Pith. sign in

REVIEW 4 major objections 4 minor 5 cited by

A dynamical system can be uniquely recovered from its trajectories only when its motion is chaotic; stable systems remain fundamentally ambiguous.

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 · deepseek-v4-flash

2026-08-03 22:44 UTC pith:IC7HFYXD

load-bearing objection Solid continuous-trajectory identifiability results, oversold as finite-observation discovery; worth refereeing with a rewrite of the abstract. the 4 major comments →

arxiv 2511.08860 v2 pith:IC7HFYXD submitted 2025-11-12 math.DS cs.AIcs.LGcs.NAmath.NAnlin.CD

When is a System Discoverable from Data? Discovery Requires Chaos

Zakhar Shumaylov , Peter Zaika , Philipp Scholl , Gitta Kutyniok , Lior Horesh , Carola-Bibiane Sch\"onlieb This is my paper
classification math.DS cs.AIcs.LGcs.NAmath.NAnlin.CD MSC 37D4537C4534A55
keywords discoverabilitychaostopological transitivityidentifiabilitysets of uniquenessLorenz systemfirst integralsHausdorff dimension
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.

This paper tries to establish a foundational condition for data-driven discovery of dynamical systems: uniqueness of the governing equations from trajectory data is possible exactly when the dynamics are chaotic. The authors prove that a system that is chaotic across its whole domain is discoverable from a single trajectory in the space of continuous functions, and that a system chaotic on a strange attractor is analytically discoverable if the attractor's Hausdorff dimension exceeds d−1. They also prove the negative result that any analytic vector field with a nonconstant global analytic first integral—a conserved quantity—cannot be analytically discovered from any trajectory. As a corollary, the Lorenz system, long treated as the standard chaotic benchmark, is analytically discoverable. The upshot is that the systems most amenable to data-driven model discovery are the unstable, chaotic ones, while stable engineering systems are fundamentally underdetermined by trajectory data alone.

Core claim

The central claim is that discoverability—the existence of a unique vector field F in a function class V consistent with observed trajectories—reduces to a geometric question: does the image of a trajectory form a set of uniqueness for V? In continuous functions, a set is a set of uniqueness iff it is dense, so any topologically transitive system has a dense trajectory and is uniquely identifiable. In real-analytic functions, the relevant criterion is that the trajectory's closure contains a set of Hausdorff dimension greater than d−1, because the zero set of a nonzero analytic function has dimension at most d−1. Chaotic attractors that satisfy this dimension condition—including the Lorenz a

What carries the argument

The load-bearing object is the set of uniqueness: a subset A of R^d such that any function in the class V vanishing on A vanishes everywhere. Proposition 3.1 converts discoverability of F from trajectory u into the statement that the trajectory image O = u(U) is a set of uniqueness for V. The analytic case is carried by the dimension criterion: any set with Hausdorff dimension greater than d−1 is an analytic set of uniqueness, since the zero set of a non-zero real-analytic function has Hausdorff dimension at most d−1. Alongside this, topological transitivity, via Birkhoff's theorem, supplies the dense trajectory needed to make a chaotic attractor a set of uniqueness.

Load-bearing premise

The proofs treat 'data' as the entire continuous trajectory curve u(t) on an open time interval; finite sets of sampled points are never treated, and for analytic functions a finite set has Hausdorff dimension 0 < d−1 and cannot be a set of uniqueness.

What would settle it

A concrete falsifier: find a continuous vector field on a compact metric space that is uniquely identifiable from a single trajectory but is not topologically transitive. The paper's Theorem 4.1 asserts discoverability from one trajectory implies topological transitivity, so such an example would break the equivalence; no such example is given in the paper.

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

If this is right

  • Any model fitted to a single dense trajectory of a chaotic system is the true vector field: no other continuous function can reproduce the same trajectory.
  • The Lorenz system is analytically discoverable, so symbolic or neural discovery methods that succeed on it are not merely overfitting—a unique analytic vector field underlies the data.
  • Stable or integrable systems—those with analytic first integrals, such as energy-conserving oscillators—cannot be uniquely identified from any finite number of trajectories, so purely data-driven discovery from them is ill-posed.
  • Finite continuous discoverability forces a decomposition of the state space into finitely many closed invariant cells, each of which is topologically transitive; one trajectory per cell suffices.
  • Known physical priors, such as a conservation law, can restore uniqueness for non-chaotic systems, but only when the law is informative enough to make the candidate vector field isolated.

Where Pith is reading between the lines

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

  • If the central claim transfers to approximately chaotic or high-dimensional turbulent flows, it would give a theoretical rationale for why weather and climate models learned from data generalize: their trajectories explore enough of the state space to pin down the dynamics.
  • In practice observations are finite samples, not continuous curves; the paper's guarantee is about infinite-resolution data. A testable inference is that finite-sample discoverability should degrade with sample density for chaotic systems, while failing completely for regular ones.
  • The first-integral obstruction suggests that Hamiltonian or conservative systems, abundant in physics, are systematically hidden from purely data-driven discovery; physics-informed constraints are not optional refinements but necessary conditions for uniqueness.
  • The Hausdorff-dimension condition is sufficient, not necessary; there may be analytically discoverable attractors of dimension at most d−1, and searching for such examples would sharpen the boundary of the result.

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

4 major / 4 minor

Summary. The paper defines discoverability of a vector field F in a function class V as uniqueness of F among all V-members matching a given trajectory (Definition 2.2), and reduces that problem to whether the trajectory image is a set of uniqueness for V (Proposition 3.1). It then proves: in C^0, a single dense trajectory is necessary and sufficient for discoverability (Theorem 4.1); in the real-analytic class, a chaotic attractor that is a set of uniqueness, e.g. by Hausdorff dimension > d-1, gives discoverability from a single trajectory (Theorem 4.13); an analytic first integral obstructs analytic discoverability (Theorem 5.1); and certain conservation laws can restore uniqueness in examples (Corollaries 5.3 and 5.5). The advertised applications are the classical Lorenz system and consequences for data-driven scientific discovery.

Significance. Read as a paper about exact observation of a continuous trajectory on an open time interval, the results are substantial and mostly coherent. The C^0 characterization is clean, the dimension criterion is simple and correct, and the first-integral obstruction is a useful, rigorous negative result. The sheaf-theoretic framework is ambitious and, where completed, provides a promising local picture of non-uniqueness; the worked examples (e.g. Example 4.10) help the reader. The main advertised application, analytic discoverability of the Lorenz system, is conditional on an external dimension result whose hypotheses are not checked in the manuscript. Overall, the paper contains enough correct mathematics to warrant revision, but the central claim as stated in the abstract and introduction is not what is proved.

major comments (4)
  1. [Abstract, §2.1 (Def. 2.2), §3 (Prop. 3.1)] The abstract promises uniqueness from a finite set of observations, but the formal framework observes a differentiable trajectory u:U→R^d on an open interval U, and Proposition 3.1 reduces uniqueness to vanishing on the full image O=u(U). For any finite sample S={x_1,...,x_N}, h(u)=∏_{j=1}^N ‖u−x_j‖² is a non-zero real-analytic function vanishing on S, and the same construction works in C^0. Hence no finite set is a set of uniqueness in either class, regardless of chaos. Theorems 4.1 and 4.13 therefore do not establish the finite-observation claim, and the Section 6 discussion of weather forecasting and digital twins relies on that unproved extrapolation. The authors should either formulate a genuine finite-observation model and prove what can be said, or explicitly reframe the paper's claim as continuous full-trajectory observation.
  2. [§4.2.3, Thm 4.13] The converse in Theorem 4.13 is only sketched. It silently assumes every trajectory converges to an attractor, without stating compact trapping-region or asymptotic-approach hypotheses. The line 'As γ⊆A' appears to be a slip for γ⊆U_A (the basin of attraction), and the extension from one to finitely many trajectories is omitted; one must take the product of the individual non-zero analytic functions and observe that the product is non-zero. Please state the precise assumptions, complete the proof, or downgrade this part to a conjecture. This matters because the theorem is advertised as an 'if and only if' statement.
  3. [§4.2.3, Cor. 4.14] The Lorenz corollary is the paper's headline application, but its proof consists of two citation steps: Moreira et al. [2020] prove dim_H>2 for a class of geometric Lorenz flows, and Tucker [1999,2002] is said to imply the original Lorenz equations belong to that class. Neither the exact statement from [Moreira et al.] nor a verification that the Lorenz system satisfies its hypotheses is provided. Since the advertised result is that the classical Lorenz system is analytically discoverable, the authors should quote the relevant theorem from [Moreira et al.] and confirm that Tucker's computer-assisted proof yields the required geometric model.
  4. [§4.2.1–4.2.2, Lemma 4.5 and Thm 4.9] These results are central to the converse direction and to the non-subanalytic-locus picture, but the proofs are not at the standard of the rest of the paper. In Lemma 4.5, the equivalence 'ϕ_∗O_D coherent iff γ subanalytic' is argued with several unstated finiteness conclusions. In Theorem 4.9, the claim that every x∈Z∩A lies in N is not established: the constructed sequence (t_n) need not diverge or have no convergent subsequence, e.g. for periodic points on the attractor. Please complete these arguments or reduce the statements to what is proved.
minor comments (4)
  1. [Throughout] Several cross-references are inconsistent: 'Theorem 3.1' should be Proposition 3.1, 'Theorem 3.7' should be Proposition 3.7, 'Theorem 4.5' should be Lemma 4.5, and definitions are sometimes cited as theorems (B.5, B.7).
  2. [§2, Def. 2.5 vs Def. 2.2] Definition 2.5 defines a trajectory on [0,+∞], while Definition 2.2 considers an arbitrary open interval U. The relationship should be clarified, especially since flows may not exist for all positive time without additional hypotheses.
  3. [§5, Note 5.2] The note states that Hénon–Heiles and the double pendulum are not chaotic in position-momentum variables and are only chaotic in cross-sections. This is incorrect for the standard Hénon–Heiles system, which has chaotic orbits on the energy hypersurface in phase space. The remark is not needed for Theorem 5.1 and should be corrected or removed.
  4. [Throughout] Typos and small inaccuracies: 'Lorentz' for 'Lorenz', 'Exerciese' for 'Exercise', and the caption of Figure 1 refers to an 'integrable Lorenz model' where the intended contrast is with a non-chaotic system.

Circularity Check

0 steps flagged

No significant circularity: the derivation chain is self-contained and built on elementary reductions plus external mathematical results.

full rationale

I walked the derivation chain. Definition 2.2 formalizes discoverability from differentiable trajectories u:U→R^d on an open interval, and Proposition 3.1 reduces uniqueness to the absence of a nonzero G∈V vanishing on O=u(U). Although this reduction is cited to Scholl et al. (overlapping authors), it is a direct equivalence: if two fields fit the trajectory, their difference vanishes on O; conversely, any G vanishing on O gives a competing field F+G. The reduction is elementary and does not smuggle in the paper's conclusions. Proposition 3.5 is proved directly from density of continuous functions, and Lemma 3.10 uses the standard fact that the zero set of a nonzero real-analytic function has Hausdorff dimension ≤ d−1. Theorem 4.13 combines Birkhoff's lemma with the set-of-uniqueness criterion, and Corollary 4.14 relies on external results by Moreira et al. and Tucker, not on this paper's own claims. Theorem 5.1 constructs H=G−c0 from the first integral itself, so non-discoverability follows by direct construction rather than by definitional fiat. The conservation-law recovery results verify explicit Hessian conditions. No fitted parameter is renamed as a prediction, and no load-bearing assertion rests solely on a self-citation. The abstract's phrase 'finite set of observations' is not the formal model—the theorems concern exact continuous trajectories—and the paper itself flags the theoretical-vs-practical gap in Section 6. That is a scope limitation, not a circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 0 invented entities

The paper is pure mathematical analysis: no parameters are fitted to data. It relies on standard analytic-geometry and dynamical-systems theorems; the only externally quantified input is the Lorenz attractor dimension estimate and the identification of the Lorenz flow with a geometric Lorenz flow.

axioms (7)
  • standard math Birkhoff's theorem: topological transitivity implies existence of a dense trajectory (hypercyclic point) on complete metric spaces.
    Used throughout Section 4 to convert chaotic transitivity into a dense single trajectory; proved in Appendix A.
  • standard math Identity theorem for real analytic functions: a nonzero analytic function cannot vanish on a nonempty open set.
    Used in Theorem 3.8 and in the proof of Lemma 3.11 to extend local agreement to global agreement.
  • standard math The zero set of a nonzero real analytic function has Hausdorff dimension at most d−1.
    Used in Lemma 3.10 to characterize sets of analytic uniqueness; cites Mityagin 2020.
  • standard math Coherence of the sheaf of real analytic functions on R^d, Cartan's Theorem A, and Oka's coherence theorem.
    Used in Section 4.2.1 and Appendix B to relate trajectory subanalyticity to uniqueness sheaves.
  • standard math Existence and uniqueness of maximal flows for locally Lipschitz vector fields, and forward invariance of trapping regions.
    Used in Section 2 to set up flows, attractors, and basins of attraction.
  • domain assumption Every forward trajectory converges to some attractor and basins of attraction are the relevant non-uniqueness sets.
    Assumed without explicit hypotheses in the converse direction of Theorem 4.13; not true for arbitrary noncompact or escaping flows.
  • domain assumption The Lorenz equations define a geometric Lorenz flow, and geometric Lorenz attractors have Hausdorff dimension strictly greater than 2.
    Used in Corollary 4.14 to conclude the classical Lorenz attractor is analytically discoverable; relies on Tucker (1999, 2002) and Moreira et al. (2020) without verifying all hypotheses in the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 33120 in / 19042 out tokens · 191685 ms · 2026-08-03T22:44:46.710775+00:00 · methodology

0 comments
read the original abstract

The deep learning revolution has spurred a rise in advances of using AI in sciences. Within physical sciences the main focus has been on discovery of dynamical systems from observational data. Yet the reliability of learned surrogates and symbolic models is often undermined by the fundamental problem of non-uniqueness. The resulting models may fit the available data perfectly, but lack genuine predictive power. This raises the question: under what conditions can the systems governing equations be uniquely identified from a finite set of observations? We show, counter-intuitively, that chaos, typically associated with unpredictability, is crucial for ensuring a system is discoverable in the space of continuous or analytic functions. The prevalence of chaotic systems in benchmark datasets may have inadvertently obscured this fundamental limitation. More concretely, we show that systems chaotic on their entire domain are discoverable from a single trajectory within the space of continuous functions, and systems chaotic on a strange attractor are analytically discoverable under a geometric condition on the attractor. As a consequence, we demonstrate for the first time that the classical Lorenz system is analytically discoverable. Moreover, we establish that analytic discoverability is impossible in the presence of first integrals, common in real-world systems. These findings help explain the success of data-driven methods in inherently chaotic domains like weather forecasting, while revealing a significant challenge for engineering applications like digital twins, where stable, predictable behavior is desired. For these non-chaotic systems, we find that while trajectory data alone is insufficient, certain prior physical knowledge can help ensure discoverability. These findings warrant a critical re-evaluation of the fundamental assumptions underpinning purely data-driven discovery.

Figures

Figures reproduced from arXiv: 2511.08860 by Carola-Bibiane Sch\"onlieb, Gitta Kutyniok, Lior Horesh, Peter Zaika, Philipp Scholl, Zakhar Shumaylov.

Figure 1
Figure 1. Figure 1: Model ambiguity in non-chaotic systems versus uniqueness in chaotic systems. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of how system dynamics affect model discovery from data. The regular motion of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of sets of uniqueness and non-uniqueness for various function spaces. Columns differentiate between function spaces: Linear ([Qiu et al., 2022, Casolo et al., 2025]), Real Analytic (Theorem 3.8), and Continuous (Theorem 3.5). Being a set of uniqueness for a function space implies a function, zero on that set, is zero everywhere. 3.1 The Continuous Case In the continuous case, uniqueness becomes al… view at source ↗
Figure 4
Figure 4. Figure 4: Proof. Here we assume that we are looking at a continuous flow on a compact metric space X. Discoverable from n trajectories implies that for some {x1, x2, . . . , xn} ⊂ X, we must have by Theorems 3.1 and 3.5 that the union of the closures of their trajectories covers the space: X = [n i=1 Traj(xi). Now, if the systems is topologically transitive, then the result of [Grosse-Erdmann and Manguillot, 2011, E… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of finite continuous discoverability being equivalent to a decomposition into cells, on each of which [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the main argument behind Theorem 5.1. An analytic first integral foliates the underlying space [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Attractor Geometry Determines the Identifiability Limits of System Discovery

    cs.LG 2026-07 conditional novelty 7.0

    The smallest eigenvalue of the attractor's invariant-measure moment matrix — not the choice of algorithm — sets the identifiability ceiling for recovering governing equations from trajectory data.

  2. How Low Can You Go? Active Learning for Sparse Model Discovery in the Ultra-Low-Data Limit

    cs.LG 2026-06 unverdicted novelty 7.0

    An active learning method based on E-SINDy identifies governing ODEs and PDEs accurately with significantly fewer data samples than random sampling across tested systems.

  3. Theory and interpretability of Quantum Extreme Learning Machines: a Pauli-transfer matrix approach

    quant-ph 2026-02 unverdicted novelty 7.0

    A Pauli-transfer-matrix analysis of QELMs reveals the full set of nonlinear Pauli features generated by encoding and transformed by quantum channels, producing an interpretable classical nonlinear vector autoregressio...

  4. Symbolic recovery of PDEs from measurement data

    cs.LG 2026-02 unverdicted novelty 7.0

    Symbolic rational-function networks recover an admissible PDE from noiseless complete measurements and select the regularization-minimizing parameterization within the architecture.

  5. Flow map learning in nonlinear vector autoregressive models: influence of the feature-library structure on the training error

    cs.LG 2026-05 unverdicted novelty 6.0

    NVAR models exhibit training error scaling laws tied to feature library representation of Lie-series coefficients, with delays reducing one-step error but aiding long-horizon forecasts only under sufficient nonlinearity.

Reference graph

Works this paper leans on

168 extracted references · 8 canonical work pages · cited by 5 Pith papers · 1 internal anchor

  1. [1]

    Nature Machine Intelligence, 7 0 (1): 0 1--1, 2025

    Machine learning solutions looking for pde problems. Nature Machine Intelligence, 7 0 (1): 0 1--1, 2025. doi:10.1038/s42256-025-00989-w. URL https://doi.org/10.1038/s42256-025-00989-w

    work page doi:10.1038/s42256-025-00989-w 2025

  2. [2]

    Learning-informed parameter identification in nonlinear time-dependent pdes

    Christian Aarset, Martin Holler, and Tram Thi Ngoc Nguyen. Learning-informed parameter identification in nonlinear time-dependent pdes. Applied Mathematics & Optimization, 88 0 (3): 0 76, 2023

    2023

  3. [3]

    Identification of the coefficient in elliptic equations

    Robert Acar. Identification of the coefficient in elliptic equations. SIAM journal on control and optimization, 31 0 (5): 0 1221--1244, 1993

    1993

  4. [4]

    Fernando

    Francesca Acquistapace, Fabrizio Broglia, and Jos \'e F. Fernando. Topics in Global Real Analytic Geometry. Springer International Publishing, Cham, 2022. ISBN 978-3-030-96666-9. doi:10.1007/978-3-030-96666-9. URL https://doi.org/10.1007/978-3-030-96666-9

    work page doi:10.1007/978-3-030-96666-9 2022

  5. [5]

    An identification problem for an elliptic equation in two variables

    Giovanni Alessandrini. An identification problem for an elliptic equation in two variables. Annali di matematica pura ed applicata, 145 0 (1): 0 265--295, 1986

    1986

  6. [6]

    Three-dimensional flows, volume 1

    V \' tor Ara \'u jo, Maria Jos \'e Pacifico, and Marcelo Viana. Three-dimensional flows, volume 1. Springer, 2010

    2010

  7. [7]

    Invariant physics-informed neural networks for ordinary differential equations

    Shivam Arora, Alex Bihlo, and Francis Valiquette. Invariant physics-informed neural networks for ordinary differential equations. Journal of Machine Learning Research, 25 0 (233): 0 1--24, 2024

    2024

  8. [8]

    O ktem, and Carola-Bibiane Sch \

    Simon Arridge, Peter Maass, Ozan \"O ktem, and Carola-Bibiane Sch \"o nlieb. Solving inverse problems using data-driven models. Acta Numerica, 28: 0 1--174, 2019

    2019

  9. [9]

    Adriano Augusto and Helio J.C

    D. Adriano Augusto and Helio J.C. Barbosa. Symbolic regression via genetic programming. In Proceedings. Vol.1. Sixth Brazilian Symposium on Neural Networks, pages 173--178, 2000. doi:10.1109/SBRN.2000.889734

    arXiv 2000

  10. [10]

    Neural operators for accelerating scientific simulations and design

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandkumar. Neural operators for accelerating scientific simulations and design. Nature Reviews Physics, 6 0 (5): 0 320--328, 2024

    2024

  11. [11]

    Reproducibility crisis

    Monya Baker. Reproducibility crisis. Nature, 533 0 (26): 0 353--66, 2016

    2016

  12. [12]

    Poincar \'e and the Three Body Problem

    June Barrow-Green. Poincar \'e and the Three Body Problem . American Mathematical Society, 1997

    1997

  13. [13]

    Representation equivalent neural operators: a framework for alias-free operator learning

    Francesca Bartolucci, Emmanuel de Bezenac, Bogdan Raonic, Roberto Molinaro, Siddhartha Mishra, and Rima Alaifari. Representation equivalent neural operators: a framework for alias-free operator learning. Advances in Neural Information Processing Systems, 36: 0 69661--69672, 2023

    2023

  14. [14]

    E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E Smidt, and Boris Kozinsky. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13 0 (1): 0 2453, 2022

    2022

  15. [15]

    On structural identifiability

    Ror Bellman and Karl Johan str \"o m. On structural identifiability. Mathematical biosciences, 7 0 (3-4): 0 329--339, 1970

    1970

  16. [16]

    A survey of projection-based model reduction methods for parametric dynamical systems

    Peter Benner, Serkan Gugercin, and Karen Willcox. A survey of projection-based model reduction methods for parametric dynamical systems. SIAM Review, 57 0 (4): 0 483--531, 2015. doi:10.1137/130932715. URL https://doi.org/10.1137/130932715

    doi:10.1137/130932715 2015

  17. [17]

    Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast. arXiv preprint arXiv:2211.02556, 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Neural symbolic regression that scales

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aur \' e lien Lucchi, and Giambattista Parascandolo. Neural symbolic regression that scales. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , volume 139 of Proceedings of Machine Learning Research, pag...

    2021

  19. [19]

    Neural flows: Efficient alternative to neural odes

    Marin Bilo s , Johanna Sommer, Syama Sundar Rangapuram, Tim Januschowski, and Stephan G \"u nnemann. Neural flows: Efficient alternative to neural odes. Advances in neural information processing systems, 34: 0 21325--21337, 2021

    2021

  20. [20]

    Birkhoff

    George D. Birkhoff. Surface transformations and their dynamical applications. Acta Mathematica, 43 0 (1): 0 1--119, December 1922. ISSN 1871-2509. doi:10.1007/BF02401754. URL https://doi.org/10.1007/BF02401754

    doi:10.1007/bf02401754 1922

  21. [21]

    Topological chaos: what may this mean ?, 2008

    François Blanchard. Topological chaos: what may this mean ?, 2008. URL https://arxiv.org/abs/0805.0232

    Pith/arXiv arXiv 2008

  22. [22]

    The control of chaos: theory and applications

    Stefanos Boccaletti, Celso Grebogi, Y-C Lai, Hector Mancini, and Diego Maza. The control of chaos: theory and applications. Physics reports, 329 0 (3): 0 103--197, 2000

    2000

  23. [23]

    Automated reverse engineering of nonlinear dynamical systems

    Josh Bongard and Hod Lipson. Automated reverse engineering of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 104 0 (24): 0 9943--9948, 2007

    2007

  24. [24]

    Deepmod: Deep learning for model discovery in noisy data

    Gert-Jan Both, Subham Choudhury, Pierre Sens, and Remy Kusters. Deepmod: Deep learning for model discovery in noisy data. Journal of Computational Physics, 428: 0 109985, 2021. ISSN 0021-9991. doi:https://doi.org/10.1016/j.jcp.2020.109985. URL https://www.sciencedirect.com/science/article/pii/S0021999120307592

    arXiv 2021

  25. [25]

    Does equivariance matter at scale? In NeurIPS 2024 Workshop on Symmetry and Geometry in Neural Representations, 2025

    Johann Brehmer, S \"o nke Behrends, Pim De Haan, and Taco Cohen. Does equivariance matter at scale? In NeurIPS 2024 Workshop on Symmetry and Geometry in Neural Representations, 2025. URL https://openreview.net/forum?id=L4gb2wvVhM

    2024

  26. [26]

    Promising directions of machine learning for partial differential equations

    Steven L Brunton and J Nathan Kutz. Promising directions of machine learning for partial differential equations. Nature Computational Science, 4 0 (7): 0 483--494, 2024

    2024

  27. [27]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    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 0 (15): 0 3932--3937, 2016

    2016

  28. [28]

    Chaos as an intermittently forced linear system

    Steven L Brunton, Bingni W Brunton, Joshua L Proctor, Eurika Kaiser, and J Nathan Kutz. Chaos as an intermittently forced linear system. Nature communications, 8 0 (1): 0 19, 2017

    2017

  29. [29]

    Levine, Iñigo Urteaga, and Youssef Marzouk

    Andrey Bryutkin, Matthew E. Levine, Iñigo Urteaga, and Youssef Marzouk. Canonical bayesian linear system identification, 2025. URL https://arxiv.org/abs/2507.11535

    Pith/arXiv arXiv 2025

  30. [30]

    Symplectic neural flows for modeling and discovery, 2024

    Priscilla Canizares, Davide Murari, Carola-Bibiane Schönlieb, Ferdia Sherry, and Zakhar Shumaylov. Symplectic neural flows for modeling and discovery, 2024. URL https://arxiv.org/abs/2412.16787

    arXiv 2024

  31. [31]

    Machine learning and the physical sciences

    Giuseppe Carleo, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborov \'a . Machine learning and the physical sciences. Reviews of Modern Physics, 91 0 (4): 0 045002, 2019

    2019

  32. [32]

    Identifiability challenges in sparse linear ordinary differential equations, 2025

    Cecilia Casolo, Sören Becker, and Niki Kilbertus. Identifiability challenges in sparse linear ordinary differential equations, 2025. URL https://arxiv.org/abs/2506.09816

    Pith/arXiv arXiv 2025

  33. [33]

    Nathan Kutz, and Steven L

    Kathleen Champion, Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Data-driven discovery of coordinates and governing equations. Proceedings of the National Academy of Sciences, 116 0 (45): 0 22445--22451, 2019. doi:10.1073/pnas.1906995116. URL https://www.pnas.org/doi/abs/10.1073/pnas.1906995116

    doi:10.1073/pnas.1906995116 2019

  34. [34]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

  35. [35]

    The nonequivalence and dimension formula for attractors of lorenz-type systems

    Yuming Chen and Qigui Yang. The nonequivalence and dimension formula for attractors of lorenz-type systems. International Journal of Bifurcation and Chaos, 23 0 (12): 0 1350200, 2013

    2013

  36. [36]

    Physics-informed learning of governing equations from scarce data

    Zhao Chen, Yang Liu, and Hao Sun. Physics-informed learning of governing equations from scarce data. Nature communications, 12 0 (1): 0 6136, 2021

    2021

  37. [37]

    Chesebro, David Hofmann, Vaibhav Dixit, Earl K

    Anthony G. Chesebro, David Hofmann, Vaibhav Dixit, Earl K. Miller, Richard H. Granger, Alan Edelman, Christopher V. Rackauckas, Lilianne R. Mujica-Parodi, and Helmut H. Strey. Scientific machine learning of chaotic systems discovers governing equations for neural populations, 2025. URL https://arxiv.org/abs/2507.03631

    arXiv 2025

  38. [38]

    The double scroll family

    LEONO Chua, Motomasa Komuro, and Takashi Matsumoto. The double scroll family. IEEE transactions on circuits and systems, 33 0 (11): 0 1072--1118, 2003

    2003

  39. [39]

    Parameter and structural identifiability concepts and ambiguities: a critical review and analysis

    Claudio Cobelli and Joseph J Distefano 3rd. Parameter and structural identifiability concepts and ambiguities: a critical review and analysis. American Journal of Physiology-Regulatory, Integrative and Comparative Physiology, 239 0 (1): 0 R7--R24, 1980

    1980

  40. [40]

    Combining data and theory for derivable scientific discovery with ai-descartes

    Cristina Cornelio, Sanjeeb Dash, Vernon Austel, Tyler R Josephson, Joao Goncalves, Kenneth L Clarkson, Nimrod Megiddo, Bachir El Khadir, and Lior Horesh. Combining data and theory for derivable scientific discovery with ai-descartes. Nature Communications, 14 0 (1): 0 1777, 2023

    2023

  41. [41]

    Evolving scientific discovery by unifying data and background knowledge with ai hilbert

    Ryan Cory-Wright, Cristina Cornelio, Sanjeeb Dash, Bachir El Khadir, and Lior Horesh. Evolving scientific discovery by unifying data and background knowledge with ai hilbert. Nature Communications, 15 0 (1): 0 5922, 2024

    2024

  42. [42]

    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

  43. [43]

    Lagrangian neural networks

    Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks. arXiv preprint arXiv:2003.04630, 2020

    Pith/arXiv arXiv 2003

  44. [44]

    Learning symbolic physics with graph networks

    Miles D Cranmer, Rui Xu, Peter Battaglia, and Shirley Ho. Learning symbolic physics with graph networks. arXiv preprint arXiv:1909.05862, 2019

    Pith/arXiv arXiv 1909

  45. [45]

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics--informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92 0 (3): 0 88, 2022

    2022

  46. [46]

    Physics and lie symmetry informed gaussian processes

    David Dalton, Dirk Husmeier, and Hao Gao. Physics and lie symmetry informed gaussian processes. In Forty-first International Conference on Machine Learning, 2024

    2024

  47. [47]

    Machine learning in drug discovery: a review

    Suresh Dara, Swetha Dhamercherla, Surender Singh Jadav, CH Madhu Babu, and Mohamed Jawed Ahsan. Machine learning in drug discovery: a review. Artificial intelligence review, 55 0 (3): 0 1947--1999, 2022

    1947

  48. [48]

    Physics-informed neural networks for data-driven simulation: Advantages, limitations, and opportunities

    F \'e lix Fern \'a ndez de la Mata, Alfonso Gij \'o n, Miguel Molina-Solana, and Juan G \'o mez-Romero. Physics-informed neural networks for data-driven simulation: Advantages, limitations, and opportunities. Physica A: Statistical Mechanics and its Applications, 610: 0 128415, 2023

    2023

  49. [49]

    Magnetic control of tokamak plasmas through deep reinforcement learning

    Jonas Degrave, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de Las Casas, et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602 0 (7897): 0 414--419, 2022

    2022

  50. [50]

    On parameter and structural identifiability: Nonunique observability/reconstructibility for identifiable systems, other ambiguities, and new definitions

    J Distefano and Claudio Cobelli. On parameter and structural identifiability: Nonunique observability/reconstructibility for identifiable systems, other ambiguities, and new definitions. IEEE Transactions on Automatic Control, 25 0 (4): 0 830--833, 1980

    1980

  51. [51]

    From digital control to digital twins in medicine: A brief review and future perspectives

    Raluca Eftimie, Andreea Mavrodin, and St \'e phane PA Bordas. From digital control to digital twins in medicine: A brief review and future perspectives. Advances in Applied Mechanics, 56: 0 323--368, 2023

    2023

  52. [52]

    Ueber diffusion

    Adolf Fick. Ueber diffusion. Annalen der physik, 170 0 (1): 0 59--86, 1855

  53. [53]

    Theorie analytique de la chaleur, par M

    Joseph Fourier. Theorie analytique de la chaleur, par M. Fourier. Chez Firmin Didot, p \`e re et fils, 1822

  54. [54]

    On determining the dimension of chaotic flows

    Harold Froehling, James P Crutchfield, Doyne Farmer, Norman H Packard, and Rob Shaw. On determining the dimension of chaotic flows. Physica D: Nonlinear Phenomena, 3 0 (3): 0 605--617, 1981

    1981

  55. [55]

    Galatolo and M

    S. Galatolo and M. J. Pacifico. Lorenz like flows: exponential decay of correlations for the poincar\'e map, logarithm law, quantitative recurrence, 2009. URL https://arxiv.org/abs/0901.0574

    Pith/arXiv arXiv 2009

  56. [56]

    Plasma surrogate modelling using Fourier neural operators

    Vignesh Gopakumar, Stanislas Pamela, Lorenzo Zanisi, Zongyi Li, Ander Gray, Daniel Brennand, Nitesh Bhatia, Gregory Stathopoulos, Matt Kusner, Marc Peter Deisenroth, Anima Anandkumar, the JOREK Team, and MAST Team. Plasma surrogate modelling using Fourier neural operators. Nuclear Fusion, 64 0 (5): 0 056025, April 2024. ISSN 0029-5515. doi:10.1088/1741-43...

    doi:10.1088/1741-4326/ad313a 2024

  57. [57]

    Measuring the strangeness of strange attractors

    Peter Grassberger and Itamar Procaccia. Measuring the strangeness of strange attractors. Physica D: nonlinear phenomena, 9 0 (1-2): 0 189--208, 1983

    1983

  58. [58]

    Symbolic regression with a learned concept library

    Arya Grayeli, Atharva Sehgal, Omar Costilla Reyes, Miles Cranmer, and Swarat Chaudhuri. Symbolic regression with a learned concept library. Advances in Neural Information Processing Systems, 37: 0 44678--44709, 2024

    2024

  59. [59]

    Hamiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. Advances in neural information processing systems, 32, 2019

    2019

  60. [60]

    Linear chaos

    Karl-G Grosse-Erdmann and Alfred Peris Manguillot. Linear chaos. Springer Science & Business Media, 2011

    2011

  61. [61]

    Can physics-informed neural networks beat the finite element method? IMA Journal of Applied Mathematics, 89 0 (1): 0 143--174, 2024

    Tamara G Grossmann, Urszula Julia Komorowska, Jonas Latz, and Carola-Bibiane Sch \"o nlieb. Can physics-informed neural networks beat the finite element method? IMA Journal of Applied Mathematics, 89 0 (1): 0 143--174, 2024

    2024

  62. [62]

    Sur les probl \`e mes aux d \'e riv \'e es partielles et leur signification physique

    Jacques Hadamard. Sur les probl \`e mes aux d \'e riv \'e es partielles et leur signification physique. Princeton university bulletin, pages 49--52, 1902

    1902

  63. [63]

    Pereira, Robert J

    Ali Hasan, Jo \ a o M. Pereira, Robert J. Ravier, Sina Farsiu, and Vahid Tarokh. Learning partial differential equations from data using neural networks. ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3962--3966, 2020

    2020

  64. [64]

    Robust identifiability for symbolic recovery of differential equations, 2024

    Hillary Hauger, Philipp Scholl, and Gitta Kutyniok. Robust identifiability for symbolic recovery of differential equations, 2024. URL https://arxiv.org/abs/2410.09938

    Pith/arXiv arXiv 2024

  65. [65]

    Poseidon: Efficient foundation models for pdes

    Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger K \"a ppeli, Roberto Molinaro, Emmanuel de B \'e zenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37: 0 72525--72624, 2024

    2024

  66. [66]

    u r ein-und ausgangsgr \

    BL Ho and Rudolf E K \'a lm \'a n. Effective construction of linear state-variable models from input/output functions: Die konstruktion von linearen modeilen in der darstellung durch zustandsvariable aus den beziehungen f \"u r ein-und ausgangsgr \"o en. at-Automatisierungstechnik, 14 0 (1-12): 0 545--548, 1966

    1966

  67. [67]

    On uniqueness in structured model learning

    Martin Holler and Erion Morina. On uniqueness in structured model learning. arXiv preprint arXiv:2410.22009, 2024

    arXiv 2024

  68. [68]

    Deep generative symbolic regression

    Samuel Holt, Zhaozhi Qian, and Mihaela van der Schaar. Deep generative symbolic regression. In The Eleventh International Conference on Learning Representations, 2023

    2023

  69. [69]

    Artificial intelligence faces reproducibility crisis, 2018

    Matthew Hutson. Artificial intelligence faces reproducibility crisis, 2018

    2018

  70. [70]

    Highly accurate protein structure prediction with alphafold

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Z \' dek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596 0 (7873): 0 583--589, 2021

    2021

  71. [71]

    D- CIPHER : Discovery of closed-form partial differential equations

    Krzysztof Kacprzyk, Zhaozhi Qian, and Mihaela van der Schaar. D- CIPHER : Discovery of closed-form partial differential equations. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=jnCPN1vpSR

    2023

  72. [72]

    Sindy-pi: a robust algorithm for parallel implicit sparse identification of nonlinear dynamics

    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, 476 0 (2242): 0 20200279, 2020

    2020

  73. [73]

    The experimental multi-arm pendulum on a cart: A benchmark system for chaos, learning, and control

    Kadierdan Kaheman, Urban Fasel, Jason J Bramburger, Benjamin Strom, J Nathan Kutz, and Steven L Brunton. The experimental multi-arm pendulum on a cart: A benchmark system for chaos, learning, and control. HardwareX, 15: 0 e00465, 2023

    2023

  74. [74]

    End-to-end symbolic regression with transformers

    Pierre - Alexandre Kamienny, St \' e phane d'Ascoli, Guillaume Lample, and Fran c ois Charton. End-to-end symbolic regression with transformers. In Advances in Neural Information Processing Systems, 2022

    2022

  75. [75]

    Leakage and the reproducibility crisis in machine-learning-based science

    Sayash Kapoor and Arvind Narayanan. Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4 0 (9), 2023

    2023

  76. [76]

    Benchmarking sparse system identification with low-dimensional chaos

    Alan A Kaptanoglu, Lanyue Zhang, Zachary G Nicolaou, Urban Fasel, and Steven L Brunton. Benchmarking sparse system identification with low-dimensional chaos. Nonlinear Dynamics, 111 0 (14): 0 13143--13164, 2023 a

    2023

  77. [77]

    Kaptanoglu, Lanyue Zhang, Zachary G

    Alan A. Kaptanoglu, Lanyue Zhang, Zachary G. Nicolaou, Urban Fasel, and Steven L. Brunton. Benchmarking sparse system identification with low-dimensional chaos, 2023 b . URL https://arxiv.org/abs/2302.10787

    Pith/arXiv arXiv 2023

  78. [78]

    Machine learning in the search for new fundamental physics

    Georgia Karagiorgi, Gregor Kasieczka, Scott Kravitz, Benjamin Nachman, and David Shih. Machine learning in the search for new fundamental physics. Nature Reviews Physics, 4 0 (6): 0 399--412, 2022

    2022

  79. [79]

    Astronomia nova

    Johann Kepler . Astronomia nova. 1609

  80. [80]

    The method of proper orthogonal decomposition for dynamical characterization and order reduction of mechanical systems: an overview

    Gaetan Kerschen, Jean-claude Golinval, Alexander F Vakakis, and Lawrence A Bergman. The method of proper orthogonal decomposition for dynamical characterization and order reduction of mechanical systems: an overview. Nonlinear dynamics, 41 0 (1): 0 147--169, 2005

    2005

Showing first 80 references.