Pith. sign in

REVIEW 3 major objections 6 minor 52 references

This paper claims that neural operators can learn the unobserved factors that distinguish dynamical systems directly from heterogeneous trajectory data, and that the learned latent space mirrors the true factor space, enabling generalizatio

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

2026-07-31 23:42 UTC pith:5GIJ2OGG

load-bearing objection Solid latent-conditioned operator work, but the zero-shot discovery claim leaks factor labels and should be reframed. the 3 major comments →

arxiv 2607.23337 v1 pith:5GIJ2OGG submitted 2026-07-25 cs.LG

Neural operator discovery from heterogeneous trajectories

Zituo Chen , Qiaofeng Li , Jiaxin Hu , Sili Deng This is my paper
classification cs.LG
keywords neural operatorlatent representationdynamical systemsself-supervised learningtrajectory-decoupled samplingfactor discoverygeneralizationPDE surrogates
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.

The paper tackles a practical gap in neural operator learning: parametric surrogates require the parameters that distinguish system instances, but real archives often lack those labels. It proposes learning both a shared solution operator and a low-dimensional latent descriptor of each system instance directly from trajectory data, using only the knowledge of which trajectories come from the same instance. The central claim is that this latent descriptor, shaped by a minimal-dimension prior and by sampling conditioning and prediction trajectories from different initial conditions of the same instance, becomes an approximate one-to-one image of the true governing factors. If that holds, a single trained model can predict unseen system instances from a short trajectory (one-shot) or from a sampled point in latent space (zero-shot), and can roll out stably over long horizons. The paper demonstrates the mechanism on Burgers' equation, cylinder wake flow, FitzHugh–Nagumo reaction–diffusion, and four additional case studies.

Core claim

On its own terms, the paper's discovery is that unobserved system-specific factors can be recovered as the minimal-dimensional latent that makes prediction work, when the encoder is forced to ignore initial-condition information via trajectory-decoupled sampling. Across the reported testbeds, the recovered latent dimension equals the number of independent governing factors, the latent space is smooth and approximately diffeomorphic to factor space, and prediction generalizes beyond training regimes (including the inviscid limit for Burgers' and outside the training hull for FitzHugh–Nagumo). The learned organization is not merely a compression artifact: perturbing the latent preserves a smoo

What carries the argument

The load-bearing construction is factorized latent conditioning: a trajectory encoder maps a past-state window to a low-dimensional code c, and a conditional decoder evolves future states given c. Two constraints make c meaningful: trajectory-decoupled sampling (the encoder sees one trajectory of an instance while the decoder predicts a different trajectory of the same instance, so only cross-trajectory invariants survive) and dimension selection (the smallest d at which prediction loss saturates, chosen by a two-stage sweep that also checks for factor-aligned latent structure). The paper further shows that the decoder's supervision geometry is part of the machinery: full-field or hierarchic

Load-bearing premise

The argument stands or falls on whether grouping trajectories by system instance plus forcing a minimal-dimensional latent is enough to make the encoder recover exactly the true governing factors, rather than other cross-trajectory constants such as sensor bias or unresolved geometry.

What would settle it

Take a system family with two independent cross-trajectory invariants: one a true physical parameter and one a pure nuisance (for example, a per-instance constant offset added to every trajectory). Train the proposed method with dimension selection and inspect the learned latent. If the two-stage sweep returns dimension 2 and the latent separates the nuisance as cleanly as the physical factor, the claim that the latent aligns with underlying governing factors is falsified; if the latent provably ignores the nuisance, the claim is supported. The paper reports no such nuisance-controlled experim

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

If this is right

  • Unlabeled experimental or legacy trajectory archives can be turned into parametric-family surrogates, with unseen instances handled by one-shot inference from a short trajectory.
  • The learned latent dimension gives a data-driven estimate of the number of independent governing factors, matching the true count in all reported cases.
  • Zero-shot inference—sampling a latent rather than observing a trajectory—lets a user target a regime never seen in training, as demonstrated for FitzHugh–Nagumo outside the training hull.
  • Long-horizon rollout error is lower than with no-conditioning or autoregressive-history baselines, and physical observables such as drag and Strouhal number are recovered correctly.
  • The dependence on decoder supervision geometry implies that practical deployments should prefer full-field or hierarchical decoders when factor identification is the goal.

Where Pith is reading between the lines

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

  • A testable extension is to couple the learned latent channel with a much larger decoder: the principle prescribes how conditioning is structured, not how large the decoder is, so hybrid models might combine the generalization of latent conditioning with the capacity of large surrogates.
  • If the identifiability intuition generalizes, the same two constraints—instance grouping plus minimal dimension—could serve as a general recipe for discovering hidden factors in any repeated-observation dataset, not just PDEs.
  • The paper leaves open how far the latent–factor map extends beyond the training hull; one could probe this by fitting the diffeomorphism on a training set and measuring how prediction degrades as sampled latents move away from the training manifold.
  • A stress test with per-instance nuisance invariants (for example, a constant sensor offset added to every trajectory) would clarify whether the latent captures physical factors specifically or any cross-trajectory invariant, a distinction the paper does not settle.

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 / 6 minor

Summary. The paper proposes Neural Operator Discovery (NODnet), a latent-conditioned neural operator trained on heterogeneous trajectories grouped by system instance but without explicit factor labels. A trajectory encoder maps a short observed trajectory to a low-dimensional latent c, and a decoder predicts future states conditioned on c. Training uses trajectory-decoupled sampling (conditioning and prediction trajectories from the same system instance but different initial conditions) and a two-stage dimension-selection procedure. The method is evaluated on Burgers' equation, cylinder flow wake, FitzHugh–Nagumo reaction–diffusion, and four supplementary systems, with claims that the learned latent space captures the intrinsic dimensionality of system variation, aligns with the true governing factors, and enables both one-shot and zero-shot generalization to unseen system instances.

Significance. If the central claims hold, the paper is a valuable contribution to operator learning: it shows that a prediction objective alone, combined with instance grouping and a low-dimensional bottleneck, can produce latent representations that correlate strongly with physical factors across multiple PDE/ODE families. The experimental methodology is a genuine strength: multi-seed sweeps with reported across-seed variability, matched-capacity baselines (SNO, ANO, PNO), ablations of trajectory-decoupled sampling and decoder supervision geometry, and public code/data DOIs. The seven case studies span convolutional, recurrent, transformer, FNO, DeepONet, and Neural-ODE backbones, giving the claimed principle architectural generality. However, the headline claims of 'zero-shot extrapolation across regimes' and fully label-free 'factor discovery' are not supported by the reported evaluation protocols: the zero-shot experiment uses the true unseen factors through a supervised fitted map, and the dimension-selection stage uses knowledge of factor-aligned structure. The identifiability discussion is analogical rather than a theorem. These issues do not invalidate the core representation-learning re

major comments (3)
  1. [Sec. 4 (FitzHugh–Nagumo), Fig. 4d, Table S7, Eq. S7] The 'zero-shot' evaluation fits c = Φ(p) by least squares on the true (k,β) labels of the 9 training system instances and then evaluates by plugging the true (k,β) values of unseen OOD-Intra/OOD-Extra systems into Φ. This is supervised regression on the very factors the method claims not to use; it does not demonstrate 'zero-shot extrapolation across regimes' from a label-free representation. The paper's own definition (Fig. 1b) says zero-shot inference means sampling directly in latent space. Please either evaluate zero-shot by sampling/interpolating along the learned latent manifold without factor labels, or revise the abstract and results to state that a label-fitted latent–factor map is used to select latent coordinates for unseen factors. This concern is load-bearing because the abstract's central claim explicitly includes label-free zero-shot extrapolation.
  2. [Main text 'Dimension selection'; S2.1/S2.2/S2.3 Stage 2] The final latent dimension d* is selected using a Stage 2 criterion that inspects whether the d+1 model produces additional 'factor-aligned structure' in PCA of per-instance latents. Concretely, Burgers' S2.1 rejects d=2 because its second axis is 'spurious' (using knowledge that viscosity is the only factor), and FN-RD S2.3 accepts d=2 because the d=3 encoder 'spontaneously collapses to a 2-D submanifold' (using knowledge that two factors are expected). Thus the claim that d* is recovered 'without access to labeled governing factors' and matches the intrinsic dimensionality is not fully supported: factor-aligned structure and factor count enter the model-selection loop. This does not invalidate the trained encoder/decoder, but the paper should either provide a fully unsupervised dimension-selection criterion or explicitly disclose this supervision leak in the main text.
  3. [S3.4] The text asserts: 'The identifiability claim NODnet supports is therefore: ... the recovered latent is unique up to permutation, sign, and component-wise reparametrization, with the dimension of the recovered representation matching the dimension of the governing factor space.' No theorem or proof is given, and the cited results (Khemakhem et al.; Locatello et al.) require assumptions — exponential-family conditionals, non-degenerate distributions, shared-factor pair sampling — that the text itself acknowledges are not satisfied ('the formal subtlety is that this conditional is degenerate'). The main-text Discussion also states that 'a formal theoretical characterization of these properties remains open.' Please either supply a theorem with explicit assumptions and a proof, or demote the S3.4 paragraph to a conjecture/analogy and align it with the stated limitation.
minor comments (6)
  1. [Abstract / Fig. 1b] The phrase 'zero-shot extrapolation across regimes' overstates the evidence given the label-fitted protocol in Table S7; please harmonize the abstract and figure caption with the actual evaluation.
  2. [Eq. (4) and Methods] The factorized realization is written as x̂_{t+1} = gθ(x_t; c), but the Burgers' decoder is a full-field DeepONet predicting û(t,x) from u0 and c, not an autoregressive next-step map. Please either generalize the notation to cover both full-field and autoregressive decoders or clarify that Eq. (4) is a schematic.
  3. [S2.2] The text says 'The dataset consists of 36 trajectories in total' and later 'training uses only 16 trajectories'; this conflates (ν,u∞) system instances with initial-condition trajectories. Please use consistent terminology (e.g., 'system instances' vs. 'rollouts').
  4. [Table 1 caption] The arrows in 'ANO↓' and 'PNO↑' are not explained in the caption. State explicitly what they denote (e.g., 'matched/lower capacity baseline' vs. 'parameter-supervised upper-bound reference').
  5. [S3.3] The GAN interpretation (Eqs. S20–S23) is heuristic and could mislead readers into thinking there is a formal adversarial equivalence. Consider removing it or clearly labeling it as an analogy only.
  6. [S2.3, Table S7] The cubic polynomial has 10 coefficients fit on 9 training points; the ridge regularization strength is not reported. Given the small gain over quadratic (0.0178 vs. 0.0179), please report the regularization parameter and a leave-one-out stability check.

Circularity Check

2 steps flagged

Zero-shot extrapolation and dimension discovery are partly constructed from supervised inputs: latent locations for unseen systems come from a polynomial fit to true factor labels, and d* is chosen using known factor-aligned structure.

specific steps
  1. fitted input called prediction [Main text, FitzHugh-Nagumo 'zero-shot generalization' paragraph; Supplementary S2.3, Eq. (S7) and Table S7]
    "To evaluate this behavior, we fit analytical mappings of increasing complexity between the latent space and the parameter space and generate predictions directly from sampled latent representations. ... The fitted quadratic diffeomorphism used in the main text takes the explicit form ... fitted by least squares on the 9 training SIs."

    The paper defines zero-shot inference as 'sampling directly in the latent space' without factor labels. The only zero-shot evaluation instead fits a polynomial map c = Phi(p) by least squares to the true (k, beta) labels of the 9 training systems, then plugs in the known true (k, beta) of the 16 held-out systems to obtain their latents. Every 'unseen factor' latent is therefore supplied by a supervised regression on exactly the labels that the NOD setting says are unavailable. The reported zero-shot extrapolation is a fitted-input prediction, not a label-free latent-space sample, so the abstract's 'zero-shot extrapolation across regimes' claim is not derived from the unsupervised representation.

  2. self definitional [Methods, Dimension selection; Supplementary S2.1 Stage 2]
    "the final dimension d* is selected as the smallest candidate whose corresponding d+1 model does not produce additional factor-aligned structure in the learned latent space, assessed using principal component analysis of the per-instance latent representations. Applied to the three main case studies, this procedure yields d* = (1,1,2), matching the number of independent governing factors in each system family."

    This Stage-2 rule cannot be applied without knowing what counts as 'factor-aligned structure', which requires the true factors and their count. Choosing the smallest d whose d+1 model has no additional factor-aligned structure is effectively choosing d to equal the known number of factors, so the reported match d* = (1,1,2) is enforced by the selection criterion rather than discovered from unlabeled trajectories. The Burgers' decision is explicitly made by preferring d=1's 'clean monotonic nu-curve' over d=2's spurious axis, an inspection that uses the known viscosity labels. The claim that NOD recovers the intrinsic dimensionality of system variation is therefore partly self-definitional.

full rationale

The core trajectory-only training of NODnet is not circular: the encoder and decoder are trained on grouped trajectories with no factor labels, and one-shot inference is tested by feeding a genuinely unseen trajectory through the encoder. Latent-factor alignment is also checked against external ground-truth factors, which is legitimate evidence. However, two load-bearing parts of the central narrative do reduce to supervised inputs. First, the only zero-shot evaluation (FHN) constructs latents for unseen systems from a polynomial map fitted by least squares to the true (k,beta) labels and evaluated at the true held-out labels, so the 'zero-shot' generalization claim is a supervised regression renamed as label-free sampling. Second, the dimension-selection Stage 2 uses knowledge of factor-aligned structure (including the known factor count) to pick d*, so reporting d* = (1,1,2) is in part a consequence of the selection rule, not an emergent unsupervised discovery. The paper also does not rely on any load-bearing self-citation or imported uniqueness theorem; the identifiability discussion in S3.4 is analogical, and the authors state that 'a formal theoretical characterization of these properties remains open.' Given these two partial reductions, the appropriate score is 6: some central predictions are partly forced by construction, but the main latent representation itself retains independent, externally validated content.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

No new physical entities are postulated. The model introduces learned latent codes and a fitted latent-factor map. The central assumptions are instance grouping, TDS feasibility, and the minimal-dimension principle; none of these is formally proved, and several are only empirically motivated.

free parameters (3)
  • latent dimension d* = d* = 1 (Burgers), 1 (cylinder wake), 2 (FitzHugh–Nagumo)
    Selected by the two-stage rule: the smallest dimension whose d+1 run lacks 'additional factor-aligned structure'. This uses knowledge of true factors and human PCA inspection, so it is not purely data-driven.
  • zero-shot polynomial diffeomorphism coefficients = quadratic map, 6 coefficients per latent coordinate (Eq. S7)
    Fitted by least squares on the 9 training (k, beta) systems. Used to generate latents for unseen systems, so the 'zero-shot' claim partly rests on a supervised post-hoc fit.
  • PNO calibration constants for sqrt(nu) transform = alpha, beta calibrated so nine trained viscosities map to [-2, +2]
    Used only to construct the parameter-supervised reference in S2.1; not part of NODnet itself, but a hand-fitted input to a comparison claim.
axioms (6)
  • domain assumption Trajectories are grouped by system instance, and all trajectories within a group share the same hidden governing factors
    The method requires instance-grouping supervision; grouping must be correct and each group must correspond to one factor setting (main text after Eq. 1).
  • domain assumption At least two trajectories per instance with distinct initial conditions are available for trajectory-decoupled sampling
    TDS (Eqs. 6–7) requires a conditioning trajectory and a prediction trajectory from the same instance but different initial conditions.
  • ad hoc to paper Minimal latent dimension equals the number of independent governing factors
    The dimension-selection procedure presumes the elbow and latent-structure criteria identify true factor count; no proof is given, only empirical case studies (Methods 'Dimension selection').
  • ad hoc to paper Decoder supervision must include full-field or multi-step horizons for factor identification
    Burgers and FitzHugh–Nagumo results show one-step AR supervision collapses the latent; this is an empirical design principle, not derived (S2.1, S2.3).
  • domain assumption Solution operator and learned decoder are continuous/differentiable, and the latent-factor map is injective with matched dimension, so Brouwer's invariance of domain applies
    Used in S3.3 to argue for a diffeomorphic latent space; injectivity and matched dimension are not proven.
  • ad hoc to paper Identifiability transfers from nonlinear ICA / weakly supervised disentanglement to NOD's deterministic degenerate setting
    S3.4 claims uniqueness up to permutation/sign/reparametrization by analogy to Khemakhem et al. and Locatello et al.; the conditional p(z|u_i) is a point mass, so the cited theorems do not directly apply.

pith-pipeline@v1.3.0-alltime-deepseek · 38785 in / 14073 out tokens · 146895 ms · 2026-07-31T23:42:06.386006+00:00 · methodology

0 comments
read the original abstract

Neural operators provide data-driven mappings for modeling dynamical systems. Extending them to families of systems typically requires explicit conditioning variables such as physical parameters, geometries, or boundary conditions. In many real-world settings, these quantities are unobserved. Here, we formulate neural operator discovery (NOD) as the problem of learning both shared solution operators and system-specific variation directly from heterogeneous trajectories without access to labeled governing factors. We introduce a factorized latent-conditioning formulation that jointly learns a neural operator and a low-dimensional latent representation through factorized prediction, trajectory-decoupled sampling, and dimension selection. Across diverse systems, the learned latent representation captures the intrinsic dimensionality of system variation and organizes system instances in a smooth and approximately invertible latent structure aligned with the underlying governing factors. This organization enables generalization to previously unseen system instances, including zero-shot extrapolation across regimes and stable long-horizon prediction. These results establish an interpretable paradigm for operator learning in the absence of explicit factor supervision.

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

52 extracted references · 1 canonical work pages

  1. [1]

    Brunton, J.L

    S.L. Brunton, J.L. Proctor, J.N. Kutz, Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the national academy of sciences113(15), 3932–3937 (2016)

    2016

  2. [2]

    Jacobs, B.W

    M. Jacobs, B.W. Brunton, S.L. Brunton, J.N. Kutz, R.V. Raut. Hypersindy: Deep generative modeling of nonlinear stochastic governing equations (2023). URL https://arxiv.org/abs/2310.04832

    Pith/arXiv arXiv 2023

  3. [3]

    d’Ascoli, S

    S. d’Ascoli, S. Becker, A. Mathis, P. Schwaller, N. Kilbertus. Odeformer: Symbolic regression of dynamical systems with transformers (2023). URL https://arxiv. org/abs/2310.05573

    Pith/arXiv arXiv 2023

  4. [4]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research24(89), 1–97 (2023)

    2023

  5. [5]

    Azizzadenesheli, N

    K. Azizzadenesheli, N. Kovachki, Z. Li, M. Liu-Schiaffini, J. Kossaifi, A. Anand- kumar, Neural operators for accelerating scientific simulations and design. Nature Reviews Physics6(5), 320–328 (2024)

    2024

  6. [6]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stu- art, A. Anandkumar, Fourier neural operator for parametric partial differential 18 equations. arXiv preprint arXiv:2010.08895 (2020)

    Pith/arXiv arXiv 2010

  7. [7]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G.E. Karniadakis, Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence3(3), 218–229 (2021)

    2021

  8. [8]

    H. Zhou, Y. Ma, H. Wu, H. Wang, M. Long, Unisolver: Pde-conditional transformers are universal pde solvers. arXiv preprint arXiv:2405.17527 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    J. Sun, Y. Liu, Z. Zhang, H. Schaeffer, Towards a foundation model for partial differential equations: Multioperator learning and extrapolation. Physical Review E111(3), 035304 (2025)

    2025

  10. [10]

    H. Wu, H. Luo, H. Wang, J. Wang, M. Long, Transolver: A fast transformer solver for pdes on general geometries. arXiv preprint arXiv:2402.02366 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Z. Fang, S. Wang, P. Perdikaris, Learning only on boundaries: a physics-informed neural operator for solving parametric partial differential equations in complex geometries. Neural computation36(3), 475–498 (2024)

    2024

  12. [12]

    Y. Cao, Y. Liu, L. Yang, R. Yu, H. Schaeffer, S. Osher, Vicon: Vision in-context operator networks for multi-physics fluid dynamics prediction. arXiv preprint arXiv:2411.16063 (2024)

    arXiv 2024

  13. [13]

    Z. Hao, C. Su, S. Liu, J. Berner, C. Ying, H. Su, A. Anandkumar, J. Song, J. Zhu, Dpot: Auto-regressive denoising operator transformer for large-scale pde pre-training. arXiv preprint arXiv:2403.03542 (2024)

    Pith/arXiv arXiv 2024

  14. [14]

    Z. Chen, S. Deng, Flow marching for a generative PDE foundation model. arXiv preprint arXiv:2509.18611 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    McCabe, B.R.S

    M. McCabe, B.R.S. Blancard, L.H. Parker, R. Ohana, M. Cranmer, A. Bietti, M. Eickenberg, S. Golkar, G. Krawezik, F. Lanusse, et al., Multiple physics pretraining for physical surrogate models. arXiv preprint arXiv:2310.02994 (2023)

    Pith/arXiv arXiv 2023

  16. [16]

    Arora, L.E

    K. Arora, L.E. Asri, H. Bahuleyan, J.C.K. Cheung. Why exposure bias matters: An imitation learning perspective of error accumulation in language generation (2023). URL https://arxiv.org/abs/2204.01171

    Pith/arXiv arXiv 2023

  17. [17]

    Z. Chen, S. Deng. Latent generative solvers for generalizable long-term physics simulation (2026). URL https://arxiv.org/abs/2602.11229

    Pith/arXiv arXiv 2026

  18. [18]

    P.Y. Lu, S. Kim, M. Soljaˇ ci´ c, Extracting interpretable physical parameters from spatiotemporal systems using unsupervised learning. Physical Review X10(3) (2020). https://doi.org/10.1103/physrevx.10.031056. URL http://dx.doi.org/10. 1103/PhysRevX.10.031056 19

    work page doi:10.1103/physrevx.10.031056 2020

  19. [19]

    Q. Li, T. Wang, V. Roychowdhury, M.K. Jawed, Metalearning generalizable dynamics from trajectories. Physical Review Letters131(6), 067301 (2023)

    2023

  20. [20]

    Oleinik, Discontinuous solutions of non-linear differential equations

    O.A. Oleinik, Discontinuous solutions of non-linear differential equations. Uspekhi Matematicheskikh Nauk12(3), 3–73 (1957)

    1957

  21. [21]

    Kruˇ zkov, First order quasilinear equations in several independent variables

    S.N. Kruˇ zkov, First order quasilinear equations in several independent variables. Mathematics of the USSR-Sbornik10(2), 217–243 (1970)

    1970

  22. [22]

    H. Zhao, L. Jiang, J. Jia, P.H. Torr, V. Koltun, in Proceedings of the IEEE/CVF international conference on computer vision (2021), pp. 16259–16268

    2021

  23. [23]

    Khosla, P

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu, D. Krishnan, Supervised contrastive learning. Advances in neural information processing systems33, 18661–18673 (2020)

    2020

  24. [24]

    Henderson, Details of the drag curve near the onset of vortex shedding

    R.D. Henderson, Details of the drag curve near the onset of vortex shedding. Physics of Fluids7(9), 2102 (1995)

    1995

  25. [25]

    Williamson, Vortex dynamics in the cylinder wake

    C.H.K. Williamson, Vortex dynamics in the cylinder wake. Annual Review of Fluid Mechanics28, 477–539 (1996)

    1996

  26. [26]

    Sch¨ afer, S

    M. Sch¨ afer, S. Turek, F. Durst, E. Krause, R. Rannacher, in Flow simulation with high-performance computers II: DFG priority research programme results 1993–1995 (Springer, 1996), pp. 547–566

    1993

  27. [27]

    FitzHugh, Impulses and physiological states in theoretical models of nerve membrane

    R. FitzHugh, Impulses and physiological states in theoretical models of nerve membrane. Biophysical journal1(6), 445–466 (1961)

    1961

  28. [28]

    Nagumo, S

    J. Nagumo, S. Arimoto, S. Yoshizawa, An active pulse transmission line simulat- ing nerve axon. Proceedings of the IRE50(10), 2061–2070 (1962)

    2061

  29. [29]

    Khemakhem, D

    I. Khemakhem, D. Kingma, R. Monti, A. Hyvarinen, in International Conference on Artificial Intelligence and Statistics (PMLR, 2020), pp. 2207–2217

    2020

  30. [30]

    Locatello, B

    F. Locatello, B. Poole, G. R¨ atsch, B. Sch¨ olkopf, O. Bachem, M. Tschannen, in International Conference on Machine Learning (PMLR, 2020), pp. 6348–6359

    2020

  31. [31]

    Hyv¨ arinen, H

    A. Hyv¨ arinen, H. Sasaki, R. Turner, in International Conference on Artificial Intelligence and Statistics (PMLR, 2019), pp. 859–868

    2019

  32. [32]

    B. Chen, K. Huang, S. Raghupathi, I. Chandratreya, Q. Du, H. Lipson, Auto- mated discovery of fundamental variables hidden in experimental data. Nature Computational Science2(7), 433–442 (2022)

    2022

  33. [33]

    Chen, B.W

    X. Chen, B.W. Soh, Z.E. Ooi, E. Vissol-Gaudin, H. Yu, K.S. Novoselov, K. Hip- palgaonkar, Q. Li, Constructing custom thermodynamics using deep learning. Nature Computational Science4(1), 66–85 (2024) 20

    2024

  34. [34]

    Goodfellow, J

    I.J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio. Generative adversarial networks (2014). URL https: //arxiv.org/abs/1406.2661

    Pith/arXiv arXiv 2014

  35. [35]

    Alkin, A

    B. Alkin, A. F¨ urst, S. Schmid, L. Gruber, M. Holzleitner, J. Brandstetter, Uni- versal physics transformers: A framework for efficiently scaling neural operators. Advances in Neural Information Processing Systems37, 25152–25194 (2024)

    2024

  36. [36]

    Peebles, S

    W. Peebles, S. Xie, in Proceedings of the IEEE/CVF international conference on computer vision (2023), pp. 4195–4205

    2023

  37. [37]

    J. Su, Y. Lu, S. Pan, A. Murtadha, B. Wen, Y. Liu. Roformer: Enhanced trans- former with rotary position embedding (2023). URL https://arxiv.org/abs/2104. 09864

    2023

  38. [38]

    Chung, C

    J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

    Pith/arXiv arXiv 2014

  39. [39]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory. Neural computation 9(8), 1735–1780 (1997)

    1997

  40. [40]

    K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, Q. Tian, Accurate medium-range global weather forecasting with 3D neural networks. Nature619(7970), 533–538 (2023)

    2023

  41. [41]

    Green, G

    R. Green, G. Stevens, Z. Abdallah, T.M.S. Filho. Stratify: Unifying multi-step forecasting strategies (2024). URL https://arxiv.org/abs/2412.20510

    Pith/arXiv arXiv 2024

  42. [42]

    R.T. Chen, Y. Rubanova, J. Bettencourt, D.K. Duvenaud, Neural ordinary differential equations. Advances in neural information processing systems31 (2018)

    2018

  43. [43]

    Lee, J.M

    J.M. Lee, J.M. Lee, Smooth manifolds (Springer, 2003)

    2003

  44. [44]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, in Proceedings of the IEEE conference on computer vision and pattern recognition (2017), pp. 4700– 4708

    2017

  45. [45]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F.A. Hamprecht, Y. Ben- gio, A. Courville, in Proceedings of the 36th International Conference on Machine Learning (ICML) (2019), pp. 5301–5310

    2019

  46. [46]

    Tancik, P.P

    M. Tancik, P.P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J.T. Barron, R. Ng, in Advances in Neural Information Processing Systems (NeurIPS), vol. 33 (2020), pp. 7537–7547

    2020

  47. [47]

    Spanier, Algebraic topology (Springer Science & Business Media, 2012) 21

    E.H. Spanier, Algebraic topology (Springer Science & Business Media, 2012) 21

    2012

  48. [48]

    Locatello, S

    F. Locatello, S. Bauer, M. Lucic, G. R¨ atsch, S. Gelly, B. Sch¨ olkopf, O. Bachem, in International Conference on Machine Learning (PMLR, 2019), pp. 4114–4124

    2019

  49. [49]

    Klindt, L

    D.A. Klindt, L. Schott, Y. Sharma, I. Ustyuzhaninov, W. Brendel, M. Bethge, D. Paiton, in International Conference on Learning Representations (2021)

    2021

  50. [50]

    Udrescu, M

    S.M. Udrescu, M. Tegmark, AI Feynman: A physics-inspired method for symbolic regression. Science advances6(16), eaay2631 (2020)

    2020

  51. [51]

    Cranmer, Interpretable machine learning for science with pysr and symboli- cregression

    M. Cranmer, Interpretable machine learning for science with pysr and symboli- cregression. jl. arXiv preprint arXiv:2305.01582 (2023)

    Pith/arXiv arXiv 2023

  52. [52]

    elbow” shape as the latent dimension increases; the “elbow

    S. Rudy, A. Alla, S.L. Brunton, J.N. Kutz, Data-driven identification of paramet- ric partial differential equations. SIAM Journal on Applied Dynamical Systems 18(2), 643–660 (2019) Acknowledgments Z.C. and S.D. acknowledge financial support from MIT Sea Grant. Q.L. and J.H. are supported by the Major Research Plan of the National Natural Science Founda- ...

    arXiv 2019