Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Moment Estimate and Variational Approach for Learning Generalized Diffusion with Non-gradient Structures

T0 review · 3 major / 5 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 The paper claims that a two-stage weak-form scheme based on first-moment evolution and energy dissipation recovers the drift, pseudo-potential, and rotation of generalized diffusions with pointwise orthogonal non-gradient structures.

desk verdict A genuinely useful two-stage weak-form method, but the advertised non-gradient class is wider than what the energy-law step actually supports: div(R)=0 is missing from the assumptions. read the letter →

arxiv 2508.01854 v2 pith:MJFOMHPT submitted 2025-08-03 physics.comp-ph cs.LGmath.APnlin.AO

Fanze Kong , Chen-Chih Lai , Yubin Lu This is my paper · ORCID
classification physics.comp-phcs.LGmath.APnlin.AO MSC 60H1060J6065C30 PACS 05.10.Gg02.60.Cb
keywords generalizeddiffusionnon-gradientdriftpseudo-potentialrotationtermFokker-Planckequationenergydissipationlawfirst-momentevolutionweak-formlearning
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

The paper sets out to identify the drift, the pseudo-potential, and the rotational component of a generalized diffusion from snapshots of its probability density, without solving or differentiating the Fokker-Planck equation. The target class is the drift decomposition $b=-\tfrac{1}{2}\sigma^2\nabla\psi+\tfrac{1}{2}\sigma^2 R$ with the pointwise orthogonality condition $\nabla\psi\cdot R=0$, so that the gradient part controls steady states and the rotation controls how the system approaches them. If the two-stage method works as claimed, a wider family of non-equilibrium stochastic systems becomes learnable by weak-form losses, which the paper shows to be more robust to noise and low-regularity data than strong-form PDE learning. The numerical experiments in dimension two support the claim by recovering fields in double-well, quadruple-well, rough-potential, and non-canonical-rotation cases.

What carries the argument

The machinery is the pair of integral identities that replace the PDE. The first is the first-moment evolution $\frac{d}{dt}\int f x_i\,dx=\int b_i f\,dx$, which is used in Stage 1 to pin down the drift because the loss is convex in the drift network. The second is the free-energy dissipation law $\frac{dF}{dt}=-\int \frac{2f}{\sigma^2}|u|^2\,dx$, which holds under the pointwise orthogonality condition and supplies the residual for Stage 2. The third piece is the weighted penalty (3.4), derived from dimensional analysis, which makes the orthogonality constraint compete with the energy loss on the same physical scale; without that weighting, the gradient and rotational components are not separated by the optimization.

What would settle it

Construct a Fokker-Planck problem with a rotation field that satisfies $\nabla\psi\cdot R=0$ but has nonzero divergence, generate exact density snapshots, and run the Stage-2 loss; if the learned potential deviates from the true one while Stage-1 drift recovery stays accurate, the energy-law premise is the point of failure.

Watch

Extended reading notes

Core claim

The central claim is that the decomposition of the drift into a dissipative gradient part and a conservative rotation part can be learned in two stages from four density snapshots per initial condition. Stage 1 minimizes loss (3.2), which encodes the first-moment evolution $\frac{d}{dt}\int f x_i\,dx = \int b_i f\,dx$; this loss is convex in the drift network and fixes $b$. Stage 2 minimizes the free-energy dissipation residual (3.3), which is derived from the identity $\frac{dF}{dt}=-\int \frac{2f}{\sigma^2}|u|^2\,dx$ that holds under the orthogonality condition, combined with the weighted penalty (3.4) enforcing $\nabla\psi\cdot R=0$ at the level of the learned fields. The rotation is then not learned independently but reconstructed as $R=\frac{2}{\sigma^2}b+\nabla\psi$. The reported experiments give relative root-mean-square errors around $10^{-2}$ for the baseline double-well case and show that the two-stage method outperforms direct joint optimization of $\psi$ and $R$.

Load-bearing premise

Stage 2 treats the energy law as exact, but that law is derived only for rotations with zero divergence, while the paper states only the pointwise orthogonality condition and never verifies divergence-freeness for the target class.

Editorial extensions

If this is right

  • If the central claim holds, non-gradient generalized diffusions can be learned from density snapshots alone, with no need to solve the Fokker-Planck equation or form its strong residual.
  • The rotation is obtained for free from the relation $R=\frac{2}{\sigma^2}b+\nabla\psi$ once the first two stages succeed, so no separate parametrization for the rotational field is required.
  • The weighted penalty is what makes the gradient-rotation separation possible; the paper's ablation shows that replacing it with a standard $L^2$ penalty leaves the potential almost unrecovered while the drift remains accurate.
  • The method's integral-form losses justify its robustness to noise levels up to $\gamma=0.3$ in the reported experiments and its ability to handle rough pseudo-potentials without differentiating high-order derivatives.
  • Longer Stage-2 observation times reduce the non-convexity of the energy loss and improve potential recovery, while Stage-1 times must stay short enough for the centroid dynamics to identify the drift.

Reading between the lines

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

  • The authors leave implicit that the divergence-free condition on $R$, needed for (2.8), may be required for all their examples; a systematic scan over rotations with $\nabla\cdot R\neq0$ would delimit the true class of admissible systems.
  • A testable extension is to replace the density-snapshot training data with Monte Carlo estimates of first moments and energy from sampled trajectories, which would let the same two-stage losses operate in higher dimensions where Fokker-Planck solvers are unavailable.
  • The dimensional-analysis principle behind penalty (2.9) is a general device: any multi-term physics loss whose terms carry different units can be reweighted by dimensional consistency, so the idea may transfer to other constrained inverse problems beyond diffusion learning.
  • Because Stage 1 is convex in the drift, one could feed the Stage-2 error back into the drift estimate for a refinement loop, though the paper presents the stages as strictly sequential.
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

3 major / 5 minor

Summary. The paper proposes a two-stage, weak-form learning framework for identifying the drift, pseudo-potential, and rotational component of generalized diffusion processes of the form dX_t = b(X_t)dt + σ(X_t)dW_t with b = -1/2 σ² ∇ψ + 1/2 σ² R. In Stage 1, the drift b is learned from the first-moment evolution (Eq. (2.3)) via the loss (3.2). In Stage 2, the pseudo-potential ψ is learned from the energy dissipation law (2.8) via the loss (3.3), supplemented by a pointwise orthogonality penalty (3.4) derived from dimensional analysis; the rotation is then reconstructed as R = 2b/σ² + ∇ψ. Numerical experiments cover canonical and non-canonical rotations, quadruple-well and rough potentials, noisy data, and ablations against direct methods and an L2 penalty. The central claim is that the method recovers ψ and R for drifts satisfying the pointwise orthogonality condition ∇ψ·R=0.

Significance. If the method is restricted to the class for which the energy law is actually valid, the paper offers a useful weak-form alternative to strong-form PDE-based quasi-potential learning: the losses are integral forms, so they avoid high-order derivatives and show robustness to noisy data, as demonstrated by the noise-level table. The two-stage idea of learning the drift from short-time moment data and the potential from intermediate-time energy data is natural and is supported by the ablation study showing that the two-stage method substantially outperforms the two direct methods. The inclusion of rough potentials, non-canonical rotations, and comparisons with different penalty forms strengthens the empirical contribution. The main caveats are that the advertised class of drifts is broader than the class actually treated, and that the Stage-2 loss as written does not match the stated energy law.

major comments (3)
  1. [Section 2, Eq. (2.8)] The claim that (2.8) follows from (2.6) under the pointwise orthogonality condition (2.7) is incomplete. Integrating by parts in the second term of (2.6) gives ∫ [R·∇(σ²f)+σ²f R·∇ψ] dx = ∫∇·(σ²f R) dx - ∫ σ²f (∇·R) dx = -∫ σ²f (∇·R) dx when the boundary term vanishes and (2.7) is used. Therefore (2.8) requires ∇·R = 0, or at least ∫ σ² f (∇·R) dx = 0 for the family of densities used in training. The paper never states or proves this divergence-free condition. Every numerical example in Section 4 uses divergence-free rotations, and the non-canonical example explicitly verifies ∇·R=0. Thus the class of drifts for which the Stage-2 loss (3.3) is a valid energy balance is narrower than the class advertised in the abstract and in Eq. (2.7). This is load-bearing because Stage 2 is the only place where ψ is learned. Please either state and prove the divergence-free assumption explicitly, or modify the energy-loss construction to account for nonzero ∇·R.
  2. [Eq. (3.3)] The discretized dissipation term in (3.3) is inconsistent with the energy law (2.6). From the definition of u in (2.6), one has 2f/σ² |u|² = f σ²/2 |∇ln(σ²f) + ∇ψ|², so the second term in (3.3) should contain (σ²(x_i)/2) f_j(x_i,T_1), not (1/(2σ²(x_i))) f_j(x_i,T_1). As written, the loss is not the Riemann-sum discretization of the right-hand side of (2.8). For the variable-coefficient examples in Sections 4.1 and 4.2 this changes the pointwise weighting by a factor of σ⁴, and for the rough-potential example with σ²=2 it changes the weight by a factor of 4. Since (3.3) is the objective used to learn ψ, this discrepancy needs to be resolved, either by correcting the coefficient in the displayed formula or by explaining a different convention that reconciles (3.3) with (2.6).
  3. [Section 3.1, Eq. (3.2)] The remark after (3.2) states that the loss is convex in (b_k)NN and therefore the optimization problem is convex and unique. This is misleading: the loss is convex as a function of the network output values, but the optimization over the network parameters θ is generally non-convex because the network output is nonlinear in θ. The statement as written overstates the theoretical guarantee for the actual training problem. This does not affect the numerical results, but the claim should be corrected or qualified.
minor comments (5)
  1. [Section 2, Eq. (2.3)] The passage from (2.2) to (2.3) drops the boundary term -1/2∫∂_{x_i}(σ²f)dx without explicitly stating the decay or boundary conditions on f and σ²f. Please state the mild decay assumptions under which this boundary term vanishes.
  2. [Section 4.2, non-canonical rotation] The non-canonical example checks both ∇ψ·R=0 and ∇·R=0. The fact that this verification is included highlights that the second condition is needed; however, the main text of Section 2 does not mention it. Please move this condition into the formulation section so that the hypotheses match the examples.
  3. [Appendix A] The dimensional analysis treats logarithms as dimensionless and assigns [ψ]=1. This is a formal argument, and the paper would benefit from a sentence clarifying that the dimensional analysis is a heuristic used only to select the form of the penalty, not a rigorous derivation.
  4. [Table 1] Experiment 8 with λ=100 gives rRMSEψ = 5.807e-01, which is much worse than λ=10 or λ=1. This suggests that the orthogonality penalty can dominate the energy-loss term and degrade the reconstruction. A brief comment on how λ should be selected would be useful.
  5. [General] There are several typographical issues, including the title 'Moment Estimate and V ariational Approach' with an extraneous space, and the repeated use of 'diffusion' with a broken ligature. These should be corrected in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the two-stage weak-form losses are fitted to synthetic data and validated against independent ground truth; the gap between Eqs. (2.6) and (2.8) is a missing divergence-free condition, not an input-output circularity.

full rationale

The derivation chain is not circular. In Stage 1, the drift is learned by minimizing the first-moment loss (3.2), which follows directly from the Fokker-Planck equation by integration by parts; the moments are computed from simulated densities, not from the target drift. In Stage 2, the pseudo-potential is learned from the energy-dissipation loss (3.3) plus the orthogonality penalty (3.4), and the rotation is reconstructed through R = 2b/sigma^2 + grad(psi), which is an algebraic rewriting of the assumed decomposition (2.4), not a fitted-to-target surrogate. All benchmarks compare against independently specified synthetic potentials, rotations, and drifts, so the reported rRMSE values are genuine generalization errors. The main load-bearing predecessor is [47], a same-group preprint supplying the free-energy evolution (2.6); this is a standard integration-by-parts calculation and is not used to define the target quantities, so the self-citation is not circular. The principal weakness is mathematical rather than circular: (2.8) follows from (2.6) only if the divergence term involving R vanishes, e.g., if div(R)=0, which is not stated alongside the orthogonality condition (2.7); every numerical example happens to use divergence-free rotations. This gap narrows the advertised class but does not cause the predictions to reduce to fitted inputs or to self-citation chains.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The method is not a derivation of a physical constant; it is a learning scheme. Its identifiable hand-chosen constant is the penalty weight lambda, which strongly affects results. The main unstated premise is that the rotation is divergence-free, and the examples also assume access to clean Fokker-Planck density snapshots. The weighted penalty's functional form is selected by dimensional analysis, which fixes dimensions but not the function.

free parameters (1)
  • lambda (orthogonality penalty weight) = 10 in baseline; 1 and 100 in ablations
    Hand-chosen multiplier in Eq (3.5); Table 1 shows rRMSE_psi varies from 0.0193 (lambda=10) to 0.5807 (lambda=100), so the method's performance depends on this choice.
assumptions (6)
  • domain assumption The target drift has the generalized fluctuation-dissipation form b = -1/2 sigma^2 grad(psi) + 1/2 sigma^2 R.
    Section 2, Eq (2.4); this restricts the class of stochastic systems the method addresses.
  • domain assumption The potential gradient and rotation satisfy the pointwise orthogonality condition grad(psi) . R = 0.
    Eq (2.7); the energy-dissipation law and the penalty construction both depend on this condition.
  • domain assumption The rotational component is divergence-free, div(R) = 0.
    Needed to pass from Eq (2.6) to Eq (2.8) but never stated in Section 2; all numerical examples happen to satisfy it, and Example 2 verifies it explicitly.
  • domain assumption The noise intensity sigma^2 is a known, positive scalar function.
    Section 3 assumes sigma^2 is known and enters both the first-moment loss and the energy-loss integrands.
  • domain assumption Observations are clean density snapshots from the Fokker-Planck equation at four times, with boundary terms vanishing at infinity.
    Eq (2.3) uses integration by parts on R^d, and Section 4 generates data by solving the Fokker-Planck equation on a bounded computational domain.
  • ad hoc to paper Dimensional consistency is sufficient to select the penalty form.
    Appendix A shows dimension matching for (2.9), but infinitely many functionals share that dimension, so the specific choice of the weighted penalty is not uniquely derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moment Estimate and Variational Approach for Learning Generalized Diffusion with Non-gradient Structures." pith.science (2026). https://pith.science/paper/MJFOMHPT

@misc{pith2026250801854,
  author       = {Pith},
  title        = {Pith review of: Moment Estimate and Variational Approach for Learning Generalized Diffusion with Non-gradient Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJFOMHPT}},
  note         = {Machine review of arXiv:2508.01854}
}
read the original abstract

This paper proposes a data-driven learning framework for identifying governing laws of generalized diffusions with non-gradient components. By combining energy dissipation laws with a physically consistent penalty and first-moment evolution, we design a two-stage method to recover the pseudo-potential and rotation in the pointwise orthogonal decomposition of a class of non-gradient drifts in generalized diffusions. Our two-stage method is applied to complex generalized diffusion processes including dissipation-rotation dynamics, rough pseudo-potentials and noisy data. Representative numerical experiments demonstrate the effectiveness of our approach for learning physical laws in non-gradient generalized diffusions.

Figures

Figures reproduced from arXiv: 2508.01854 by the authors.

Figure 1
Figure 1. The flowchart of the two-stage method. We remark that in our two-stage learning framework for nonlinear stochastic dynamics, we consider two temporal steps: in Stage 1, the short time is chosen for training to capture the drift b, while in Stage 2, we choose the intermediate time to learn the pseudo-potential ψ. The reason is that, in Stage 1, short-time data remain distinct enough to be a family of effective test f… view at source ↗
Figure 2
Figure 2. Comparison of the learned drift (left) bNN with the ground truth (right) b(x, y) = 1 2(1+x 2+y 2 ) " −(x 3 − x) + y −y − (x 3 − x) # . The relative root mean square error is 2.266 × 10−2 . 13 [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the learned potential function ψNN with the ground truth ψ(x, y) = 1 4 (x 2 − 1)2 + 1 2 y 2 . The heatmaps, shown from left to right, correspond to ψNN, the ground truth, and their pointwise difference. The relative root mean square error is 1.929 × 10−2 . −   x −    −   x −   [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Comparison of the learned rotation (left) RNN with the ground truth (right) R(x, y) = ∇ψ ⊥ = " y −(x 3 − x) # . The relative root mean square error is 3.495 × 10−2 . −   x −    −   x −   [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the learned rotation (left) RNN with the ground truth (right) R(x, y) =  1 16 (x 2 − 1)2 + 1 8 y 2  " y −(x 3 − x) # , using M = 80. The relative root mean square error is 1.692 × 10−1 . 14 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the learned potential function ψNN with the ground truth ψ(x, y) = 1 8 (x 2 − 1)2 + 1 8 (y 2 − 1)2 . The heatmaps, shown from left to right, correspond to ψNN, the ground truth, and their pointwise difference. The relative root mean square error is 1.513 …
Figure 7
Figure 7. Figure 7: The potential function ψ(x) = 1 4 (x 2 − 1)2 + 1 2 y 2 + ε 4 sin( 2πx ε ) sin( 2πy ε ) with parameter ε = 0.4. − x − [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: The first and second partial derivatives of the potential function ψ(x) = 1 4 (x 2 − 1)2 + 1 2 y 2 + ε 4 sin( 2πx ε ) sin( 2πy ε ) with parameter ε = 0.4. Upper row: The first derivatives ∂xψ and ∂yψ. Lower row: The second derivatives ∂ 2 xψ and ∂ 2 yψ. 15 [PITH_FULL_…
Figure 9
Figure 9. Figure 9: Comparison of the learned potential function ψNN with the ground truth ψ(x, y) = 1 4 (x 2 − 1)2 + 1 2 y 2 + ε 4 sin( 2πx ε ) sin( 2πy ε ). The heatmaps, shown from left to right, correspond to ψNN, the ground truth, and their pointwise difference. The relative root mea…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Moment Estimates and DeepRitz Methods on Learning Diffusion Systems with Non-gradient Drifts

    cs.LG 2025-08 conditional novelty 4.0 of 10

    Moment-DeepRitz learns the potential and rotational parts of a non-gradient drift in diffusion systems by combining first-moment dynamics with a DeepRitz variational loss.

Reference graph

Works this paper leans on

69 extracted references · 67 canonical work pages · cited by 1 Pith paper

  1. [1]

    Batlle, Y

    P . Batlle, Y . Chen, B. Hosseini, H. Owhadi, and A. M. Stuart. Error analysis of kernel /GP methods for nonlinear and parametric PDEs. J. Comput. Phys., 520:113488, 2025

    work page 2025

  2. [2]

    Bertalan, F

    T. Bertalan, F. Dietrich, I. Mezi´ c, and I. G. Kevrekidis . On learning Hamiltonian systems from data. Chaos., 29(12):121107, 2019

    work page 2019

  3. [3]

    Brunton, J

    S. Brunton, J. Proctor, and J. Kutz. Discovering governi ng equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. , 113(15):3932–3937, 2016

    work page 2016

  4. [4]

    Chen and M

    R. Chen and M. Tao. Data-driven prediction of general Ham iltonian dynamics via learning exactly-symplectic maps. In Proceedings of the 38th International Conference on Machin e Learning, 2021

    work page 2021

  5. [5]

    X. Chen, B. Soh, Z.-E. Ooi, E. Vissol-Gaudin, H. Y u, K. Nov oselov, K. Hippalgaonkar, and Q. Li. Constructing custom thermodynamics using deep learning. Nat. Comput. Sci. , 4(1):66–85, 2024

    work page 2024

  6. [6]

    X. Chen, L. Y ang, J. Duan, and G. E. Karniadakis. Solving i nverse stochastic problems from discrete particle obser- vations using the fokker–planck equation and physics-informed neural networks. SIAM J. Sci. Comput., 43(3):B811– B830, 2021

    work page 2021

  7. [7]

    Y . Chen, B. Hosseini, H. Owhadi, and A. M. Stuart. Solving and learning nonlinear PDEs with Gaussian processes. J. Comput. Phys., 447:110668, 2021

    work page 2021

  8. [8]

    Chen and D

    Y . Chen and D. Xiu. Learning stochastic dynamical system via flow map operator. J. Comput. Phys. , 508:112984, 2024

    work page 2024

Show all 69 references
  1. [9]

    Z. Chen, J. Zhang, M. Arjovsky, and L. Bottou. Symplectic recurrent neural networks. In International Conference on Learning Representations, 2020

    2020

  2. [10]

    Churchill and D

    V . Churchill and D. Xiu. Flow map learning for unknown dy namical systems: Overview, implementation, and benchmarks. J. Mach. Learn. Model. Comput. , 4(2):173–201, 2023

    2023

  3. [11]

    G. Crooks. Entropy production fluctuation theorem and t he nonequilibrium work relation for free energy differences. Phys. Rev. E, 60(3):2721, 1999. 9

    1999

  4. [12]

    Dietrich, A

    F. Dietrich, A. Makeev, G. Kevrekidis, N. Evangelou, T. Bertalan, S. Reich, and I. G. Kevrekidis. Learning e ffec- tive stochastic di fferential equations from microscopic simulations: Linking stochastic numerics to deep learning. Chaos., 33(2), 2023

    2023

  5. [13]

    L. Ding, W . Li, S. Osher, and W . Yin. A mean field game inver se problem. J. Sci. Comput., 92(1):7, 2022

    2022

  6. [14]

    J. Dorfman. An introduction to chaos in nonequilibrium statistical mec hanics. Number 14. Cambridge university press, 1999

    1999

  7. [15]

    S. Drobot. On the foundations of dimensional analysis. Stud. Math., 14(1):84–99, 1953

    1953

  8. [16]

    Esposito

    M. Esposito. Stochastic thermodynamics under coarse g raining. Phys. Rev. E., 85(4):041125, 2012

    2012

  9. [17]

    Esposito and C

    M. Esposito and C. V . den Broeck. Three detailed fluctuat ion theorems. Phys. Rev. Lett., 104(9):090601, 2010

    2010

  10. [18]

    L. Evans. An introduction to stochastic di fferential equations, volume 82. American Mathematical Soc., 2012

    2012

  11. [19]

    J. Feng, C. Kulick, and S. Tang. Data-driven model selec tions of second-order particle dynamics via integrating Gaussian processes with low-dimensional interacting stru ctures. Physica D, 461:134097, 2024

    2024

  12. [20]

    Finzi, K

    M. Finzi, K. A. Wang, and A. G. Wilson. Simplifying Hamil tonian and Lagrangian neural networks via explicit constraints. In Advances in Neural Information Processing Systems , 2020

    2020

  13. [21]

    H. Gao, M. J. Zahr, and J.-X. Wang. Physics-informed gra ph neural Galerkin networks: A unified framework for solving PDE-governed forward and inverse problems. Comput. Methods Appl. Mech. Eng. , 390:114502, 2022

    2022

  14. [22]

    Y . Gao, Q. Lang, and F. Lu. Self-test loss functions for l earning weak-form operators and gradient flows. arXiv preprint arXiv:2412.03506, 2024

    2024

  15. [23]

    Ge and H

    H. Ge and H. Qian. Physical origins of entropy productio n, free energy dissipation, and their mathematical repre- sentations. Phys. Rev. E, 81(5):051133, 2010

    2010

  16. [24]

    R. Graham. Statistical theory of instabilities in stat ionary nonequilibrium systems with applications to lasers and nonlinear optics. In In Quantum Statistics in Optics and Solid-state Physics, Sp ringer Tracts in Modern Physics , pages 1–97. Springer, 1973

    1973

  17. [25]

    Graham and H

    R. Graham and H. Haken. Generalized thermodynamic pote ntial for marko ff systems in detailed balance and far from thermal equilibrium. Z. Physik, 243:289–302, 1971

    1971

  18. [26]

    Greydanus, M

    S. Greydanus, M. Dzamba, and J. Y osinski. Hamiltonian n eural networks. In Advances in Neural Information Processing Systems, 2019

    2019

  19. [27]

    Grigorio and M

    L. Grigorio and M. Alqahtani. Quasi-potential and drif t decomposition in stochastic systems by sparse identification. arXiv preprint arXiv:2409.06886 , 2024

    2024 arXiv

  20. [28]

    Gruber, M

    A. Gruber, M. Gunzburger, L. Ju, and Z. Wang. Energetica lly consistent model reduction for metriplectic systems. Comput. Methods Appl. Mech. Eng. , 404:115709, 2023

    2023

  21. [29]

    Gruber, K

    A. Gruber, K. Lee, H. Lim, N. Park, and N. Trask. E fficiently parameterized neural metriplectic systems. In The Thirteenth International Conference on Learning Represen tations, 2025

    2025

  22. [30]

    Hu, J.-P

    J. Hu, J.-P . Ortega, and D. Yin. A structure-preserving kernel method for learning hamiltonian systems. Mathematics of Computation, 2025

    2025

  23. [31]

    Z. Hu, C. Liu, Y . Wang, and Z. Xu. Energetic variational n eural network discretizations of gradient flows. SIAM J. Sci. Comput., 46(4):A2528–A2556, 2024

    2024

  24. [32]

    Huang, Q

    C.-K. Huang, Q. Tang, Y . K. Batygin, O. Beznosov, J. Burb y, A. Kim, S. Kurennoy, T. Kwan, and H. N. Rakotoariv- elo. Symplectic neural surrogate models for beam dynamics. J. Phys.: Conf. Ser ., 2687(6):062026, 2024

    2024

  25. [33]

    Huang, Z

    S. Huang, Z. He, N. Dirr, J. Zimmer, and C. Reina. Statist ical-physics-informed neural networks (stat-pinns): A machine learning strategy for coarse-graining dissipativ e dynamics. J. Mech. Phys. Solids , page 105908, 2024

    2024

  26. [34]

    Huang, Z

    S. Huang, Z. He, and C. Reina. V ariational Onsager Neura l Networks (VONNs): A thermodynamics-based varia- tional learning strategy for non-equilibrium PDEs. J. Mech. Phys. Solids , 163:104856, 2022. 10

    2022

  27. [35]

    Jiang and D.-H

    D.-Q. Jiang and D.-H. Jiang. Mathematical theory of nonequilibrium steady states: on th e frontier of probability and dynamical systems. Springer Science & Business Media, 2004

    2004

  28. [36]

    P . Jin, Z. Zhang, A. Zhu, Y . Tang, and G. Karniadakis. SympNets: Intrinsic structure-preserving symplectic networks for identifying hamiltonian systems. Neural Netw., 132:166–179, 2020

    2020

  29. [37]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, and G. E. Karniadakis. hp-vpinns : V ariational physics-informed neural networks with domain decomposition. Comput. Methods Appl. Mech. Eng. , 374:113547, 2021

    2021

  30. [38]

    J. Kurchan. Fluctuation theorem for stochastic dynami cs. J. Phys. A: Math. Gen. , 31(16):3719, 1998

    1998

  31. [39]

    Lang and F

    Q. Lang and F. Lu. Learning interaction kernels in mean- field equations of first-order systems of interacting particles. SIAM J. Sci. Comput. , 44(1):A260–A285, 2022

    2022

  32. [40]

    K. Lee, N. Trask, and P . Stinis. Machine learning struct ure preserving brackets for forecasting irreversible processes. In Advances in Neural Information Processing Systems , 2021

    2021

  33. [41]

    Y . Li, S. Xu, J. Duan, X. Liu, and Y . Chu. A machine learnin g method for computing quasi-potential of stochastic dynamical systems. Nonlinear Dyn., 109(3):1877–1886, 2022

    2022

  34. [42]

    B. Lin, Q. Li, and W . Ren. A data driven method for computi ng quasipotentials. pages 652–670, 2022

    2022

  35. [43]

    Y . Liu, Y . Chen, D. Xiu, and G. Zhang. A training-free conditional diffusion model for learning stochastic dynamical systems. arXiv preprint arXiv:2410.03108 , 2024

    2024 arXiv

  36. [44]

    F. Lu, Q. An, and Y . Y u. Nonparametric learning of kernel s in nonlocal operators. J. Peridyn. Nonlocal Model. , 2023

    2023

  37. [45]

    F. Lu, M. Maggioni, and S. Tang. Learning interaction ke rnels in stochastic systems of interacting particles from multiple trajectories. F ound. Comput. Math., 22(4):1013–1067, 2022

    2022

  38. [46]

    F. Lu, M. Zhong, S. Tang, and M. Maggioni. Nonparametric inference of interaction laws in systems of agents from trajectory data. Proc. Natl. Acad. Sci. , 116(29):14424–14433, 2019

    2019

  39. [47]

    Y . Lu, X. Li, C. Liu, Q. Tang, and Y . Wang. Learning generalized diffusions using an energetic variational approach. arXiv preprint arXiv:2412.04480 , 2024

    2024

  40. [48]

    S. Ma, S. Liu, H. Zha, and H. Zhou. Learning stochastic be haviour from aggregate data. In M. Meila and T. Zhang, editors, Proceedings of the 38th International Conference on Machin e Learning , volume 139 of Proceedings of Machine Learning Research, pages 7258–7267. PMLR, 18–...

    2021

  41. [49]

    Marconi, A

    U. Marconi, A. Puglisi, L. Rondoni, and A. Vulpiani. Flu ctuation–dissipation: response theory in statistical phy sics. Phys. Rep., 461(4-6):111–195, 2008

    2008

  42. [50]

    Mattheakis, D

    M. Mattheakis, D. Sondak, A. S. Dogra, and P . Protopapas . Hamiltonian neural networks for solving equations of motion. Phys. Rev. E, 105:065305, 2022

    2022

  43. [51]

    D. A. Messenger and D. M. Bortz. Weak SINDy for partial di fferential equations. J. Comput. Phys. , 443:110525, 2021

    2021

  44. [52]

    D. A. Messenger and D. M. Bortz. Learning mean-field equa tions from particle data using WSINDy. Physica D, 439:133406, 2022

    2022

  45. [53]

    D. A. Messenger, J. W . Burby, and D. M. Bortz. Coarse-gra ining hamiltonian systems using wsindy. Scientific Reports, 14(1):14457, 2024

    2024

  46. [54]

    Miller, S

    J. Miller, S. Tang, M. Zhong, and M. Maggioni. Learning t heory for inferring interaction kernels in second-order interacting agent systems. Sampling Theory Signal Process. Data Anal. , 21(1):21, 2023

    2023

  47. [55]

    M. Opper. V ariational inference for stochastic di fferential equations. Annalen der Physik , 531(3):1800233, 2019

    2019

  48. [56]

    H. Qian. A decomposition of irreversible di ffusion processes without detailed balance. J. Math. Phys., 54(5), 2013

    2013

  49. [57]

    Qian and Z

    M. Qian and Z. Wang. The entropy production of di ffusion processes on manifolds and its circulation decomposi - tions. Commun. Math. Phys. , 206:429–445, 1999. 11

    1999

  50. [58]

    Qian and M

    M.-P . Qian and M. Qian. Circulation for recurrent marko v chains. Z. W ahrsch. V erw. Gebiete, 59(2):203–210, 1982

    1982

  51. [59]

    M.-P . Qian, M. Qian, and G. Gong. The reversibility and t he entropy production of markov processes. Contemp. Math, 118:255–261, 1991

    1991

  52. [60]

    Raissi, P

    M. Raissi, P . Perdikaris, and G. E. Karniadakis. Physic s-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear p artial differential equations. J. Comput. Phys., 378:686– 707, 2019

    2019

  53. [61]

    T. D. Ryck, S. Mishra, and R. Molinaro. Weak physics info rmed neural networks for approximating entropy solutions of hyperbolic conservation laws. In Seminar f¨ ur Angewandte Mathematik, Eidgen¨ ossische T echnische Hochschule, Z¨ urich, Switzerland, Rep, volume 35, page 2022, 2022

    2022

  54. [62]

    S¨ arkk¨ a and A

    S. S¨ arkk¨ a and A. Solin.Applied stochastic differential equations, volume 10. Cambridge University Press, 2019

    2019

  55. [63]

    M. O. Williams, I. G. Kevrekidis, and C. W . Rowley. A data –driven approximation of the koopman operator: Extending dynamic mode decomposition. J. Nonlinear Sci., 25:1307–1346, 2015

    2015

  56. [64]

    Z. Xu, D. Long, Y . Xu, G. Y ang, S. Zhe, and H. Owhadi. Towar d efficient kernel-based solvers for nonlinear PDEs. In F orty-second International Conference on Machine Learning, 2025

    2025

  57. [65]

    Y ang, D

    L. Y ang, D. Zhang, and G. E. Karniadakis. Physics-infor med generative adversarial networks for stochastic di ffer- ential equations. SIAM J. Sci. Comput. , 42(1):A292–A317, 2020

    2020

  58. [66]

    H. Y u, X. Tian, W . E, and Q. Li. OnsagerNet: Learning stab le and interpretable dynamics using a generalized Onsager principle. Phys. Rev. Fluids, 6:114402, 2021

    2021

  59. [67]

    Y . Zang, G. Bao, X. Y e, and H. Zhou. Weak adversarial networks for high-dimensional partial differential equations. J. Comput. Phys., 411:109409, 2020

    2020

  60. [68]

    Zhang, S

    J. Zhang, S. Zhang, J. Shen, and G. Lin. Energy-dissipat ive evolutionary deep operator neural networks. J. Comput. Phys., 498:112638, 2024

    2024

  61. [69]

    Zhang, Y

    Z. Zhang, Y . Shin, and G. E. Karniadakis. Gfinns: Generic formalism informed neural networks for deterministic and stochastic dynamical systems. Philos. Trans. Roy. Soc. A, 380(2229):20210207, 2022. A Dimensional Analysis: Penalty in Energy Laws In this appendix, we discuss ou...

    2022

Pith tools

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