arxiv: 2605.00394 · v1 · submitted 2026-05-01 · 💻 cs.LG

Recognition: unknown

Mesh Field Theory: Port-Hamiltonian Formulation of Mesh-Based Physics

Satoshi Noguchi , Yoshinobu Kawahara

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:18 UTC · model grok-4.3

classification 💻 cs.LG

keywords mesh field theoryport-hamiltonian systemsmesh-based physicsstructure-preserving learningneural physical simulationenergy conservationtopological reductionconstitutive relations

0 comments

The pith

Mesh-based physics admits a local factorization into port-Hamiltonian form where mesh topology alone fixes the conservative interconnection.

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

The paper proves that minimal physical principles suffice to separate the dynamics on meshes into a part fixed entirely by topology and a part that depends on metric properties. This factorization takes the explicit form of a port-Hamiltonian system whose conservative interconnection graph is determined uniquely by the mesh. A reader would care because the separation tells exactly which pieces must be hard-coded for physical fidelity and which pieces can safely be learned from data. The resulting neural architecture therefore inherits energy balance and conservation laws automatically rather than having to discover them.

Core claim

We prove a reduction theorem for mesh-based physics. Under these conditions, the physical dynamics admit a local factorization into a port-Hamiltonian form: the conservative interconnection is fixed uniquely by mesh topology, whereas metric effects enter only through constitutive relations and dissipation.

What carries the argument

The reduction theorem that factors mesh dynamics into port-Hamiltonian form with topology-determined conservative interconnection and metric-dependent constitutive relations.

If this is right

MeshFT-Net needs to learn only the metric-dependent constitutive relations and dissipation terms.
Simulations exhibit near-zero energy drift while preserving dispersion relations and momentum.
The model extrapolates robustly outside the training distribution and requires fewer data points.
Non-physical degrees of freedom are eliminated by construction rather than penalized during training.

Where Pith is reading between the lines

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

The same topological factorization may apply to other discrete structures such as graphs or simplicial complexes that carry orientation and locality.
Architectures that hard-code only the topology-derived interconnection could be tested on hybrid continuum-discrete problems where part of the domain is meshed and part is not.
If the reduction holds for time-dependent metrics, it would allow online adaptation of material properties without retraining the entire interconnection structure.

Load-bearing premise

The four minimal physical principles of locality, permutation equivariance, orientation covariance, and energy balance/dissipation inequality are already enough to guarantee that the conservative interconnection depends only on mesh topology.

What would settle it

A concrete mesh-based physical system obeying locality, permutation equivariance, orientation covariance, and the energy balance inequality whose conservative interconnection nevertheless changes when the metric is altered.

Figures

Figures reproduced from arXiv: 2605.00394 by Satoshi Noguchi, Yoshinobu Kawahara.

**Figure 1.** Figure 1: Core concept of this study—comparison of MeshFT and MGN by underlying physical assumptions: (L) locality, (P) Permutation equivariance, (O) Orientation covariance, (E) Non-increasing energy. MGN attains (L) and (P) by architecture design, whereas MeshFT additionally enforces (O) and (E), yielding a clear modeling guideline: fix topology (incidence-based interconnection) and learn metricdependent structure… view at source ↗

**Figure 2.** Figure 2: Relationship between the size of training dataset and onestep MSE (top) and rollout energy drift (bottom) for grid mesh. MeshFT-Net’s topological interconnection with a symplectic step yields orders of magnitude greater robustness. Implementation details and additional results appear in Appendix C.1. We also test a Rayleigh-damped setting (amplitude ∝ e −γt); details appear in Appendix C.2. For HNN, we… view at source ↗

**Figure 3.** Figure 3: Pressure snapshots at equal time steps ordered right to left. The rightmost frame is the initial state (shared colormap). 5.4. Acoustic Scattering Benchmark from The Well To assess transfer beyond synthetic data, we evaluate MeshFT-Net on a subset of The Well—Acoustic Scattering (Ohana et al., 2024; Mandli et al., 2016) which is near-Hamiltonian but includes discontinuous media and open/reflective boundar… view at source ↗

**Figure 4.** Figure 4: Relationship between the size of training dataset and one-step MSE (left) and rollout energy drift (right) for random delaunay mesh. Metrics and Hyperparameters. For evaluation, a shared theory Hodge (M, W) = (V0, c2V −1 1 ) defines the physical norm used for relative error and for energy-drift over open-loop rollouts (∆t=0.002, T=200). Training runs for 10 epochs with a mini-batch size of 8 on 2000 traini… view at source ↗

read the original abstract

We present Mesh Field Theory (MeshFT) and its neural realization, MeshFT-Net: a structure-preserving framework for mesh-based continuum physics that cleanly separates the physics' topological structure from its metric structure. Imposing minimal physical principles (locality, permutation equivariance, orientation covariance, and energy balance/dissipation inequality), we prove a reduction theorem for mesh-based physics. Under these conditions, the physical dynamics admit a local factorization into a port-Hamiltonian form: the conservative interconnection is fixed uniquely by mesh topology, whereas metric effects enter only through constitutive relations and dissipation. This reduction clarifies what must be fixed and what should be learned, directly informing MeshFT-Net's design. Across evaluations on analytic and realistic datasets, physics-consistency tests, and out-of-distribution validation, MeshFT-Net achieves near-zero energy drift and strong physical fidelity (correct dispersion and momentum conservation) along with robust extrapolation and high data efficiency. By eliminating non-physical degrees of freedom and learning only metric-dependent structure, MeshFT provides a principled inductive bias for stable, faithful, and data-efficient learning-based physical simulation.

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.

Desk Editor's Note private letter to a colleague

The reduction theorem in this paper looks promising for structure-preserving simulation but hinges on whether those four principles really pin down the topology-only interconnection without more assumptions.

read the letter

The main point is that they have derived a reduction theorem showing mesh-based physics can be put into port-Hamiltonian form where the skew-symmetric interconnection comes only from the mesh topology, while metric stuff goes into the constitutive laws. This leads to MeshFT-Net that respects this split and gets good conservation in practice. What the paper does well is make the separation explicit and use it to guide the model design, avoiding learning unnecessary non-physical parameters. The results on energy drift being near zero and maintaining physical properties like dispersion and momentum across different datasets are concrete evidence that the approach works for stable long-term simulation. The out-of-distribution validation also suggests the inductive bias helps with generalization. The soft spot is around the uniqueness in the reduction. The four principles—locality, permutation equivariance, orientation covariance, and energy balance—are supposed to fix the interconnection uniquely by topology. But as the stress-test notes, it's possible that the way the state space or the local factorization is set up adds extra constraints that do some of the work. The paper includes the proof, so one can check if it avoids that, but it would be worth verifying if the axioms are sufficient on their own. The experiments could use more baseline comparisons and statistical details to make the fidelity claims stronger, though the energy preservation is a good sign. This paper is for people building neural models for physical simulations on meshes, like in engineering or fluid dynamics. A reader looking for structure-preserving methods with a theoretical backing will get value from the framework and the net architecture. It deserves a serious referee because it has a formal result plus empirical validation, even if the proof needs polishing for clarity. I recommend putting it through peer review.

Referee Report

2 major / 2 minor

Summary. The paper introduces Mesh Field Theory (MeshFT) and its neural realization MeshFT-Net for mesh-based continuum physics. Imposing four minimal principles (locality, permutation equivariance, orientation covariance, and energy balance/dissipation inequality), it proves a reduction theorem asserting that the dynamics admit a local factorization into port-Hamiltonian form, with the conservative (skew-symmetric) interconnection fixed uniquely by mesh topology while metric effects appear only in constitutive relations and dissipation. This separation guides the architecture of MeshFT-Net, which is reported to achieve near-zero energy drift, correct dispersion relations, momentum conservation, and strong out-of-distribution performance on analytic and realistic datasets.

Significance. If the reduction theorem holds, the work supplies a clean theoretical separation between topological and metric structure that directly informs inductive biases for structure-preserving neural simulators. The reported empirical outcomes (near-zero energy drift together with physical fidelity metrics) would constitute a practical advance for stable, data-efficient learning of continuum physics on meshes.

major comments (2)

[Reduction theorem section] The reduction theorem (abstract and the section deriving the port-Hamiltonian factorization): the assertion that the four listed principles alone force the conservative interconnection to be fixed uniquely by combinatorial topology must be shown to exclude other admissible skew-symmetric operators that still satisfy power balance and the stated axioms but incorporate additional data (e.g., a reference metric or non-canonical pairing). Without an explicit characterization or counter-example ruling out such alternatives, the claimed clean separation between topology and metric remains unverified.
[Evaluations section] Evaluations section (physics-consistency and OOD tests): the manuscript reports near-zero energy drift and strong fidelity but supplies neither explicit error bars, quantitative baseline comparisons against non-structure-preserving or generic graph networks, nor details on data exclusion criteria. These omissions prevent assessment of whether the observed performance is attributable to the topological bias or to other modeling choices.

minor comments (2)

[Notation and preliminaries] Notation for the Dirac structure and the mesh incidence operators should be introduced once and used consistently; occasional re-use of symbols for both combinatorial and metric quantities creates ambiguity.
[Figures] Figure captions for energy-drift plots would benefit from insets or log-scale insets to make the near-zero behavior visually quantifiable.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. The comments have helped us strengthen the presentation of the reduction theorem and the experimental reporting. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

Referee: [Reduction theorem section] The reduction theorem (abstract and the section deriving the port-Hamiltonian factorization): the assertion that the four listed principles alone force the conservative interconnection to be fixed uniquely by combinatorial topology must be shown to exclude other admissible skew-symmetric operators that still satisfy power balance and the stated axioms but incorporate additional data (e.g., a reference metric or non-canonical pairing). Without an explicit characterization or counter-example ruling out such alternatives, the claimed clean separation between topology and metric remains unverified.

Authors: We appreciate the referee's emphasis on rigor here. The reduction theorem (Theorem 3.1) derives the port-Hamiltonian factorization directly from the four axioms and shows that any operator satisfying locality, permutation equivariance, orientation covariance, and power balance must coincide with the combinatorial incidence structure of the mesh (the boundary operator). The proof proceeds by first establishing skew-symmetry from power balance, then using locality and equivariance to restrict the support and symmetry of the operator, and finally invoking orientation covariance to fix the signs and exclude metric-dependent pairings. To make the uniqueness explicit, we have added a new corollary (Corollary 3.2) that characterizes all admissible skew-symmetric operators under the axioms: they are precisely the topological pairings induced by the mesh complex without reference to any metric. We also include a short counter-example paragraph demonstrating that inserting a non-constant reference metric violates either permutation equivariance (under mesh automorphisms) or orientation covariance (under orientation-reversing maps). These additions confirm the claimed separation without altering the original theorem statement. The revised section now contains this characterization and counter-example. revision: partial
Referee: [Evaluations section] Evaluations section (physics-consistency and OOD tests): the manuscript reports near-zero energy drift and strong fidelity but supplies neither explicit error bars, quantitative baseline comparisons against non-structure-preserving or generic graph networks, nor details on data exclusion criteria. These omissions prevent assessment of whether the observed performance is attributable to the topological bias or to other modeling choices.

Authors: The referee is correct that these reporting details were missing. We have revised the Evaluations section (now Section 5) as follows: (i) all quantitative results now include explicit error bars (mean ± one standard deviation over five independent runs with different random seeds); (ii) we added direct comparisons against two baselines—a standard graph convolutional network without port-Hamiltonian structure and a generic MLP operating on flattened mesh features—showing that MeshFT-Net achieves orders-of-magnitude lower energy drift and superior OOD accuracy; (iii) we added a dedicated paragraph on data protocol, specifying that 5% of samples with extreme boundary conditions were excluded for solver stability, that the split is 70/15/15, and that OOD test sets use disjoint initial-condition distributions. These changes allow readers to evaluate the contribution of the topological bias. The revised manuscript contains the new tables, baseline results, and protocol description. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the reduction theorem

full rationale

The paper derives its central reduction theorem directly from four stated minimal physical principles (locality, permutation equivariance, orientation covariance, energy balance/dissipation inequality) to conclude that conservative interconnection is fixed uniquely by mesh topology while metric effects are isolated to constitutive relations. No equations or claims in the provided abstract or description reduce the theorem to a self-definition, a fitted parameter renamed as prediction, or a load-bearing self-citation whose content is itself unverified. The factorization is presented as a consequence of the axioms rather than presupposed by them, rendering the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests primarily on the domain assumptions of locality, equivariance, covariance, and energy balance; the paper introduces the MeshFT framework and MeshFT-Net as new constructs without explicit free parameters or invented physical entities beyond the framework itself.

axioms (1)

domain assumption locality, permutation equivariance, orientation covariance, and energy balance/dissipation inequality
These are the minimal physical principles imposed to prove the reduction theorem as described in the abstract.

invented entities (2)

Mesh Field Theory (MeshFT) no independent evidence
purpose: Structure-preserving framework separating topological and metric structure in mesh-based physics
Newly introduced theory in the paper to organize mesh physics.
MeshFT-Net no independent evidence
purpose: Neural realization of MeshFT for learning metric-dependent parts
The neural network architecture designed based on the reduction theorem.

pith-pipeline@v0.9.0 · 5487 in / 1614 out tokens · 66333 ms · 2026-05-09T20:18:45.935366+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

70 extracted references · 3 canonical work pages · 1 internal anchor

[1]

Large-Scale Kernel Machines , publisher =

Yoshua Bengio and Yann LeCun , title =. Large-Scale Kernel Machines , publisher =
[2]

Hinton and Simon Osindero and Yee Whye Teh , title =

Geoffrey E. Hinton and Simon Osindero and Yee Whye Teh , title =. Neural Computation , volume =
[3]

Ian Goodfellow and Yoshua Bengio and Aaron Courville , title =
[4]

International Conference on Learning Representations , year =

Learning Mesh-Based Simulation with Graph Networks , author =. International Conference on Learning Representations , year =
[5]

Hirani and Melvin Leok and Jerrold E

Mathieu Desbrun and Anil N. Hirani and Melvin Leok and Jerrold E. Marsden , title =. 2005 , note =

2005
[6]

Hirani , title =

Anil N. Hirani , title =
[7]

IEEE Transactions on Antennas and Propagation , volume =

Kane Yee , title =. IEEE Transactions on Antennas and Propagation , volume =
[8]

Harley Flanders , title =
[9]

Raissi and P

M. Raissi and P. Perdikaris and G. E. Karniadakis , title =. Journal of Computational Physics , volume =. 2019 , doi =

2019
[10]

Communications in Mathematics and Statistics , volume =

Weinan E and Bing Yu , title =. Communications in Mathematics and Statistics , volume =
[11]

International Conference on Learning Representations , year =

Fourier Neural Operator for Parametric Partial Differential Equations , author =. International Conference on Learning Representations , year =
[12]

Nature Machine Intelligence , volume =

Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , title =. Nature Machine Intelligence , volume =
[13]

Journal of Machine Learning Research , volume =

Nikola Kovachki and Zongyi Li and Burigede Liu and Kamyar Azizzadenesheli and Kaushik Bhattacharya and Andrew Stuart and Anima Anandkumar , title =. Journal of Machine Learning Research , volume =
[14]

Advances in Neural Information Processing Systems , volume =

Gaurav Gupta and Xiongye Xiao and Paul Bogdan , title =. Advances in Neural Information Processing Systems , volume =
[15]

Proceedings of the 37th International Conference on Machine Learning , series =

Learning to Simulate Complex Physics with Graph Networks , author =. Proceedings of the 37th International Conference on Machine Learning , series =. 2020 , publisher =

2020
[16]

International Conference on Learning Representations , year =

Lagrangian Fluid Simulation with Continuous Convolutions , author =. International Conference on Learning Representations , year =
[17]

Relational inductive biases, deep learning, and graph networks

Peter W. Battaglia and Jessica B. Hamrick and Victor Bapst and Alvaro Sanchez-Gonzalez and Vinicius Zambaldi and Mateusz Malinowski and Andrea Tacchetti and David Raposo and Adam Santoro and Ryan Faulkner and others , title =. arXiv preprint arXiv:1806.01261 , year =

work page internal anchor Pith review arXiv
[18]

ACM SIGGRAPH 2005 Courses , pages =

Mathieu Desbrun and Eva Kanso and Yiying Tong , title =. ACM SIGGRAPH 2005 Courses , pages =. 2005 , publisher =

2005
[19]

Arnold and Richard S

Douglas N. Arnold and Richard S. Falk and Ragnar Winther , title =. Acta Numerica , volume =
[20]

Alain Bossavit , title =
[21]

Hagness and Melinda Piket-May , title =

Allen Taflove and Susan C. Hagness and Melinda Piket-May , title =. The Electrical Engineering Handbook , publisher =
[22]

Advances in Neural Information Processing Systems , volume =

Samuel Greydanus and Misko Dzamba and Jason Yosinski , title =. Advances in Neural Information Processing Systems , volume =
[23]

Symplectic

Yaofeng Desmond Zhong and Biswadip Dey and Amit Chakraborty , booktitle =. Symplectic
[24]

Lagrangian neural networks.arXiv:2003.04630,

Miles Cranmer and Sam Greydanus and Stephan Hoyer and Peter Battaglia and David Spergel and Shirley Ho , title =. arXiv preprint arXiv:2003.04630 , year =

work page arXiv 2003
[25]

Symplectic Learning for Hamiltonian Neural Networks , journal =

Marco David and Florian M. Symplectic Learning for Hamiltonian Neural Networks , journal =
[26]

Pseudo-Hamiltonian Neural Networks for Learning Partial Differential Equations , journal =

S. Pseudo-Hamiltonian Neural Networks for Learning Partial Differential Equations , journal =. 2024 , doi =

2024
[27]

Advances in Neural Information Processing Systems , volume =

Ruben Ohana and Michael McCabe and Lucas Meyer and Rudy Morel and Fruzsina Agocs and Miguel Beneitez and Marsha Berger and Blakesly Burkhart and Stuart Dalziel and Drummond Fielding and others , title =. Advances in Neural Information Processing Systems , volume =
[28]

Mandli and Aron J

Kyle T. Mandli and Aron J. Ahmadia and Marsha Berger and Donna Calhoun and David L. George and Yiannis Hadjimichael and David I. Ketcheson and Grady I. Lemoine and Randall J. LeVeque , title =. PeerJ Computer Science , volume =
[29]

Computer Methods in Applied Mechanics and Engineering , volume =

Satoshi Noguchi and Misumi Nakamichi and Kenji Oguni , title =. Computer Methods in Applied Mechanics and Engineering , volume =
[30]

Zienkiewicz and Robert L

Olgierd C. Zienkiewicz and Robert L. Taylor , title =
[31]

SIAM Journal on Numerical Analysis , volume =

Gilbert Strang , title =. SIAM Journal on Numerical Analysis , volume =. 1968 , doi =

1968
[32]

LeVeque , title =

Randall J. LeVeque , title =
[33]

IBM Journal of Research and Development , volume =

Richard Courant and Kurt Friedrichs and Hans Lewy , title =. IBM Journal of Research and Development , volume =
[34]

Tenenbaum and Tao Du and Chuang Gan and Wojciech Matusik , title =

Pingchuan Ma and Peter Yichen Chen and Bolei Deng and Joshua B. Tenenbaum and Tao Du and Chuang Gan and Wojciech Matusik , title =. International Conference on Machine Learning , pages =. 2023 , publisher =

2023
[35]

Morrison and John M

Philip J. Morrison and John M. Greene , title =. Physical Review Letters , volume =
[36]

Morrison , title =

Philip J. Morrison , title =. AIP Conference Proceedings , volume =
[37]

Journal of computational physics , volume=

DGM: A deep learning algorithm for solving partial differential equations , author=. Journal of computational physics , volume=. 2018 , publisher=

2018
[38]

Choose a Transformer: Fourier or Galerkin , volume =

Cao, Shuhao , booktitle =. Choose a Transformer: Fourier or Galerkin , volume =
[39]

Journal of Nonlinear Science , volume=

Time integration and discrete Hamiltonian systems , author=. Journal of Nonlinear Science , volume=. 1996 , publisher=

1996
[40]

Brunton and Joshua L

Steven L. Brunton and Joshua L. Proctor and J. Nathan Kutz , title =. Proceedings of the National Academy of Sciences , volume =
[41]

Nathan Kutz and Steven L

Bethany Lusch and J. Nathan Kutz and Steven L. Brunton , title =. Nature Communications , volume =
[42]

Proctor and Steven L

Joshua L. Proctor and Steven L. Brunton and J. Nathan Kutz , title =. SIAM Journal on Applied Dynamical Systems , volume =
[43]

Peter Van Overschee and Bart De Moor , title =
[44]

van der Schaft and Dimitri Jeltsema , title =

Arjan J. van der Schaft and Dimitri Jeltsema , title =. Foundations and Trends in Systems and Control , volume =
[45]

Introduction to smooth manifolds , pages=

Smooth manifolds , author=. Introduction to smooth manifolds , pages=. 2003 , publisher=

2003
[46]

Qing Han and Fanghua Lin , title =
[47]

Rajendra Bhatia , title =
[48]

Evans and Ronald F

Lawrence C. Evans and Ronald F. Gariepy , title =
[49]

International Conference on Machine Learning , pages=

Graph-coupled oscillator networks , author=. International Conference on Machine Learning , pages=
[50]

2024 , author =

Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes , journal =. 2024 , author =

2024
[51]

2022 , author =

Enforcing exact physics in scientific machine learning: A data-driven exterior calculus on graphs , journal =. 2022 , author =

2022
[52]

2023 , booktitle =

Gruber, Anthony and Lee, Kookjin and Trask, Nathaniel , title =. 2023 , booktitle =

2023
[53]

Philosophical Transactions of the Royal Society A , volume=

GFINNs: GENERIC formalism informed neural networks for deterministic and stochastic dynamical systems , author=. Philosophical Transactions of the Royal Society A , volume=. 2022 , publisher=

2022
[54]

2021 , author =

Structure-preserving neural networks , journal =. 2021 , author =

2021
[55]

Advances in Neural Information Processing Systems , editor=

Machine learning structure preserving brackets for forecasting irreversible processes , author=. Advances in Neural Information Processing Systems , editor=
[56]

C., Arratia, P

Data-driven particle dynamics: Structure-preserving coarse-graining for emergent behavior in non-equilibrium systems , author=. arXiv preprint arXiv:2508.12569 , year=

work page arXiv
[57]

International Conference on Machine Learning , pages =

GRAND: Graph Neural Diffusion , author =. International Conference on Machine Learning , pages =
[58]

Deep Sets , volume =

Zaheer, Manzil and Kottur, Satwik and Ravanbakhsh, Siamak and Poczos, Barnabas and Salakhutdinov, Russ R and Smola, Alexander J , booktitle =. Deep Sets , volume =
[59]

International Conference on Learning Representations , year=

Invariant and Equivariant Graph Networks , author=. International Conference on Learning Representations , year=
[60]

Advances in neural information processing systems , volume=

Universal invariant and equivariant graph neural networks , author=. Advances in neural information processing systems , volume=
[61]

U-Net: Convolutional Networks for Biomedical Image Segmentation

Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015. 2015

2015
[62]

LeVeque, Randall J. , year=. Finite Volume Methods for Hyperbolic Problems , publisher=
[63]

Weak Form Generalized Hamiltonian Learning , volume =

Course, Kevin and Evans, Trefor and Nair, Prasanth , booktitle =. Weak Form Generalized Hamiltonian Learning , volume =
[64]

2026 , howpublished =

Anonymous prior work , author =. 2026 , howpublished =

2026
[65]

Acta Numerica , author=

Finite element exterior calculus, homological techniques, and applications , volume=. Acta Numerica , author=. 2006 , pages=

2006
[66]

Proceedings of the National Academy of Sciences , volume =

Rose Yu and Rui Wang , title =. Proceedings of the National Academy of Sciences , volume =. 2024 , doi =

2024
[67]

Advances in Neural Information Processing Systems , year=

Neural Conservation Laws: A Divergence-Free Perspective , author=. Advances in Neural Information Processing Systems , year=
[68]

Neural Symplectic Form: Learning Hamiltonian Equations on General Coordinate Systems , volume =

Chen, Yuhan and Matsubara, Takashi and Yaguchi, Takaharu , booktitle =. Neural Symplectic Form: Learning Hamiltonian Equations on General Coordinate Systems , volume =
[69]

Port-Hamiltonian neural networks for learning explicit time-dependent dynamical systems , author =. Phys. Rev. E , volume =. 2021 , publisher =

2021
[70]

Learning for Dynamics and Control Conference , pages=

Compositional learning of dynamical system models using port-Hamiltonian neural networks , author=. Learning for Dynamics and Control Conference , pages=. 2023 , organization=

2023