pith. machine review for the scientific record.
sign in

arxiv: 2511.06609 · v3 · submitted 2025-11-10 · 💻 cs.LG · math.DS

A Weak Penalty Neural ODE for Learning Chaotic Dynamics from Noisy Time Series

Xuyang Li , John Harlim , Dibyajyoti Chakraborty , Romit Maulik This is my paper

Pith reviewed 2026-05-18 00:01 UTC · model grok-4.3

classification 💻 cs.LG math.DS
keywords Neural ODEchaotic dynamicsweak formulationnoisy time seriesdynamical systemsforecastingmachine learningclimate data
0
0 comments X

The pith

A weak formulation of the loss lets Neural ODEs filter noise and forecast chaotic dynamics accurately.

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

The paper develops a training method for Neural ODEs that replaces or augments the usual pointwise loss with a weak penalty term derived from the differential equation residual. This term is formed by integrating the residual against selected test functions over chosen domains, which the authors show acts like a built-in smoother that removes high-frequency noise while leaving the underlying flow intact. As a result, the trained model produces short-term predictions that remain accurate even when initial errors would normally explode in chaotic systems, and it also maintains correct long-term statistical behavior. A reader would care because real measurements of weather, fluids, or other nonlinear systems are almost always noisy, and this offers a direct, solver-independent way to obtain usable forecasts without separate denoising steps.

Core claim

The authors show that the weak form of the Neural ODE residual, when used as a penalty in the training objective with suitable test functions and integration intervals, filters measurement noise in a manner equivalent to fitting the model to pre-filtered data. This formulation stabilizes training for chaotic attractors, where standard L2 losses on noisy trajectories lead to rapid divergence, and produces forecasts whose short-term accuracy and long-term invariants both improve. The method is demonstrated to be computationally efficient and independent of the particular ODE integrator chosen during training or prediction.

What carries the argument

The weak penalty loss, formed by integrating the Neural ODE residual against test functions over finite domains so that high-frequency noise contributions average toward zero.

If this is right

  • The trained models achieve both short-term forecast accuracy and preservation of long-term invariant properties on standard chaotic benchmarks.
  • The same strategy produces usable forecasts on a real-world climate dataset despite observational noise.
  • Training remains efficient and works with any choice of numerical ODE solver.
  • The approach avoids the exponential error growth that plagues standard Neural ODE training on noisy chaotic time series.

Where Pith is reading between the lines

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

  • The same weak-penalty idea could be applied to other neural architectures that learn differential equations, such as neural SDEs or graph-based dynamical models.
  • Optimal selection of test functions for a given noise spectrum or system dimension remains an open design choice that future work could automate.
  • Because the method operates directly on the integral form, it may combine naturally with existing data-assimilation pipelines used in operational forecasting.

Load-bearing premise

The weak formulation with a proper choice of test function and integration domain effectively filters noisy data.

What would settle it

Train the model on a benchmark chaotic system with added noise, then verify whether the long-term statistics of simulated trajectories, such as the shape of the attractor or the leading Lyapunov exponent, match those obtained from clean data to within a small tolerance.

Figures

Figures reproduced from arXiv: 2511.06609 by Dibyajyoti Chakraborty, John Harlim, Romit Maulik, Xuyang Li.

Figure 1
Figure 1. Figure 1: FIG. 1. Forecasting of the Lorenz-63 system using WP-NODE under 5% training data noise. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Invariant measure of the predicted Lorenz-63 system (for 100 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Performance comparison of WP-NODE and baselines models on the KS system, under 5% data noise. DeepSkip is [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Comparison of the long-term behavior of the KS system under 5% data noise, using the invariant measure (i.e., joint [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. NODE predictions performance on real-world ERA5 dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

The accurate forecasting of complex, high-dimensional dynamical systems from observational data is a fundamental task across numerous scientific and engineering disciplines. A significant challenge arises from noise-corrupted measurements, which severely degrade the performance of data-driven models. In chaotic dynamical systems, where small initial errors amplify exponentially, it is particularly difficult to develop a model from noisy data that achieves short-term accuracy while preserving long-term invariant properties. To overcome this, we consider the weak formulation as a complementary approach to the classical $L2$-loss function for training models of dynamical systems. We empirically verify that the weak formulation, with a proper choice of test function and integration domain, effectively filters noisy data. This insight explains why a weak form loss function is analogous to fitting a model to filtered data and provides a practical way to parameterize the weak form. Subsequently, we demonstrate how this approach overcomes the instability and inaccuracy of standard Neural ODE (NODE) in modeling chaotic systems. Through numerical examples, we show that our proposed training strategy, the Weak Penalty NODE, is computationally efficient, solver-agnostic, and yields accurate and robust forecasts across benchmark chaotic systems and a real-world climate dataset.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the Weak Penalty Neural ODE (WP-NODE), which augments standard Neural ODE training with a weak-form loss term. The central claim is that, for a suitable choice of test function and integration domain, the weak formulation filters noise in time-series observations of chaotic systems, yielding short-term forecast accuracy while preserving long-term invariants; the method is further asserted to be computationally efficient and solver-agnostic, with supporting numerical results on benchmark chaotic systems and a real-world climate dataset.

Significance. If the noise-filtering property and long-term stability claims hold under broader conditions, the work would offer a practical, preprocessing-free route to training NODEs on noisy chaotic data. This could be useful in climate modeling and other domains where explicit denoising is costly or distorts invariants, and the solver-agnostic character would be a clear practical advantage.

major comments (2)
  1. [Abstract and weak-formulation section] Abstract and method description of the weak formulation: the assertion that 'the weak formulation, with a proper choice of test function and integration domain, effectively filters noisy data' is load-bearing for the entire approach, yet no general rule, stability analysis, or selection criterion is supplied for these choices. Without such guidance the reported success on the chosen benchmarks may reflect case-specific tuning rather than a structural guarantee, leaving the robustness claim for varying noise statistics or system dimensionality unproven.
  2. [Numerical experiments] Numerical experiments on long-term forecasts: while short-term accuracy and computational efficiency are shown, quantitative evidence that long-term invariants (e.g., Lyapunov exponents or attractor geometry) are preserved independently of the specific test-function parameters is limited. This weakens the claim that the method reliably overcomes standard NODE instability for chaotic systems.
minor comments (2)
  1. [Method] The notation used for the weak penalty loss should be contrasted explicitly with the classical L2 NODE loss to make the algorithmic difference immediately clear.
  2. [Numerical experiments] A short table or paragraph quantifying the sensitivity of forecast error to the integration-domain length would help readers assess practical implementation cost.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive and insightful comments. We address each major comment point by point below, indicating where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and weak-formulation section] Abstract and method description of the weak formulation: the assertion that 'the weak formulation, with a proper choice of test function and integration domain, effectively filters noisy data' is load-bearing for the entire approach, yet no general rule, stability analysis, or selection criterion is supplied for these choices. Without such guidance the reported success on the chosen benchmarks may reflect case-specific tuning rather than a structural guarantee, leaving the robustness claim for varying noise statistics or system dimensionality unproven.

    Authors: We appreciate this observation. The manuscript presents the approach as an empirical method and supplies a practical parameterization by showing that the weak-form loss corresponds to fitting filtered data, along with guidance on matching test-function support to system time scales (Section 3). We agree that more explicit selection criteria would improve clarity. In the revised manuscript we have added a dedicated paragraph in the method section that outlines concrete rules: choose the integration domain to exceed the noise correlation time but remain shorter than the system's Lyapunov time, and select test functions whose frequency content overlaps the dominant dynamics while attenuating high-frequency noise. A full stability analysis under arbitrary noise statistics and dimensionality is not derived, as it would require additional theoretical assumptions and development beyond the paper's empirical focus; robustness is instead demonstrated across multiple chaotic benchmarks with varying noise levels and dimensions. revision: partial

  2. Referee: [Numerical experiments] Numerical experiments on long-term forecasts: while short-term accuracy and computational efficiency are shown, quantitative evidence that long-term invariants (e.g., Lyapunov exponents or attractor geometry) are preserved independently of the specific test-function parameters is limited. This weakens the claim that the method reliably overcomes standard NODE instability for chaotic systems.

    Authors: We concur that stronger quantitative support for parameter independence would reinforce the long-term stability claim. The original experiments reported results for effective parameter choices. To address the concern, the revised manuscript now includes a sensitivity study: Lyapunov exponents and attractor geometry metrics (e.g., correlation dimension) are computed over a grid of test-function widths and integration lengths for each benchmark system. The results show that these invariants remain consistent within a practically useful range of parameters, with degradation only outside that range. Corresponding figures and tables have been added to the numerical experiments section. revision: yes

standing simulated objections not resolved
  • A rigorous theoretical stability analysis of the weak formulation for arbitrary noise statistics and system dimensions is not provided.

Circularity Check

0 steps flagged

No significant circularity; method relies on empirical verification of weak-form filtering rather than self-referential derivation

full rationale

The paper presents the Weak Penalty NODE as a training strategy that leverages the weak formulation of dynamical systems to handle noisy data. The key assertion—that the weak form with suitable test functions and domains filters noise and is analogous to fitting filtered data—is stated as an empirical insight verified through numerical examples on benchmark chaotic systems and climate data. No load-bearing step reduces by construction to a fitted parameter renamed as prediction, a self-citation chain, or an ansatz smuggled via prior work. The derivation chain consists of proposing the penalty-augmented loss, demonstrating solver-agnostic training, and reporting forecast accuracy; these rest on external benchmarks and observed performance rather than tautological reparameterization of inputs. Self-contained against the provided numerical validation, this yields a normal non-finding of circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach depends on the effectiveness of the weak form in filtering noise, which is taken as given and verified empirically rather than derived from first principles.

free parameters (1)
  • test function and integration domain
    Proper choice is required to filter noise effectively as stated in the abstract.
axioms (1)
  • domain assumption Weak formulation filters noisy data in dynamical systems when test functions and domains are chosen properly
    Invoked to explain why the weak form loss works analogously to fitting filtered data.

pith-pipeline@v0.9.0 · 5514 in / 1057 out tokens · 41573 ms · 2026-05-18T00:01:17.677929+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · 1 internal anchor

  1. [1]

    Evaluation is based on a validation threshold ofε= 0.3

    In our numerical experiment, we integrate this system of ODEs using the RK4 scheme with a time step of ∆t= 0.01sfor 100s, yieldingN= 10 4 samples. Evaluation is based on a validation threshold ofε= 0.3. Under these parameters, the largest Lyapunov exponent is Λ≈0.91. First, 5% data noise is applied to the observation data, where a zero-mean Gaussian is de...

    work page 2000

  2. [2]

    Given an initial conditionu(t 0) =u 0, the evolu- tion of the state is governed by: du(t) dt =f(u(t), t;θ),(5) wherefis a neural network parameterized byθ

    generalize residual networks by replacing discrete- layer architectures with a continuous-time dynamics model. Given an initial conditionu(t 0) =u 0, the evolu- tion of the state is governed by: du(t) dt =f(u(t), t;θ),(5) wherefis a neural network parameterized byθ. The so- lution at a later timetis determined by solving this initial value problem using s...

    work page

  3. [3]

    Ghadami and B

    A. Ghadami and B. I. Epureanu, Data-driven prediction in dynamical systems: recent developments, Philosophi- cal Transactions of the Royal Society A380, 20210213 (2022)

    work page 2022

  4. [4]

    Morrison,The art of modeling dynamic systems: fore- casting for chaos, randomness and determinism(Courier Corporation, 2012)

    F. Morrison,The art of modeling dynamic systems: fore- casting for chaos, randomness and determinism(Courier Corporation, 2012)

    work page 2012

  5. [5]

    Kaheman, E

    K. Kaheman, E. Kaiser, B. Strom, J. N. Kutz, and S. L. Brunton, Learning discrepancy models from experimen- tal data, arXiv preprint arXiv:1909.08574 (2019)

    work page arXiv 1909

  6. [6]

    J. Hart, S. A. McQuarrie, Z. Morrow, and B. v. B. Waanders, Toward real-time optimization through model reduction and model discrepancy sensitivities, arXiv preprint arXiv:2508.21792 (2025)

    work page arXiv 2025

  7. [7]

    S. L. Brunton and J. N. Kutz,Data-driven science and engineering: Machine learning, dynamical systems, and control(Cambridge University Press, 2022)

    work page 2022

  8. [8]

    R. Wang, D. Maddix, C. Faloutsos, Y. Wang, and R. Yu, Bridging physics-based and data-driven modeling for learning dynamical systems, inLearning for dynamics and control(PMLR, 2021) pp. 385–398

    work page 2021

  9. [9]

    Floryan and M

    D. Floryan and M. D. Graham, Data-driven discovery of intrinsic dynamics, Nature Machine Intelligence4, 1113 (2022)

    work page 2022

  10. [10]

    M. J. Colbrook and A. Townsend, Rigorous data-driven computation of spectral properties of koopman operators for dynamical systems, Communications on Pure and Ap- plied Mathematics77, 221 (2024)

    work page 2024

  11. [11]

    A. A. Kaptanoglu, J. L. Callaham, A. Aravkin, C. J. Hansen, and S. L. Brunton, Promoting global stability in data-driven models of quadratic nonlinear dynamics, Physical Review Fluids6, 094401 (2021)

    work page 2021

  12. [12]

    Legaard, T

    C. Legaard, T. Schranz, G. Schweiger, J. Drgoˇ na, B. Falay, C. Gomes, A. Iosifidis, M. Abkar, and P. Larsen, Constructing neural network based models for simulat- ing dynamical systems, ACM Computing Surveys55, 1 (2023)

    work page 2023

  13. [13]

    Gilpin, Chaos as an interpretable benchmark for forecasting and data-driven modelling, arXiv preprint arXiv:2110.05266 (2021)

    W. Gilpin, Chaos as an interpretable benchmark for forecasting and data-driven modelling, arXiv preprint arXiv:2110.05266 (2021)

    work page arXiv 2021

  14. [14]

    Levine and A

    M. Levine and A. Stuart, A framework for machine learning of model error in dynamical systems, Commu- nications of the American Mathematical Society2, 283 (2022)

    work page 2022

  15. [15]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning representations by back-propagating errors, na- ture323, 533 (1986)

    work page 1986

  16. [16]

    Chattopadhyay, P

    A. Chattopadhyay, P. Hassanzadeh, and D. Subrama- nian, Data-driven prediction of a multi-scale lorenz 96 chaotic system using deep learning methods: Reservoir computing, ann, and rnn-lstm, Nonlinear Processes in Geophysics Discussions2020, 1 (2020)

    work page 2020

  17. [17]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber, Long short-term mem- ory, Neural computation9, 1735 (1997)

    work page 1997

  18. [18]

    Harlim, S

    J. Harlim, S. W. Jiang, S. Liang, and H. Yang, Machine learning for prediction with missing dynamics, Journal of Computational Physics428, 109922 (2021)

    work page 2021

  19. [19]

    M. U. Kobayashi, K. Nakai, Y. Saiki, and N. Tsutsumi, Dynamical system analysis of a data-driven model con- structed by reservoir computing, Physical Review E104, 044215 (2021)

    work page 2021

  20. [20]

    M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, Emerging opportunities and challenges for the future of reservoir computing, Nature Communications 15, 2056 (2024)

    work page 2056

  21. [21]

    F. M. Bianchi, S. Scardapane, S. Løkse, and R. Jenssen, Reservoir computing approaches for representation and classification of multivariate time series, IEEE transac- tions on neural networks and learning systems32, 2169 (2020)

    work page 2020

  22. [22]

    C. Lea, R. Vidal, A. Reiter, and G. D. Hager, Tempo- ral convolutional networks: A unified approach to action segmentation, inEuropean conference on computer vision (Springer, 2016) pp. 47–54

    work page 2016

  23. [23]

    Perumal, D

    V. Perumal, D. Abueidda, S. Koric, and A. Kontsos, Temporal convolutional networks for data-driven ther- mal modeling of directed energy deposition, Journal of Manufacturing Processes85, 405 (2023)

    work page 2023

  24. [24]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, At- tention is all you need, Advances in neural information processing systems30(2017)

    work page 2017

  25. [25]

    Y. He, Y. Yang, X. Cheng, H. Wang, X. Xue, B. Chen, and Y. Hu, Chaos meets attention: Transform- ers for large-scale dynamical prediction, arXiv preprint arXiv:2504.20858 (2025)

    work page arXiv 2025

  26. [26]

    Patil, J

    A. Patil, J. Viquerat, and E. Hachem, Autoregressive transformers for data-driven spatiotemporal learning of turbulent flows, APL Machine Learning1(2023)

    work page 2023

  27. [27]

    Mandal and G

    P. Mandal and G. A. Gottwald, Learning dynamical systems with hit-and-run random feature maps, Nature Communications16, 5961 (2025)

    work page 2025

  28. [28]

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

    work page 2018

  29. [29]

    Y. Oh, S. Kam, J. Lee, D.-Y. Lim, S. Kim, and A. Bui, Comprehensive review of neural differential equations for time series analysis, arXiv preprint arXiv:2502.09885 (2025)

    work page arXiv 2025

  30. [30]

    Kidger, On neural differential equations, arXiv preprint arXiv:2202.02435 (2022)

    P. Kidger, On neural differential equations, arXiv preprint arXiv:2202.02435 (2022)

    work page arXiv 2022

  31. [31]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox, U-net: Con- volutional networks for biomedical image segmentation, inInternational Conference on Medical image computing and computer-assisted intervention(Springer, 2015) pp. 234–241

    work page 2015

  32. [32]

    Universal Differential Equations for Scientific Machine Learning

    C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ramadhan, and A. Edelman, Universal differential equations for scien- tific machine learning, arXiv preprint arXiv:2001.04385 (2020)

    work page internal anchor Pith review arXiv 2001

  33. [33]

    Y. Ma, V. Dixit, M. J. Innes, X. Guo, and C. Rack- 11 auckas, A comparison of automatic differentiation and continuous sensitivity analysis for derivatives of differen- tial equation solutions, in2021 IEEE High Performance Extreme Computing Conference (HPEC)(IEEE, 2021) pp. 1–9

    work page 2021

  34. [34]

    Chakraborty, S

    D. Chakraborty, S. W. Chung, T. Arcomano, and R. Maulik, Divide and conquer: Learning chaotic dy- namical systems with multistep penalty neural ordinary differential equations, Computer Methods in Applied Me- chanics and Engineering432, 117442 (2024)

    work page 2024

  35. [35]

    Chakraborty, S

    D. Chakraborty, S. W. Chung, A. Chattopadhyay, and R. Maulik, Improved deep learning of chaotic dynami- cal systems with multistep penalty losses, arXiv preprint arXiv:2410.05572 (2024)

    work page arXiv 2024

  36. [36]

    Allauzen, T

    A. Allauzen, T. P. M. Dardis, and H. Plath, Exper- imental study of neural ode training with adaptive solver for dynamical systems modeling, arXiv preprint arXiv:2211.06972 (2022)

    work page arXiv 2022

  37. [37]

    A. J. Linot, J. W. Burby, Q. Tang, P. Balaprakash, M. D. Graham, and R. Maulik, Stabilized neural ordinary dif- ferential equations for long-time forecasting of dynamical systems, Journal of Computational Physics474, 111838 (2023)

    work page 2023

  38. [38]

    Fronk and L

    C. Fronk and L. Petzold, Training stiff neural ordinary differential equations with explicit rational taylor series methods, Chaos: An Interdisciplinary Journal of Nonlin- ear Science35(2025)

    work page 2025

  39. [39]

    Fronk and L

    C. Fronk and L. Petzold, Training stiff neural ordinary differential equations with implicit single-step methods, Chaos: An Interdisciplinary Journal of Nonlinear Science 34(2024)

    work page 2024

  40. [40]

    Cheng, C

    S. Cheng, C. Quilodr´ an-Casas, S. Ouala, A. Farchi, C. Liu, P. Tandeo, R. Fablet, D. Lucor, B. Iooss, J. Bra- jard,et al., Machine learning with data assimilation and uncertainty quantification for dynamical systems: a re- view, IEEE/CAA Journal of Automatica Sinica10, 1361 (2023)

    work page 2023

  41. [41]

    Mukhopadhyay and S

    S. Mukhopadhyay and S. Banerjee, Learning dynami- cal systems in noise using convolutional neural networks, Chaos: An Interdisciplinary Journal of Nonlinear Science 30(2020)

    work page 2020

  42. [42]

    D. A. Messenger and D. M. Bortz, Weak sindy for partial differential equations, Journal of Computational Physics 443, 110525 (2021)

    work page 2021

  43. [43]

    J. Hsin, S. Agarwal, A. Thorpe, L. Sentis, and D. Fridovich-Keil, Symbolic regression on sparse and noisy data with gaussian processes, in2025 American Control Conference (ACC)(IEEE, 2025) pp. 3170–3175

    work page 2025

  44. [44]

    S. Yang, S. W. Wong, and S. Kou, Inference of dy- namic systems from noisy and sparse data via manifold- constrained gaussian processes, Proceedings of the Na- tional Academy of Sciences118, e2020397118 (2021)

    work page 2021

  45. [45]

    Stegle, S

    O. Stegle, S. V. Fallert, D. J. MacKay, and S. Brage, Gaussian process robust regression for noisy heart rate data, IEEE Transactions on Biomedical Engineering55, 2143 (2008)

    work page 2008

  46. [46]

    Seleznev, D

    A. Seleznev, D. Mukhin, A. Gavrilov, E. Loskutov, and A. Feigin, Bayesian framework for simulation of dynam- ical systems from multidimensional data using recurrent neural network, Chaos: An Interdisciplinary Journal of Nonlinear Science29(2019)

    work page 2019

  47. [47]

    L. Yang, X. Meng, and G. E. Karniadakis, B-pinns: Bayesian physics-informed neural networks for forward and inverse pde problems with noisy data, Journal of Computational Physics425, 109913 (2021)

    work page 2021

  48. [48]

    Z. Guo, I. Cialenco, and M. Zhong, Learning stochas- tic dynamics from data, arXiv preprint arXiv:2403.02595 (2024)

    work page arXiv 2024

  49. [49]

    Ding and C

    L. Ding and C. Wen, High-order extended kalman filter for state estimation of nonlinear systems, Symmetry16, 617 (2024)

    work page 2024

  50. [50]

    Stephany and C

    R. Stephany and C. Earls, Weak-pde-learn: A weak form based approach to discovering pdes from noisy, lim- ited data, Journal of Computational Physics506, 112950 (2024)

    work page 2024

  51. [51]

    D. M. Bortz, D. A. Messenger, and A. Tran, Weak form- based data-driven modeling: Computationally efficient and noise robust equation learning and parameter infer- ence, inHandbook of Numerical Analysis, Vol. 25 (Else- vier, 2024) pp. 53–82

    work page 2024

  52. [52]

    Rummel, D

    N. Rummel, D. A. Messenger, S. Becker, V. Dukic, and D. M. Bortz, Wendy for nonlinear-in-parameters odes, arXiv preprint arXiv:2502.08881 (2025)

    work page arXiv 2025

  53. [53]

    De Ryck, S

    T. De Ryck, S. Mishra, and R. Molinaro, wpinns: Weak physics informed neural networks for approximating en- tropy solutions of hyperbolic conservation laws, SIAM Journal on Numerical Analysis62, 811 (2024)

    work page 2024

  54. [54]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, and G. E. Karniadakis, hp- vpinns: Variational physics-informed neural networks with domain decomposition, Computer Methods in Ap- plied Mechanics and Engineering374, 113547 (2021)

    work page 2021

  55. [55]

    D. A. Messenger and D. M. Bortz, Weak sindy: Galerkin- based data-driven model selection, Multiscale Modeling & Simulation19, 1474 (2021)

    work page 2021

  56. [56]

    H. Zhao, Y. Wang, H. Qi, Z. Huang, H. Zhao, L. Sha, and H. Shao, Accelerating neural odes: A variational formulation-based approach, inThe Thirteenth Interna- tional Conference on Learning Representations(2025)

    work page 2025

  57. [57]

    Kullback and R

    S. Kullback and R. A. Leibler, On information and suf- ficiency, The annals of mathematical statistics22, 79 (1951)

    work page 1951

  58. [58]

    E. N. Lorenz, Deterministic nonperiodic flow 1, inUni- versality in Chaos, 2nd edition(Routledge, 2017) pp. 367–378

    work page 2017

  59. [59]

    E. N. Lorenz, Predictability: A problem partly solved, in Proc. Seminar on predictability, Vol. 1 (Readinhig, 1996) pp. 1–18

    work page 1996

  60. [60]

    R. A. Edson, J. E. Bunder, T. W. Mattner, and A. J. Roberts, Lyapunov exponents of the kuramoto– sivashinsky pde, The ANZIAM Journal61, 270 (2019)

    work page 2019

  61. [61]

    H. Fan, J. Jiang, C. Zhang, X. Wang, and Y.-C. Lai, Long-term prediction of chaotic systems with machine learning, Physical Review Research2, 012080 (2020)

    work page 2020

  62. [62]

    Hersbach, B

    H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Hor´ anyi, J. Mu˜ noz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers,et al., The era5 global reanaly- sis, Quarterly journal of the royal meteorological society 146, 1999 (2020)

    work page 1999

  63. [63]

    Arcomano, I

    T. Arcomano, I. Szunyogh, A. Wikner, J. Pathak, B. R. Hunt, and E. Ott, A hybrid approach to atmospheric modeling that combines machine learning with a physics- based numerical model, Journal of Advances in Modeling Earth Systems14, e2021MS002712 (2022)

    work page 2022

  64. [64]

    H. Guan, T. Arcomano, A. Chattopadhyay, and R. Maulik, Lucie: A lightweight uncoupled climate emulator with long-term stability and physical consis- tency for o (1000)-member ensembles, arXiv preprint arXiv:2405.16297 (2024). 12

    work page arXiv 2024

  65. [65]

    J. R. Dormand and P. J. Prince, A family of embedded runge-kutta formulae, Journal of computational and ap- plied mathematics6, 19 (1980)

    work page 1980

  66. [66]

    J. C. Butcher,Numerical methods for ordinary differen- tial equations(John Wiley & Sons, 2016). 13 Supplementary Information SUPPLEMENT AR Y NOTE 1: LORENZ 63 INV ARIANT MEASURE ANAL YSIS To complement the quantitative KL divergence results presented in Table 1 of the article, this note provides a detailed visual comparison of the predicted and ground-tru...

    work page 2016

  67. [67]

    For the endpoints: ℓ0(s) =    s1 −s s1 −s 0 , s∈[s 0, s1] 0,otherwise ℓM−1(s) =    s−s M−2 sM−1 −s M−2 , s∈[s M−2 , sM−1] 0,otherwise

    Linear Basis Functions.Define the piecewise linear Lagrange basis functionsℓ i(s), ℓi(s) =    s−s i−1 si −s i−1 , s∈[s i−1, si] si+1 −s si+1 −s i , s∈[s i, si+1] 0,otherwise for 1≤i≤M−2. For the endpoints: ℓ0(s) =    s1 −s s1 −s 0 , s∈[s 0, s1] 0,otherwise ℓM−1(s) =    s−s M−2 sM−1 −s M−2 , s∈[s M−2 , sM−1] 0,otherwise

    work page

  68. [68]

    Note that ifx i, fi ∈R D are vectors andℓ i(s)∈Ris a scalar, the termsx iℓi(s) andf iℓi(s) represent the standard scalar-vector products

    Interpolant Construction.We define the linear interpolants of the functionsx(s) andf(u(s), s;θ) as: ˜x(s) = M−1X i=0 xiℓi(s), ˜f(s) = M−1X i=0 fiℓi(s).(S2) These functions are continuous and piecewise linear. Note that ifx i, fi ∈R D are vectors andℓ i(s)∈Ris a scalar, the termsx iℓi(s) andf iℓi(s) represent the standard scalar-vector products

    work page

  69. [69]

    Projection Against a Test Function.We approximate the following integrals using Eq. (S2) as follows, L 2 Z sM−1 s0 f(x(s), s;θ)ϕ(s)ds≈ L 2 Z sM−1 s0 ˜f(s)ϕ(s)ds= M−1X i=0 fi L 2 Z sM−1 s0 ℓi(s)ϕ(s)ds = M−1X i=0 fiwrhs,i, where the weights are defined as: wrhs,i = L 2 Z sM−1 s0 ℓi(s)ϕ(s)ds. 22 Similarly, Z sM−1 s0 x(s)ϕ′(s)ds≈ Z sM−1 s0 ˜x(s)ϕ′(s)ds= M−1X ...

    work page

  70. [70]

    Explicit Expression for Weights.For interior indices 1≤i≤M−2, we have: wrhs,i = L 2 Z si si−1 s−s i−1 si −s i−1 ϕ(s)ds+ L 2 Z si+1 si si+1 −s si+1 −s i ϕ(s)ds = L 2(si −s i−1) [Ψ(si)−Ψ(s i−1)−s i−1(Φ(si)−Φ(s i−1))] + L 2(si+1 −s i) [si+1(Φ(si+1)−Φ(s i))−Ψ(s i+1) + Ψ(si)], where Φp,d(s) = R ϕ(d) p (s)dsand Ψ p,d(s) = R sϕ(d) p (s)ds. For the endpoints: wrh...

    work page 2048