CoSynFlow: Conformal Symplectic Neural Flows for Cross-System Prediction of Dissipative Hamiltonian Dynamics

Baige Xu; Takaharu Yaguchi

REVIEW 2 major objections 4 minor 22 references

CoSynFlow: Conformal Symplectic Neural Flows for Cross-System Prediction of Dissipative Hamiltonian Dynamics

T0 review · 2 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A neural flow architecture that makes every learned time-t map exactly conformal symplectic: the symplectic form is scaled by the prescribed dissipation factor e^{γt} for any trainable parameters, by construction.

desk verdict CoSynFlow is a genuinely new construction with a correct exact conformal-symplectic guarantee and careful experiments, but the title overreaches beyond the canonical conformal class and missing code/data keep it at conditional acceptance. read the letter →

arxiv 2608.00571 v1 pith:QC5Z5GGR submitted 2026-08-01 cs.LG

Baige Xu , Takaharu Yaguchi This is my paper · ORCID

classification cs.LG MSC 37M1565P1068T07

keywords conformalsymplecticdissipativeHamiltoniansystemsneuralflowsstructure-preservinglearningoperatorscross-systempredictionphysics-informedtrainingintegrators

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

CoSynFlow learns continuous-time solution maps for dissipative Hamiltonian systems with the geometric law of the dynamics built into the architecture: every time-t map is exactly conformal symplectic, meaning it pulls the symplectic form back to $e^{{γt}}$ω for any choice of trainable parameters. The construction composes symplectic shear maps generated by scalar potentials with explicit conformal scalings whose factors are normalized to multiply to $e^{{γt}}$, so the correct phase-space contraction rate is guaranteed rather than learned or penalized. Because the maps are differentiable in the query time, the same model can be evaluated at any t, composed arbitrarily many times beyond the training horizon while keeping the exact conformal factor, and trained with a physics-informed residual. A single model conditioned on a finite-dimensional Hamiltonian descriptor and the dissipation parameter predicts solution maps for unseen systems without retraining. This matters for long-horizon simulation of damped mechanical systems, where ordinary symplectic models cannot contract phase-space volume and unstructured models learn the wrong rate.

What carries the argument

The load-bearing object is the symmetric block F^t_ℓ = $C^{{t/2}}$_ℓ ∘ $P^{{t/2}}$_ℓ ∘ Q^t_ℓ ∘ $P^{{t/2}}$_ℓ ∘ $C^{{t/2}}$_ℓ, built from two kinds of elementary maps on phase space z=(q,p). The shear maps $P^{{t/2}}$_ℓ(q,p) = (q, p − (t/2)∇_q Â_ℓ(q;t)) and Q^t_ℓ(q,p) = (q + t∇_p B̂_ℓ(p;t), p) are symplectic because they are gradients of scalar potentials, so their Hessians are symmetric; the scaling map $C^{{t/2}}$_ℓ(q,p) = (q, $e^{{γt αℓ/2}}$ p) multiplies ω by $e^{{γt αℓ/2}}$. Since the allocations satisfy Σ_ℓ αℓ(ξ) = 1 for every system condition, composing the blocks multiplies the factors to exactly $e^{{γt}}$, independent of all trained parameters. The system conditioner O_θ maps the descriptor (h_H, γ) to the potentials and allocations, and the explicit factors of t make the whole map differentiable in the query time, which is what enables continuous-time evaluation, composition, and physics-informed training.

What would settle it

Evaluate the structure error ‖DΦ̂^t_θ(z)^⊤ J DΦ̂^t_θ(z) − $e^{{γt}}$J‖_F for a trained or randomly initialized CoSynFlow in double precision at several t: if it exceeds values near $10^{{-13}}$ (the round-off level reported in the paper), Theorem 1 fails. A second check is to apply the model to a dissipative system whose damping coefficient depends on position or momentum, or varies in time, and compare the measured contraction rate of the symplectic form with $e^{{γt}}$; the mismatch would show that the exact geometric guarantee does not describe that system's dynamics.

Watch

Extended reading notes

Core claim

The paper's central claim is that for arbitrary trainable parameters, each time-t map of CoSynFlow satisfies (Φ̂^t_θ(·;H,ξ))^*ω = $e^{{γt}}$ω, equivalently DΦ̂^t_θ(z)^⊤ J DΦ̂^t_θ(z) = $e^{{γt}}$J for every phase-space point z, so the learned flow contracts the symplectic form at exactly the prescribed dissipation rate. This is achieved by composing L symmetric blocks, each shaped as $C^{{t/2}}$_ℓ ∘ $P^{{t/2}}$_ℓ ∘ Q^t_ℓ ∘ $P^{{t/2}}$_ℓ ∘ $C^{{t/2}}$_ℓ, where P and Q are gradient shear maps that preserve ω and C is an explicit scaling of the momentum variables; the scaling exponents αℓ depend on the system condition but always sum to one, which makes the total factor exactly $e^{{γt}}$. The same argument gives an explicit inverse, determinant $e^{{dγt}}$, and the identity map at t=0, and Corollary 1 extends the exact factor to any finite composition of maps, even when the total elapsed time exceeds the training horizon. A second theorem shows the exact constraint costs no expressivity: any C² conformal Hamiltonian flow can be approximated uniformly on compact sets by such maps with the conformal factor intact, via a change of variables that converts the conformal flow into a time-dependent symplectic flow.

Load-bearing premise

The target dynamics must be conformal Hamiltonian systems of the exact canonical form ż = J∇H(z) + γ diag(0, I_d)z, with the Hamiltonian H known and the dissipation parameter γ constant and acting only linearly on the momenta; if the real damping has a different form, the enforced $e^{{γt}}$ factor is the wrong geometry.

Editorial extensions

If this is right

For any trained checkpoint, the conformal symplectic structure error stays at double-precision round-off, so no penalty term or post-hoc projection is needed to enforce the geometry.
Composing the learned map across time windows keeps the exact factor e^{γt}, so long-horizon predictions contract phase-space volume at the prescribed rate e^{dγt} even beyond the training interval.
One shared model conditioned on a Hamiltonian descriptor and dissipation parameter predicts solution maps for held-out systems, not just the systems seen in training.
Because the flow is differentiable in t, the residual of the governing equation can be used as a training objective, which cuts data requirements by roughly four times in the low-data experiment reported.
Ordinary symplectic models are not a weaker version of the right constraint here; enforcing volume preservation exactly on a dissipative system is the wrong geometry and yields worse long-horizon error than no constraint at all.

Reading between the lines

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

The architecture's factorization—structure-preserving shears carrying the learned dynamics plus one analytic scaling carrying the geometric weight—is a template that could be carried to other geometric structures, such as Poisson or metriplectic systems, though the paper only suggests this as future work.
A natural testable extension is to allow the dissipation parameter to vary with time or state: the exact factor would become e^{∫γ(t)dt}, and it is an open question whether normalized allocations can still realize that factor with the same block construction.
The cross-system guarantee depends on the descriptor h_H capturing enough of the Hamiltonian; if two systems share a descriptor on the fixed sensor grid but differ away from it, the structure guarantee still holds for both, but predictive accuracy on the unseen system may degrade—this is a property of the conditioning, not of the symplectic exactness.
One can probe the limits of the linear-in-p damping assumption by training CoSynFlow on a Rayleigh-damped system with state-dependent damping; the exactly enforced e^{γt} factor would remain exact but would no longer match the true contraction, separating the geometry guarantee from the modeling assumption.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

CoSynFlow is worth taking seriously. The central construction is new and the main geometric claim checks out: composing symplectic shears with explicit conformal scalings gives a map whose pullback of omega is exactly e^{gamma t} omega for any trainable parameters. Theorem 1 is a clean, correct calculation, and the architecture genuinely delivers the first exactly conformal-symplectic neural flow, plus a system conditioner that permits cross-system prediction. The experiments are more careful than most: three seeds, matched parameter counts, structure error at round-off, long-horizon composition, dissipation extrapolation, and a physics-informed training variant. The ablation that shows ordinary Symplectic Flow is worse than an unstructured MLP Flow is a nice demonstration that enforcing the wrong geometry is actively harmful.

The soft spots are real but not fatal. The scope is narrower than the title suggests. The exact guarantee only applies to systems in the canonical conformal form (3): constant gamma, dissipation linear in momenta. The stress-test concern about Rayleigh damping or time-varying gamma is correct as a boundary issue; for those systems the enforced e^{gamma t} factor is not the right geometry. The abstract does say "conformal symplectic structure," so the paper is not hiding the class, but the title and framing overreach and the boundary of applicability is never discussed. A serious revision should add the qualifier and state the limitation explicitly.

Reproducibility is the second weakness: no code or data release. The dataset is synthetic and requires the exact Hamiltonian for each system; releasing the generator and training code is necessary for independent verification. Third, Theorem 2's sketch is honest but depends on a chain of cited results; the load-bearing guarantee is Theorem 1, so this is a minor concern. Fourth, the related-work citation [17] is wrong: it points to a camouflaged-object-detection paper, not discrete-gradient integrators. Sloppy, easy to fix.

People working on structure-preserving neural operators for Hamiltonian systems and solution operators for dissipative dynamics will get the most from this. Overall, it deserves peer review. The construction is solid enough to warrant referee time, and the claims can be fixed with framing and artifacts. I'd send it with a request for a revised title, an applicability statement, and code/data release.

Referee Report

2 major / 4 minor

Summary. CoSynFlow proposes a neural flow architecture for learning solution maps of dissipative Hamiltonian systems in the conformal symplectic class defined by Eq. (3), namely ż = J∇H(z) + γ diag(0, I_d)z with constant γ < 0. Each layer composes two gradient shear maps and an explicit momentum scaling; the scaling allocations are normalized to sum to one, so the composed time-t map satisfies (Φ̂^t_θ)^*ω = e^{γt}ω for arbitrary trainable parameters (Theorem 1), and the same holds under arbitrary composition (Corollary 1). The architecture is conditioned on a finite-dimensional Hamiltonian descriptor and the dissipation parameter through an FNO, enabling one model to predict solution maps for unseen systems. Theorem 2 claims uniform approximation of the true conformal flow on compact sets without losing the exact conformal factor. Experiments compare CoSynFlow with five baselines on four analytic benchmark systems, reporting low long-horizon state and energy errors, structure error at double-precision round-off, extrapolation in the dissipation parameter, and a physics-informed training variant that exploits differentiability in the query time.

Significance. If the claims hold, the paper offers a clean and useful construction: exact conformal symplecticity is enforced by the architecture rather than by a penalty, and the reported structure error at machine precision is a direct consequence of Theorem 1 rather than of training. The experimental protocol is careful in several respects: three seeds are used, all baselines match the parameter count within 3%, benchmark systems are held out from the training distribution, and long-horizon predictions are produced by a well-defined composition protocol. The cross-system conditioning and continuous-time queries are practically valuable, and the physics-informed training experiment gives a concrete benefit in the low-data regime. The main reservations are the scope of the title-level claim and the non-self-contained proof of Theorem 2.

major comments (2)

[Sections 1 and 3.2, Eq. (3)] The exact geometric guarantee (Theorem 1) applies only to conformal Hamiltonian systems of the canonical form (3), i.e., constant γ and dissipation acting linearly on momenta. The title and the first contribution bullet in Section 1 phrase the claim as 'dissipative Hamiltonian dynamics' without this qualifier. For non-conformal dissipative systems, such as Rayleigh damping ˙p = −∇_qH − C(q,p)∇_pR or position-dependent dissipation, the true flow does not satisfy Dφ^t(z)^T J Dφ^t(z) = e^{γt}J, so CoSynFlow would enforce the wrong geometry; this is exactly the failure mode the paper demonstrates for ordinary Symplectic Flow in Section 6.2. The authors should either narrow the title and abstract or add a prominent statement of the class boundary and a discussion of why the construction is not intended for general dissipative Hamiltonian systems.
[Appendix A.2, Theorem 2] The proof of Theorem 2 is not self-contained. It invokes [6, Theorem 1 and Appendix F] and then asserts that the proof of [6] works under a weakened containment assumption 'through a Gronwall estimate together with a stopping-time argument on a compact neighbourhood,' without reproducing that argument. This is load-bearing because Theorem 2 is the theoretical support for the claim that exact conformal symplecticity does not limit approximation capability. The authors should either make the necessary parts of the argument self-contained, including the Gronwall/stopping-time step and the five-step approximation chain, or state Theorem 2 as conditional on a verified version of the cited result with the weakened assumption made fully explicit.

minor comments (4)

[Appendix B.4] The sentence 'Latency figures reported elsewhere in this document were measured on the same partition...' refers to latency figures that do not appear anywhere in the manuscript; please either include those figures or remove the sentence.
[Section 2, reference [17]] Reference [17], 'Deep gradient learning for efficient camouflaged object detection,' does not appear to support the claim about discrete-gradient models and energy-behavior-preserving integrators; this citation seems to be mismatched and should be replaced with the intended discrete-gradient reference.
[Tables 1 and 2] Several baseline entries report uncertainties such as '±0.0×10−2', which suggests the standard deviation is below the displayed precision; please specify the number of significant digits or report the actual standard deviation so the reader can interpret the entry.
[Appendix B.1] The stated descriptor grid spacing of 6/32 = 0.1875 corresponds to cell centers of a 32×32 grid over [−3,3]^2, not to a grid that includes the boundary points, for which the spacing would be 6/31 ≈ 0.1935; please clarify which convention is used, since the resolution argument in B.1 depends on it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the conformal-symplectic guarantee is enforced by construction and the predictive results are measured on held-out systems.

full rationale

No circular step was found. The exact conformal symplecticity in Theorem 1 is an algebraic consequence of the architecture rather than a fitted datum: each gradient shear is symplectic because its Jacobian contains a symmetric Hessian, each scaling contributes a factor e^{gamma*t*alpha_l/2}, and the normalization sum_l alpha_l = 1 multiplies these factors to exactly e^{gamma*t}. The machine-precision structure errors in Tables 1 and 7 are numerical checks of this identity, not fitted parameters renamed as predictions. Predictive accuracy is measured on held-out random systems and on the four benchmark systems B1-B4, which are outside the training distribution, so the accuracy claims have independent empirical content. The two self-citations, [13] and [18], appear only in the related-work discussion and are not load-bearing for any theorem, approximation argument, or experimental construction; the approximation result in Theorem 2 relies on external results [6,20,21,22]. No uniqueness theorem is imported from the authors' own prior work, and no ansatz is smuggled in via self-citation. The stated restriction to the canonical conformal Hamiltonian form (3) limits the scope of the title-level claim but does not make the derivation circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central geometric guarantee does not depend on fitted parameter values, so free parameters only enter through trained components that affect accuracy. The axioms are primarily domain restrictions about the class of dissipative Hamiltonian systems and background density theorems inherited from the cited literature.

free parameters (3)

Trainable coefficients a_l(ξ), b_l(ξ) and scaling allocations α_l(ξ) = trained on data; individual values not reported
These modulate the known H_q and H_p terms and split the conformal factor into per-layer weights. Theorem 1 holds for any values, so they only affect accuracy.
Trainable shear potentials δA_l and δB_l with FNO and FiLM network weights = trained on data; network weights not reported
Learned corrections in each shear block. They are central to empirical accuracy but irrelevant to the exact conformal symplecticity guarantee.
Architecture hyperparameters: L=12 blocks, FNO width 32, modes 8, layers 4, correction width 64 = chosen by hand; no hyperparameter search reported
Capacity settings that affect generalization and training cost, but not the structural guarantee.

assumptions (5)

domain assumption The target dynamics are conformal Hamiltonian systems of the exact canonical form (3), ż=J∇H(z)+γ diag(0,I_d)z, with constant γ.
Section 3.2, Eq. (3). The enforced factor e^{γt} is only the correct geometry for this class; other damping models do not satisfy φ*ω=e^{γt}ω.
domain assumption Hamiltonians are of separable natural form H(q,p)=H_q(q)+H_p(p), with H_q and H_p known and used directly in the shear potentials.
Section 4.3, Eqs. (9)-(10). The model relies on known H_q and H_p; q-p coupling would have to be learned by the correction potentials and is not tested.
domain assumption C^2 smoothness of potentials and compact forward containment of the flow: there is a compact K' with φ^s(K)⊂int K' for all s in [0,T].
Theorem 2 and Appendix A.2. Needed for the uniform approximation argument; standard for non-blow-up flows.
standard math Background density and approximation results: MLP density in C^1, Turaev's polynomial approximation, and SympFlow universality [6, Appendix F].
Invoked in the proof of Theorem 2; not reproved and not machine-checked.
domain assumption The 32x32 descriptor grid resolves the random Gaussian-process potentials (GP length scale at least 3.2 times the grid spacing).
Appendix B.1. If the descriptor could not distinguish Hamiltonians, the conditioner could not identify different systems.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoSynFlow: Conformal Symplectic Neural Flows for Cross-System Prediction of Dissipative Hamiltonian Dynamics." pith.science (2026). https://pith.science/paper/QC5Z5GGR

@misc{pith2026260800571,
  author       = {Pith},
  title        = {Pith review of: CoSynFlow: Conformal Symplectic Neural Flows for Cross-System Prediction of Dissipative Hamiltonian Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QC5Z5GGR}},
  note         = {Machine review of arXiv:2608.00571}
}

read the original abstract

Learning solution operators for differential equations is a central problem in scientific machine learning. However, many neural operator methods optimize prediction accuracy without explicitly enforcing the geometric structure of the dynamics. Structure-preserving models such as SympNets and Symplectic Neural Flows address this issue for conservative Hamiltonian systems by preserving the symplectic form. In dissipative Hamiltonian systems with conformal symplectic structure, however, the symplectic form evolves according to a conformal factor determined by the dissipation. We propose CoSynFlow, a conformal symplectic neural flow for learning continuous-time solution maps of dissipative Hamiltonian dynamics. CoSynFlow composes symplectic shear maps with explicit conformal scaling, preserving the conformal symplectic structure by construction. By conditioning it on a finite-dimensional Hamiltonian descriptor and the dissipation parameter, a single trained model predicts solution maps for unseen systems without retraining. CoSynFlow keeps the structure error at machine precision, attains the lowest long-horizon error, and admits physics-informed training.

Figures

Figures reproduced from arXiv: 2608.00571 by the authors.

**Figure 2.** Trajectories in the position plane [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** First position component q1 versus time for the same trajectories as [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: State error Ez averaged over t ∈ (0, T], on a logarithmic scale. The shaded region lies outside the training range |γ| ∈ [0.1, 0.4] and extends to three times its upper end [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: shows the resulting potentials. −3 −2 −1 0 1 2 3 q1 −3 −2 −1 0 1 2 3 q2 anisotropic oscillator −3 −2 −1 0 1 2 3 q1 −3 −2 −1 0 1 2 3 q2 coupled Duffing −3 −2 −1 0 1 2 3 q1 −3 −2 −1 0 1 2 3 q2 Mexican-hat −3 −2 −1 0 1 2 3 q1 −3 −2 −1 0 1 2 3 q2 quartic-coupled 0.0 1.2 2.…

**Figure 6.** Figure 6: Energy error EH(t) for CoSynFlow and the five baselines over 0 ≤ t ≤ 10T at |γ| = 0.1, averaged over 64 initial conditions, on a logarithmic scale. 0 5 10 15 20 25 30 35 40 t 10 −4 10 −3 10 −2 10 −1 10 0 CoSynFlow SympFlow Fixed Allocation No Correction B1, anisotropic…

**Figure 7.** Figure 7: Energy error EH(t) for CoSynFlow, Symplectic Flow and the two ablations, under the protocol of [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 10 canonical work pages

[17]

Deep gradient learning for efficient camouflaged object detection.Machine Intelligence Research, 20(1):92–108, 2023

Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection.Machine Intelligence Research, 20(1):92–108, 2023

work page 2023
[6]

Symplectic neural flows for modeling and discovery.arXiv preprint arXiv:2412.16787, 2024

Priscilla Canizares, Davide Murari, Carola-Bibiane Schönlieb, Ferdia Sherry, and Zakhar Shumaylov. Symplectic neural flows for modeling and discovery.arXiv preprint arXiv:2412.16787, 2024

arXiv 2024
[1]

Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

arXiv 1910
[2]

Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

arXiv 2010
[3]

Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

2023
[4]

Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

2019
[5]

Sympnets: Intrinsic structure- preserving symplectic networks for identifying hamiltonian systems.Neural Networks, 132:166–179, 2020

Pengzhan Jin, Zhen Zhang, Aiqing Zhu, Yifa Tang, and George Em Karniadakis. Sympnets: Intrinsic structure- preserving symplectic networks for identifying hamiltonian systems.Neural Networks, 132:166–179, 2020

work page 2020
[7]

Conformal hamiltonian systems.Journal of Geometry and Physics, 39(4):276–300, 2001

Robert McLachlan and Matthew Perlmutter. Conformal hamiltonian systems.Journal of Geometry and Physics, 39(4):276–300, 2001

work page 2001

Show all 22 references

[8]

Physics-informed neural operator for learning partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

2024
[9]

Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366, 2024

Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366, 2024

2024 arXiv
[10]

Lagrangian neural networks.arXiv preprint arXiv:2003.04630, 2020

Miles Cranmer, Sam Greydanus, Stephan Hoyer, Peter Battaglia, David Spergel, and Shirley Ho. Lagrangian neural networks.arXiv preprint arXiv:2003.04630, 2020

2003 arXiv
[11]

Symplectic ode-net: Learning hamiltonian dynamics with control.arXiv preprint arXiv:1909.12077, 2019

Yaofeng Desmond Zhong, Biswadip Dey, and Amit Chakraborty. Symplectic ode-net: Learning hamiltonian dynamics with control.arXiv preprint arXiv:1909.12077, 2019

1909 arXiv
[12]

Learning poisson systems and trajectories of autonomous systems via poisson neural networks.IEEE Transactions on Neural Networks and Learning Systems, 34(11):8271–8283, 2022

Pengzhan Jin, Zhen Zhang, Ioannis G Kevrekidis, and George Em Karniadakis. Learning poisson systems and trajectories of autonomous systems via poisson neural networks.IEEE Transactions on Neural Networks and Learning Systems, 34(11):8271–8283, 2022

2022
[13]

Neural symplectic form: Learning hamiltonian equations on general coordinate systems.Advances in Neural Information Processing Systems, 34:16659–16670, 2021

Yuhan Chen, Takashi Matsubara, and Takaharu Yaguchi. Neural symplectic form: Learning hamiltonian equations on general coordinate systems.Advances in Neural Information Processing Systems, 34:16659–16670, 2021. 19

2021
[14]

Dissipative hamiltonian neural networks: Learning dissipative and conservative dynamics separately.arXiv preprint arXiv:2201.10085, 2022

Andrew Sosanya and Sam Greydanus. Dissipative hamiltonian neural networks: Learning dissipative and conservative dynamics separately.arXiv preprint arXiv:2201.10085, 2022

2022 arXiv
[15]

Dissipative symoden: Encoding hamiltonian dynamics with dissipation and control into deep learning.arXiv preprint arXiv:2002.08860, 2020

Yaofeng Desmond Zhong, Biswadip Dey, and Amit Chakraborty. Dissipative symoden: Encoding hamiltonian dynamics with dissipation and control into deep learning.arXiv preprint arXiv:2002.08860, 2020

2002 arXiv
[16]

Port-hamiltonian neural networks for learning explicit time-dependent dynamical systems.Physical Review E, 104(3):034312, 2021

Shaan A Desai, Marios Mattheakis, David Sondak, Pavlos Protopapas, and Stephen J Roberts. Port-hamiltonian neural networks for learning explicit time-dependent dynamical systems.Physical Review E, 104(3):034312, 2021

2021
[18]

Elena Celledoni, Brynjulf Owren, Chong Shen, Baige Xu, and Takaharu Yaguchi. Uepi: Universal energy- behavior-preserving integrators for energy conservative/dissipative differential equations.Advances in Neural Information Processing Systems, 38:43717–43751, 2026

2026
[19]

Gfinns: Generic formalism informed neural net- works for deterministic and stochastic dynamical systems.Philosophical Transactions of the Royal Society A, 380(2229):20210207, 2022

Zhen Zhang, Yeonjong Shin, and George Em Karniadakis. Gfinns: Generic formalism informed neural net- works for deterministic and stochastic dynamical systems.Philosophical Transactions of the Royal Society A, 380(2229):20210207, 2022

2022
[20]

Polynomial approximations of symplectic dynamics and richness of chaos in non-hyperbolic area-preserving maps.Nonlinearity, 16(1):123–135, 2003

Dmitry Turaev. Polynomial approximations of symplectic dynamics and richness of chaos in non-hyperbolic area-preserving maps.Nonlinearity, 16(1):123–135, 2003

2003
[21]

Approximation theory of the mlp model in neural networks.Acta numerica, 8:143–195, 1999

Allan Pinkus. Approximation theory of the mlp model in neural networks.Acta numerica, 8:143–195, 1999

1999
[22]

Approximation capabilities of multilayer feedforward networks.Neural networks, 4(2):251–257, 1991

Kurt Hornik. Approximation capabilities of multilayer feedforward networks.Neural networks, 4(2):251–257, 1991. 20

1991

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [17]

Deep gradient learning for efficient camouflaged object detection.Machine Intelligence Research, 20(1):92–108, 2023

Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection.Machine Intelligence Research, 20(1):92–108, 2023

work page 2023

[2] [6]

Symplectic neural flows for modeling and discovery.arXiv preprint arXiv:2412.16787, 2024

Priscilla Canizares, Davide Murari, Carola-Bibiane Schönlieb, Ferdia Sherry, and Zakhar Shumaylov. Symplectic neural flows for modeling and discovery.arXiv preprint arXiv:2412.16787, 2024

arXiv 2024

[3] [1]

Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019

arXiv 1910

[4] [2]

Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

arXiv 2010

[5] [3]

Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

2023

[6] [4]

Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

2019

[7] [5]

Sympnets: Intrinsic structure- preserving symplectic networks for identifying hamiltonian systems.Neural Networks, 132:166–179, 2020

Pengzhan Jin, Zhen Zhang, Aiqing Zhu, Yifa Tang, and George Em Karniadakis. Sympnets: Intrinsic structure- preserving symplectic networks for identifying hamiltonian systems.Neural Networks, 132:166–179, 2020

work page 2020

[8] [7]

Conformal hamiltonian systems.Journal of Geometry and Physics, 39(4):276–300, 2001

Robert McLachlan and Matthew Perlmutter. Conformal hamiltonian systems.Journal of Geometry and Physics, 39(4):276–300, 2001

work page 2001