Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Learning Stochastic Hamiltonian Systems via Stochastic Generating Function Neural Network

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An autoencoder that learns the stochastic generating function of a noisy Hamiltonian system can imitate the system's symplectic flow map.

desk verdict New architecture that learns a stochastic generating function via an autoencoder; the Gaussian-latent assumption is unproven and the validation is partly circular, but the empirical gains are consistent and the idea merits serious review. read the letter →

arxiv 2507.14467 v1 pith:TYFTNG37 submitted 2025-07-19 math.DS cs.LG

Chen Chen , Lijin Wang , Yanzhao Cao , Xupeng Cheng This is my paper
classification math.DScs.LG MSC 65P1037M1565C3068T07
keywords stochasticHamiltoniansystemsgeneratingfunctionsymplecticstructureautoencoderstructure-preservingneuralnetworkflowmaplearninglong-termprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes the stochastic generating function neural network (SGFNN), an autoencoder that learns the one-step solution map of an unknown stochastic Hamiltonian system from paired observations. The encoder extracts a latent random variable z that is meant to be a standard Gaussian independent of the current state, and the decoder reconstructs the stochastic generating function S(p1,q0,z). Because any map generated by such a function is symplectic almost surely, the learned predictor preserves phase-space geometry by construction. Across four test systems, including additive and multiplicative noise, separable and non-separable Hamiltonians, and single or multiple noise sources, the paper reports that SGFNN outperforms stochastic flow map learning in mean, standard deviation, long-term density, and structural invariants.

What carries the argument

The load-bearing object is the stochastic generating function of the first kind, a scalar random function S(P,q,ω) with the property that the map (p,q) → (P,Q) defined implicitly by p = P + ∂S/∂q^T and Q = q + ∂S/∂P^T is symplectic almost surely. The autoencoder architecture is the second piece: the encoder maps each observed pair (x0,x1) to a latent z, constrained by a KDE-based distribution loss to be standard Gaussian and independent of x0, while the decoder outputs Sθ(p1,q0,z). The final prediction is obtained by fixed-point iteration on the implicit relation for p1, then explicit update for q1, so the learned one-step map is symplectic by construction.

What would settle it

Train SGFNN on transitions generated by a stochastic Hamiltonian system with a deliberately non-Gaussian jump distribution, such as one driven by Poisson jumps rather than Brownian noise, and compare the predicted one-step density to the true transition density; if the encoder cannot map the pairs to independent standard Gaussians, the mismatch will appear directly in the density and in long-term moments. A simpler check is to compute the mutual information between the encoder's z and x0 on held-out pairs, which the method assumes to be zero.

Watch

Extended reading notes

Core claim

The central claim is that the random transition of a stochastic Hamiltonian system can be learned as an implicit symplectic map by fitting a generating function S(p1,q0,z), with the randomness localized in a low-dimensional standard Gaussian z recovered from adjacent data pairs. Given a trained S, prediction uses the implicit relations p0 = p1 + ∂S/∂q0^T and q1 = q0 + ∂S/∂p1^T, solved by fixed-point iteration with fresh Gaussian samples; this map is symplectic almost surely by Theorem 3.1. The paper's numerical experiments indicate that this structure-preserving approach yields smaller terminal errors in mean and standard deviation than sFML, correct phase-space geometry (circular orbits for the Kubo oscillator), and preservation of the theoretical linear growth of the second moment for the linear stochastic oscillator.

Load-bearing premise

The load-bearing premise is that the randomness in each one-step transition can be fully captured by a fixed number of standard Gaussian random variables that are independent of the current state and that the encoder can recover them from the observed pair; for many stochastic Hamiltonian systems the jump distribution is not Gaussian, and then the learned generating function is only an approximation.

Editorial extensions

If this is right

  • Long-term trajectory statistics follow the true system's geometry: phase-space volumes stay intact, the Kubo oscillator's circular invariant is preserved, and the linear stochastic oscillator's second moment grows linearly.
  • The predicted probability density of the solution is closer to the true density than sFML at the tested horizons, not just at the one-step level.
  • The same trained decoder can generate many trajectories with fresh Gaussian samples, giving a generative model of the stochastic flow.
  • The method handles additive and multiplicative noise, separable and non-separable Hamiltonians, and multiple noise sources without knowing the drift or diffusion functions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The method's success depends on the transition law being representable by a finite-dimensional standard Gaussian latent variable; for systems with non-Gaussian or path-dependent noise the encoder can only approximate the true randomness, and the predicted densities will inherit a bias.
  • A natural stress test is to apply SGFNN to a system driven by a non-Gaussian Lévy process or to a hypoelliptic stochastic Hamiltonian system where the one-step map is not a diffeomorphism; failure there would delimit the class of systems the architecture can learn.
  • Because the generating function is learned locally per time step, the approach could be extended to time-inhomogeneous or control-driven stochastic Hamiltonian systems by conditioning the generating function on time or control inputs.
  • The decoder defines an implicit map, so implicit differentiation could replace fixed-point iteration to yield faster and more accurate gradients for training on longer trajectories.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper proposes SGFNN, an autoencoder-based method to learn a stochastic generating function S(p1,q0,z) of an unknown stochastic Hamiltonian system from paired observations (x0,x1). The encoder maps each pair to a latent variable z, which the loss constrains to be standard Gaussian and independent of x0; the decoder outputs S, and the relation (3.1) defines an implicit symplectic one-step map. Prediction iterates (3.9)-(3.10) with fresh Gaussian draws. The method is tested on four stochastic Hamiltonian systems (linear oscillator, Kubo oscillator, non-separable SHS, synchrotron oscillations) and compared with sFML on mean, standard deviation, density histograms, and long-term moment growth. The abstract claims exact symplectic structure preservation and higher accuracy, especially in long-term predictions.

Significance. Learning structure-preserving surrogates for stochastic Hamiltonian systems is an open and worthwhile problem, and the paper's idea of learning the stochastic generating function directly, rather than drift and diffusion Hamiltonians, is a natural and potentially powerful extension of deterministic GFNN. The construction of the decoder from Theorem 3.1 is principled: it guarantees symplecticity of the implicit map provided the invertibility condition holds. The numerical study covers additive and multiplicative, separable and non-separable systems with single and multiple noises, and it consistently shows better moment and density estimates than sFML. The main weaknesses are that the finite-dimensional Gaussian latent representation is assumed without proof, the invertibility condition is not checked, and the experiments use only Delta=0.01, so the one-step transitions are near-Gaussian. If these gaps are closed, the method would be a significant contribution to data-driven structure-preserving simulation of stochastic Hamiltonian systems.

major comments (4)
  1. [§3.2.1, §3.2.3, Eq. (3.2) and (3.9)-(3.10)] The central claim that SGFNN learns the exact stochastic generating function rests on the assumption that the transition x0 to x1 can be encoded by a finite-dimensional variable z that is independent of x0 and standard Gaussian. The loss (3.3)-(3.5) forces the encoder output to be marginally Gaussian, and prediction draws omega from N(0,1) regardless of the true transition law. For a general SHS (2.1), the one-step flow depends on the full Brownian path on [t0, t0+Delta], and the transition density is typically non-Gaussian; Theorem 3.1 only guarantees that maps generated by a stochastic generating function S(P,q,omega) are symplectic, not that S can be parametrized by a finite-dimensional standard Gaussian z. The paper offers no identifiability or approximation argument for this reduction, so the trained decoder is at best a symplectic approximation of the true map. Figures 5, 9, 12, and 15 confirm the training constraint rather than a property of the data, and the experiments at Delta=0.01 produce near-Gaussian one-step transitions, so they do not test the assumption.
  2. [§3.1, Theorem 3.1, and §3.2.3] The invertibility condition required by Theorem 3.1 is never checked. Prediction uses the implicit relations (3.9)-(3.10) with a fixed-point iteration, but the theorem requires partial^2(P^T q + S(P,q,omega))/(partial P partial q) to be invertible almost surely for the map to be well defined and symplectic. The manuscript does not report any diagnostic of this Jacobian, nor does it discuss convergence or uniqueness of the fixed-point iteration for the trained S_theta2. If the condition fails, the predicted map may not be the symplectic map generated by S, and the claimed structure preservation is not guaranteed.
  3. [§3.2.1, encoder part] The independence of z from x0 is asserted but not demonstrated. The encoder is a deterministic function of the observed pair (x0,x1), and since x1 depends on x0 and the noise, the learned z may inherit dependence on x0. The sub-sampling strategy described in §3.2.1 only groups nearby initial points and does not test or guarantee global independence. The paper should either prove a statistical independence condition for the encoder output or add a diagnostic, such as comparing conditional distributions of z across different x0 regions.
  4. [§4, Figures 4-15, Table 1] The empirical comparison is not yet sufficient to support the accuracy claims. Table 1 reports single point estimates without error bars, repeated runs, or sensitivity to the unstated hyperparameters lambda, tau, nz, and KDE bandwidth in Eqs. (3.3)-(3.5). All experiments use Delta=0.01, so the one-step transitions are close to Gaussian; a test with larger Delta or with a system having a strongly non-Gaussian one-step law would be needed to distinguish exact learning from near-Gaussian approximation. The density comparisons are visual only; quantitative metrics such as KL divergence or Wasserstein distance would strengthen the claim of higher accuracy across various prediction metrics.
minor comments (6)
  1. [Algorithm 1, Eq. (3.13)] Equation (3.13) in Algorithm 1 uses i=1,...,M inside a batch, while Eqs. (3.6) and the batch size K suggest i=1,...,K; the notation is inconsistent.
  2. [§4, Eqs. (3.3)-(3.5)] The hyperparameters lambda, tau, nz, and the KDE bandwidth are not specified; without these values the experiments cannot be reproduced.
  3. [Table 1] The linear-oscillator row appears to contain both sigma=0.1 and T=50 in the T column; please separate parameter columns from the prediction time.
  4. [AMS subject classifications] The listed AMS subject classifications (52B10, 65D18, 68U05, 68U07) appear to be from computational geometry; more relevant classes such as 65P10, 37M15, and 60H35 would be appropriate.
  5. [§4, Figures 10-15] Some figure captions do not state the parameter values (e.g., the non-separable and synchrotron examples), making it harder to assess the experimental settings.
  6. [§5, Conclusion] The conclusion states that the model 'possesses the capacity of maintaining the system's symplectic structure,' but since symplecticity is enforced by the generating-function architecture, this should be described as a structural property of the construction rather than an empirical achievement.

Circularity Check

2 steps flagged · score 4.0 of 10

Accuracy vs sFML and ground truth is external and non-circular; however, two headline claims validate constraints imposed by construction: z is forced standard-Gaussian by LD and then 'validated' as 'simulating the latent stochasticity,' and 'symplectic predictions' follow for any learned S from Theorem 3.1 plus Eqs. (3.9)-(3.10), making the abstract's lead claim definitional.

  1. fitted input called prediction [Section 3.2.1 (Eqs. 3.3-3.5) and Section 4.3, Figure 12 caption; also Figures 5, 9, 15.]
    "Figure 12 illustrates the distribution of the latent variable z extracted by the encoder, which well resembles the standard normal distribution as desired, showing effectiveness of the SGFNN model in simulating the latent stochasticity even for complex systems."

    The Gaussianity displayed is the explicit training objective of LD (Eqs. 3.3-3.5), which forces the encoder output z to be standard Gaussian. The figure therefore confirms only that the constraint imposed by the model's own loss was satisfied ('as desired'), yet the caption and Section 4.3 read this as evidence of 'effectiveness ...

  2. self definitional [Abstract; Section 3.1 (Theorem 3.1 and Eq. 3.1); Section 3.2.3 (Eqs. 3.9-3.10).]
    "SGFNN preserves symplectic structure of the underlying stochastic Hamiltonian system and produces symplectic predictions."

    Theorem 3.1 states that any smooth random map S(P,q,omega) generates a symplectic map via (3.1) whenever the invertibility condition holds. The prediction map is defined by (3.9)-(3.10), which is exactly (3.1) with the learned S_theta2; hence 'produces symplectic predictions' is true for every trained network, regardless of what was learned from data. The abstract's lead claim that SGFNN 'preserves symplectic structure ... and produces symplectic predictions' therefore reduces by definition to the generating-function ansatz: the property is structurally guaranteed rather than detected or empirically demonstrated.

full rationale

The paper's quantitative core is validated against the true SHS and is not circular: predicted means, standard deviations, pdfs, and long-time behavior are compared with ground truth (Figures 2-4, 6-8, 10-11, 13-14; Table 1), and the invariant and second-moment checks (Figures 3 and 7) test whether the learned generating function reproduces the true dynamics, which is not implied by mere symplecticity of (3.1). The sFML baseline provides an independent external benchmark, so the accuracy claims are not forced by construction. The circularity burden lies in two by-construction confirmations presented as validation, as detailed in the steps. First, the standard-Gaussian latent z is imposed by the distribution loss LD (3.3)-(3.5), yet Figures 5, 9, 12, and 15 display the resulting Gaussian histograms as evidence that the model effectively simulates the latent stochasticity, and the abstract credits the encoder with identifying the randomness; the validated property is the imposed objective, not a property established about the data, and the load-bearing assumption that general nonlinear SHS transitions admit a finite-dimensional standard-Gaussian latent is never independently checked at test time, where fresh N(0,1) draws replace the encoder. Second, the headline claim that SGFNN produces symplectic predictions holds by construction, because Theorem 3.1 (quoted from the independent reference [10]) makes any map defined by (3.1) symplectic, and (3.9)-(3.10) are exactly (3.1); no learning is needed to obtain the guarantee. Self-citations ([8], [41]-[43]) are present but not load-bearing: the central theorem is attributed to [10], and the subsampling and distributional constraint devices are attributed to [38], so the score is not raised on self-citation grounds. Because the main quantitative results stand on independent external ground, the circularity is partial, warranting a moderate score of 4 rather than a higher one.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central method rests on four categories of assumptions: the cited generating-function theorem, the local existence of S for the true flow, the ad hoc Gaussian latent representation, and the accuracy of the data-generation and prediction pipeline. The first two are standard or domain assumptions; the Gaussian latent representation is the main unproven addition; the network weights and hyperparameters are fitted quantities.

free parameters (5)
  • λ (loss balance) = not reported
    Weight balancing LMSE and LD in Eq. (3.8); chosen by hand, value omitted, affects the trained generating function.
  • τ (distribution loss balance) = not reported
    Weight in Eq. (3.3); value omitted, affects how strongly z is forced to be Gaussian.
  • nz (latent dimension) = 1 for first three examples; 2 for synchrotron
    Chosen to match the number of noises; for unknown systems this is a modeling choice with no stated selection rule.
  • encoder/decoder weights θ1, θ2 = learned via Adam
    Neural network parameters fitted to data; identifiability, overfitting, and sensitivity are not analyzed.
  • KDE bandwidth = not reported
    Kernel density estimation in Eq. (3.4) requires bandwidth selection; the value is omitted and can affect training.
assumptions (5)
  • standard math Any smooth random map generated by a stochastic generating function via Eq. (3.1) is symplectic almost surely (Theorem 3.1, from ref 10).
    Background theorem cited from ref 10; accepted as standard in the stochastic symplectic integration literature.
  • domain assumption The true one-step solution flow of the SHS admits a local stochastic generating function of the first kind over the sampled domain.
    Ensures the target S exists; local existence follows from refs 8 and 10 but is not stated or verified for the test systems.
  • ad hoc to paper The latent variable z is independent of x0 and follows a unit Gaussian distribution, and one or two Gaussian variables fully describe the transition randomness.
    Imposed by loss and subsampling heuristics in Section 3.2.1; not derived from SDE theory and false for general non-Gaussian transitions.
  • domain assumption Observational pairs are exact one-step flow samples of an SHS with time step Δ, with no observation noise and with equal-length trajectories.
    Data are generated by the midpoint method; the model assumes these represent true flow values at the sampled points.
  • domain assumption Fixed-point iteration in Eq. (3.9) converges to the unique solution of the implicit generating-function map within tolerance 1e-12.
    Required at prediction time; no convergence proof or failure cases are reported for the implicit solve.
invented entities (1)
  • Latent Gaussian variable z (and z=(z1,z2))
    purpose: Encodes the unobserved noise driving each transition so the decoder can define the stochastic generating function.
    The paper postulates z ~ N(0,1) independent of x0 and uses it to generate predictions; no independent falsifiable handle is given, and the map from observed x0,x1 to z is not shown to be identifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Stochastic Hamiltonian Systems via Stochastic Generating Function Neural Network." pith.science (2026). https://pith.science/paper/TYFTNG37

@misc{pith2026250714467,
  author       = {Pith},
  title        = {Pith review of: Learning Stochastic Hamiltonian Systems via Stochastic Generating Function Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYFTNG37}},
  note         = {Machine review of arXiv:2507.14467}
}
read the original abstract

In this paper we propose a novel neural network model for learning stochastic Hamiltonian systems (SHSs) from observational data, termed the stochastic generating function neural network (SGFNN). SGFNN preserves symplectic structure of the underlying stochastic Hamiltonian system and produces symplectic predictions. Our model utilizes the autoencoder framework to identify the randomness of the latent system by the encoder network, and detects the stochastic generating function of the system through the decoder network based on the random variables extracted from the encoder. Symplectic predictions can then be generated by the stochastic generating function. Numerical experiments are performed on several stochastic Hamiltonian systems, varying from additive to multiplicative, and from separable to non-separable SHSs with single or multiple noises. Compared with the benchmark stochastic flow map learning (sFML) neural network, our SGFNN model exhibits higher accuracy across various prediction metrics, especially in long-term predictions, with the property of maintaining the symplectic structure of the underlying SHSs.

Figures

Figures reproduced from arXiv: 2507.14467 by the authors.

Figure 1
Figure 1. An illustration of the algorithm flow and loss function of the SGFNN Network [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Linear Oscillator: comparison of the mean and standard deviation of 10,000 [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Linear Oscillator: growth of the second moment of the solution predicted by the [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Linear Oscillator: histogram of the probability density function (pdf) of the [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Linear Oscillator: histogram of the pdf of the latent variable [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Kubo Oscillator: comparison of the mean and standard deviation of 10,000 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Kubo Oscillator: predicted (p(t), q(t)) on [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Kubo Oscillator: histogram of the pdf of predicted and true solutions at [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Kubo Oscillator: histogram of the pdf of the latent variable [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Non-Separable SHS: comparison of the mean and standard deviation of 10,000 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Non-Separable SHS: histogram of the pdf of predicted and true solutions at [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Non-Separable SHS: histogram of the pdf of the latent variable [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Synchrotron Oscillations: comparison of the mean and standard deviation of [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Synchrotron Oscillations: histogram of the pdf of predicted and true solutions [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Synchrotron Oscillations: histogram of the pdf of the latent variables [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [1]

    Symplectic geometric algorithms for Hamiltonian systems[M]

    Feng K, Qin M. Symplectic geometric algorithms for Hamiltonian systems[M]. Zhejiang Pub- lishing United Group, 2003

    work page 2003

  2. [2]

    Structure-Preserving Algorithms for Ordinary Differential Equations[M]

    Hairer E , Lubich C , Wanner G .Geometric Numerical Integration. Structure-Preserving Algorithms for Ordinary Differential Equations[M]. Springer, 2006

    work page 2006

  3. [3]

    M´ ecanique al´ eatoire

    Bismut, J.M. M´ ecanique al´ eatoire. In Ecole d’Et´ e de Probabilit´ es de Saint-Flour X—1980. Lecture Notes in Mathematics; Hennequin, P.L., Ed.; Springer: Berlin/Heidelberg, Germany, 1982; Volume 929

    work page 1980

  4. [4]

    Conserved quantities and symmetries related to stochastic dynamical systems[J]

    Misawa T. Conserved quantities and symmetries related to stochastic dynamical systems[J]. Annals of the Institute of Statistical Mathematics, 1999, 51: 779-802

    work page 1999

  5. [5]

    Mean-square symplectic methods for Hamiltonian systems with multiplicative noise

    Milstein G N, Repin Y M, Tretyakov M V. Mean-square symplectic methods for Hamiltonian systems with multiplicative noise. Preprint: Weierstraß-Institut f¨ ur Angewandte Analysis und Stochastik. https://doi.org/10.20347/WIAS.PREPRINT.670. 2001

    work page doi:10.20347/wias.preprint.670

  6. [6]

    Symplectic methods for Hamiltonian systems with additive noise[J]

    Milstein G N, Repin Y M, Tretyakov M V. Symplectic methods for Hamiltonian systems with additive noise[J]. SIAM Journal on Numerical Analysis,2002,39(6): 2066-2088

    work page 2002

  7. [7]

    Predictor–corrector methods for a linear stochastic oscillator with additive noise[J]

    Hong J, Scherer R, Wang L. Predictor–corrector methods for a linear stochastic oscillator with additive noise[J]. Mathematical and computer modelling, 2007, 46(5-6): 738-764

    work page 2007

  8. [8]

    Variational integrators and generating functions for stochastic Hamiltonian sys- tems[D]

    Wang L. Variational integrators and generating functions for stochastic Hamiltonian sys- tems[D]. Karlsruhe, Univ, Diss, 2007

    work page 2007

Show all 46 references
  1. [9]

    Stochastic variational integrators[J]

    Bou-Rabee N, Owhadi H. Stochastic variational integrators[J]. IMA Journal of Numerical Analysis, 2009, 29(2): 421-443

    2009

  2. [10]

    High-order symplectic schemes for stochastic Hamiltonian systems[J]

    Deng J, Anton C, Wong Y S. High-order symplectic schemes for stochastic Hamiltonian systems[J]. Communications in Computational Physics, 2014, 16(1): 169-200

    2014

  3. [11]

    Symplectic integration of stochastic Hamiltonian systems[M]

    Hong J, Sun L. Symplectic integration of stochastic Hamiltonian systems[M]. Springer, 2022

    2022

  4. [12]

    Neural ordinary differential equations[J]

    Chen R T Q, Rubanova Y, Bettencourt J, et al. Neural ordinary differential equations[J]. In Proceedings of the Advances in Neural Information Processing Systems, Montr´ eal, QC, Canada, 3–8 December 2018: 6571–6583

    2018

  5. [13]

    Symplectic ode-net: Learning hamiltonian dynamics with control

    Zhong Y D, Dey B, Chakraborty A. Symplectic ode-net: Learning hamiltonian dynamics with control. arXiv preprint arXiv:1909.12077, 2019. 24

    1909 arXiv

  6. [14]

    Pde-net: Learning pdes from data[J]

    Long Z, Lu Y, Ma X, et al. Pde-net: Learning pdes from data[J]. International conference on machine learning. PMLR, 2018: 3208-3216

    2018

  7. [15]

    PDE-Net 2.0: Learning PDEs from data with a numeric-symbolic hybrid deep network[J]

    Long Z, Lu Y, Dong B. PDE-Net 2.0: Learning PDEs from data with a numeric-symbolic hybrid deep network[J]. Journal of Computational Physics, 2019, 399: 108925

    2019

  8. [16]

    Sde-net: Equipping deep neural networks with uncertainty estimates

    Kong L, Sun J, Zhang C. Sde-net: Equipping deep neural networks with uncertainty estimates. arXiv preprint arXiv:2008.10546, 2020

    2008 arXiv

  9. [17]

    Neural SDE: Stabilizing Neural ODE Networks with Stochastic Noise

    Liu X , Xiao T , Si S ,et al. Neural SDE: Stabilizing Neural ODE Networks with Stochastic Noise. DOI:10.48550/arXiv.1906.02355, 2019

    1906 doi arXiv

  10. [18]

    Hamiltonian neural networks[J]

    Greydanus S, Dzamba M, Yosinski J. Hamiltonian neural networks[J]. In Proceedings of the Conference on Neural Information Processing Systems, Vancouver, BC, Canada, December 2019: 8-14

    2019

  11. [19]

    Symplectic recurrent neural networks[J]

    Chen Z, Zhang J, Arjovsky M, et al. Symplectic recurrent neural networks[J]. In Proceedings of the International Conference on Learning Representations, New Orleans, LA, USA, May 2019: 6-9

    2019

  12. [20]

    Deep Hamiltonian networks based on symplectic integrators

    Zhu A, Jin P, Tang Y. Deep Hamiltonian networks based on symplectic integrators. arXiv preprint arXiv:2004.13830, 2020

    2004 arXiv

  13. [21]

    SympNets: Intrinsic structure-preserving symplectic networks for identifying Hamiltonian systems[J]

    Jin P, Zhang Z, Zhu A, Tang Y, et al. SympNets: Intrinsic structure-preserving symplectic networks for identifying Hamiltonian systems[J]. Neural Networks, 2020, 132: 166-179

    2020

  14. [22]

    Journal of Computational Physics, 2021, 437(4):110325

    Tong Y , Xiong S , He X ,et al.Symplectic neural networks in Taylor series form for Hamil- tonian systems[J]. Journal of Computational Physics, 2021, 437(4):110325

    2021

  15. [23]

    Hamiltonian generative networks

    Toth P, Rezende D J, Jaegle A, et al. Hamiltonian generative networks. arXiv preprint arXiv:1909.13789, 2019

    1909 arXiv

  16. [24]

    Data-driven prediction of general Hamiltonian dynamics via learning exactly- symplectic maps[C]

    Chen R, Tao M. Data-driven prediction of general Hamiltonian dynamics via learning exactly- symplectic maps[C]. International Conference on Machine Learning. PMLR, 2021: 1717-1727

    2021

  17. [25]

    Nonseparable symplectic neural networks

    Xiong S, Tong Y, He X, et al. Nonseparable symplectic neural networks. arxiv preprint arxiv:2010.12636, 2020

    2010 arXiv

  18. [26]

    Structure-preserving method for reconstructing unknown Hamiltonian systems from trajectory data[J]

    Wu K, Qin T, Xiu D. Structure-preserving method for reconstructing unknown Hamiltonian systems from trajectory data[J]. SIAM Journal on Scientific Computing, 2020, 42(6): A3704- A3729

    2020

  19. [27]

    Learning Hamiltonian systems considering system symmetries in neural networks[J]

    Dierkes E, Flaßkamp K. Learning Hamiltonian systems considering system symmetries in neural networks[J]. IF AC-PapersOnLine, 2021, 54(19): 210-216

    2021

  20. [28]

    Dietrich F , Makeev A , Kevrekidis G ,et al.Learning effective stochastic differential equa- tions from microscopic simulations: Linking stochastic numerics to deep learning[J].Chaos, 2023.DOI:10.1063/5.0113632

    2023 doi

  21. [29]

    Detecting stochastic governing laws with observation on stationary distributions[J]

    Chen X, Wang H, Duan J. Detecting stochastic governing laws with observation on stationary distributions[J]. Physica D: Nonlinear Phenomena, 2023, 448: 133691

    2023

  22. [30]

    Scalable inference in sdes by direct matching of the fokker–planck–kolmogorov equation[J]

    Solin A, Tamir E, Verma P. Scalable inference in sdes by direct matching of the fokker–planck–kolmogorov equation[J]. Advances in Neural Information Processing Systems, 2021, 34: 417-429

    2021

  23. [31]

    Variational inference for stochastic differential equations[J]

    Opper M. Variational inference for stochastic differential equations[J]. Annalen der Physik, 2019, 531(3): 1800233

    2019

  24. [32]

    Black-box variational inference for stochastic differential equations[C]

    Ryder T, Golightly A, McGough A S, et al. Black-box variational inference for stochastic differential equations[C]. International Conference on Machine Learning. PMLR, 2018: 4423- 4432

    2018

  25. [33]

    Modeling continuous stochastic processes with dynamic normalizing flows[J]

    Deng R, Chang B, Brubaker M A, et al. Modeling continuous stochastic processes with dynamic normalizing flows[J]. Advances in neural information processing systems, 2020, 33: 7805-7815. 25

    2020

  26. [34]

    Normalizing field flows: Solving forward and inverse stochastic dif- ferential equations using physics-informed flow models[J]

    Guo L, Wu H, Zhou T. Normalizing field flows: Solving forward and inverse stochastic dif- ferential equations using physics-informed flow models[J]. Journal of Computational Physics, 2022, 461: 111202

    2022

  27. [35]

    Data-driven Effective Modeling of Multiscale Stochastic Dynamical Systems

    Chen Y, Xiu D. Data-driven Effective Modeling of Multiscale Stochastic Dynamical Systems. arXiv preprint arXiv:2408.14821, 2024

    2024 arXiv

  28. [36]

    Learning stochastic dynamical system via flow map operator[J]

    Chen Y, Xiu D. Learning stochastic dynamical system via flow map operator[J]. Journal of Computational Physics, 2024, 508: 112984

    2024

  29. [37]

    Modeling Unknown Stochastic Dynamical System Subject to External Exci- tation

    Chen Y, Xiu D. Modeling Unknown Stochastic Dynamical System Subject to External Exci- tation. arXiv preprint arXiv:2406.15747, 2024

    2024 arXiv

  30. [38]

    Modeling unknown stochastic dynamical system via autoen- coder[J]

    Xu Z, Chen Y, Chen Q, et al. Modeling unknown stochastic dynamical system via autoen- coder[J]. Journal of Machine Learning for Modeling and Computing, 2024, 5(3): 87-112

    2024

  31. [39]

    A Training-Free Conditional Diffusion Model for Learning Stochastic Dynamical Systems

    Liu Y, Chen Y, Xiu D, et al. A Training-Free Conditional Diffusion Model for Learning Stochastic Dynamical Systems. arXiv preprint arXiv:2410.03108, 2024

    2024 arXiv

  32. [40]

    Diffusion-Model-Assisted Supervised Learning of Generative Models for Density Estimation[J]

    Liu Y, Yang M, Zhang Z, et al. Diffusion-Model-Assisted Supervised Learning of Generative Models for Density Estimation[J]. Journal of Machine Learning for Modeling and Computing, 2024, 5(1): 25-38

    2024

  33. [41]

    Learning a class of stochastic differential equations via numerics- informed Bayesian denoising[J]

    Wang Z, Wang L, Cao Y. Learning a class of stochastic differential equations via numerics- informed Bayesian denoising[J]. International Journal for Uncertainty Quantification, 2025, 15(3): 0-20

    2025

  34. [42]

    Learning Parameters of a Class of Stochastic Lotka-Volterra Systems with Neural Networks[J]

    Wang Z, Wang L. Learning Parameters of a Class of Stochastic Lotka-Volterra Systems with Neural Networks[J]. Journal of University of Chinese Academy of Sciences, 2025, 42(1): 20-25

    2025

  35. [43]

    Quadrature Based Neural Network Learning of Stochastic Hamil- tonian Systems[J]

    Cheng X, Wang L, Cao Y. Quadrature Based Neural Network Learning of Stochastic Hamil- tonian Systems[J]. Mathematics, 2024, 12(16): 2438

    2024

  36. [44]

    Journal of Computational Physics, 2023, 494: 112495

    David M, M´ ehats F, Symplectic Learning for Hamiltonian Neural Networks[J]. Journal of Computational Physics, 2023, 494: 112495

    2023

  37. [45]

    Numerical methods for stochastic systems preserv- ing symplectic structure[J]

    Milstein G N, Repin Y M, Tretyakov M V. Numerical methods for stochastic systems preserv- ing symplectic structure[J]. SIAM Journal on Numerical Analysis, 2002, 40(4): 1583-1604

    2002

  38. [46]

    Numerical simulation of a linear stochastic oscillator with additive noise, Appl.Numer.Math

    Stømmen A H, Higham D J. Numerical simulation of a linear stochastic oscillator with additive noise, Appl.Numer.Math. 2004, 51(1): 89–99

    2004

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.