=

Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold , author=. arXiv preprint arXiv:2602.06021 , year=

work page internal anchor Pith review arXiv
[27]

Diffusion generative model- ing for spatially resolved gene expression inference from histology images.arXiv preprint arXiv:2501.15598, 2025

Diffusion generative modeling for spatially resolved gene expression inference from histology images , author=. arXiv preprint arXiv:2501.15598 , year=

work page arXiv
[28]

ACM computing surveys , volume=

Diffusion models: A comprehensive survey of methods and applications , author=. ACM computing surveys , volume=. 2023 , publisher=

work page 2023
[29]

SIAM Review , volume=

Strategies for reduced-order models for predicting the statistical responses and uncertainty quantification in complex turbulent dynamical systems , author=. SIAM Review , volume=. 2018 , publisher=

work page 2018
[30]

Journal of the Atmospheric Sciences , volume=

Low-dimensional reduced-order models for statistical response and uncertainty quantification: Two-layer baroclinic turbulence , author=. Journal of the Atmospheric Sciences , volume=

work page
[31]

2017 , school=

Strategies for Reduced-Order Models in Uncertainty Quantification of Complex Turbulent Dynamical Systems , author=. 2017 , school=

work page 2017
[32]

arXiv preprint arXiv:2507.08972 , year=

Simulating three-dimensional turbulence with physics-informed neural networks , author=. arXiv preprint arXiv:2507.08972 , year=

work page arXiv
[33]

Chen and E

Scale-adaptive generative flows for multiscale scientific data , author=. arXiv preprint arXiv:2509.02971 , year=

work page arXiv
[34]

Nathan Kutz and Steven L

Kathleen Champion and Bethany Lusch and J. Nathan Kutz and Steven L. Brunton , title =. Proceedings of the National Academy of Sciences , volume =

work page
[35]

arXiv preprint arXiv:2506.20607 , year=

H-FEX: A Symbolic Learning Method for Hamiltonian Systems , author=. arXiv preprint arXiv:2506.20607 , year=

work page arXiv
[36]

Journal of Computational Physics , pages=

Solving high-dimensional partial integral differential equations: The finite expression method , author=. Journal of Computational Physics , pages=. 2025 , publisher=

work page 2025
[37]

arXiv preprint arXiv:2508.06834 , year=

A Score-based Diffusion Model Approach for Adaptive Learning of Stochastic Partial Differential Equation Solutions , author=. arXiv preprint arXiv:2508.06834 , year=

work page arXiv
[38]

Journal of Computational Physics , pages=

Generative AI models for learning flow maps of stochastic dynamical systems in bounded domains , author=. Journal of Computational Physics , pages=. 2025 , publisher=

work page 2025
[39]

Variational methods for machine learning with applications to deep networks , pages=

Variational autoencoder , author=. Variational methods for machine learning with applications to deep networks , pages=. 2021 , publisher=

work page 2021
[40]

A Bradford Book , year=

Reinforcement learning: An introduction , author=. A Bradford Book , year=

work page
[41]

arXiv preprint arXiv:1912.04871 (2019)

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

work page arXiv 1912
[42]

Advances in neural information processing systems , volume=

Generative adversarial nets , author=. Advances in neural information processing systems , volume=

work page
[43]

arXiv:1905.09883 , year=

Neural stochastic differential equations: Deep latent gaussian models in the diffusion limit , author=. arXiv preprint arXiv:1905.09883 , year=

work page arXiv 1905
[44]

Auto-Encoding Variational Bayes

Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[45]

Stationary density estimation of It

Gu, Yiqi and Harlim, John and Liang, Senwei and Yang, Haizhao , journal=. Stationary density estimation of It. 2023 , publisher=

work page 2023
[46]

Physical Review X , volume=

Learning force fields from stochastic trajectories , author=. Physical Review X , volume=. 2020 , publisher=

work page 2020
[47]

Journal of Machine Learning Research , volume=

Rademacher and Gaussian Complexities: Risk Bounds and Structural Results , author=. Journal of Machine Learning Research , volume=

work page
[48]

2022 , publisher=

Lecture notes for machine learning theory (CS229M/STATS214) , author=. 2022 , publisher=

work page 2022
[49]

2006 , publisher=

Nonlinear dynamics and statistical theories for basic geophysical flows , author=. 2006 , publisher=

work page 2006
[50]

2017 , publisher=

Atmospheric and oceanic fluid dynamics , author=. 2017 , publisher=

work page 2017
[51]

1983 , publisher=

Introduction to plasma theory , author=. 1983 , publisher=

work page 1983
[52]

Tellus , volume=

The predictability of a flow which possesses many scales of motion , author=. Tellus , volume=

work page
[53]

Proceedings of the National Academy of Sciences , volume=

Quantifying uncertainty in climate change science through empirical information theory , author=. Proceedings of the National Academy of Sciences , volume=

work page
[54]

JFM , volume=

Reynolds averaged turbulence modelling using deep neural networks with embedded invariance , author=. JFM , volume=

work page
[55]

Annual Review of Fluid Mechanics , volume=

Turbulence modeling in the age of data , author=. Annual Review of Fluid Mechanics , volume=

work page
[56]

Measurement Science and Technology , volume=

Turbulent flows , author=. Measurement Science and Technology , volume=

work page
[57]

IEEE Engineering in Medicine and Biology Magazine , volume=

Multiscale modeling of the primary visual cortex , author=. IEEE Engineering in Medicine and Biology Magazine , volume=. 2009 , publisher=

work page 2009