Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Reduced-order modeling of Hamiltonian dynamics based on symplectic neural networks

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Composing HénonNet nonlinear layers with G-reflector linear layers gives an exactly symplectic encoder–decoder for Hamiltonian model reduction, and the resulting latent flow map predicts trajectories and preserves energy over long horizons.

desk verdict The HénonNet-based symplectic autoencoder idea is novel and structurally sound, but the numerical sections are internally inconsistent and the experimental claims as printed cannot be reproduced. read the letter →

arxiv 2508.11911 v2 pith:O7E3ULPW submitted 2025-08-16 math.NA cs.LGcs.NAphysics.comp-ph

Yongsheng Chen , Wei Guo , Qi Tang , Xinghui Zhong This is my paper
classification math.NAcs.LGcs.NAphysics.comp-ph MSC 37M1565P1068T07
keywords HamiltoniansystemsmodelorderreductionsymplecticneuralnetworksHénonNetsG-reflectorsstructurepreservationlong-horizonpredictionnonlinearSchrödingerequation
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 tries to establish that reduced-order models of high-dimensional Hamiltonian systems can be built as a single end-to-end neural architecture that is exactly symplectic by construction: a HénonNet-based autoencoder, optionally augmented with G-reflector layers, maps between full and latent phase spaces, and a HénonNet flow map advances the latent state. Since every component is symplectic, the whole reduction is exactly symplectic no matter what weights are learned. This geometric guarantee is the reason the method is claimed to reconstruct trajectories accurately, predict stably beyond the training horizon, and keep the Hamiltonian nearly conserved. The numerical experiments on linear wave, parametric wave, and nonlinear Schrödinger systems support that claim and show HénonNet-based embeddings reduce reconstruction error by two to three orders of magnitude over linear symplectic embeddings.

What carries the argument

The load-bearing identity is the Hénon layer $\mathcal{H}(V,\eta) = [H(V,\eta)]^4$, built from the elementary symplectic map $H(V,\eta)(x,y) = (y+\eta,\, x+\nabla V(y))$, which has an analytic inverse; composing such layers gives the HénonNet, an invertible symplectic network with universal approximation for symplectic diffeomorphisms. The optional G-reflector $G = I_{2n} + \beta u u^T J_{2n}$ supplies a parameter-efficient linear symplectic map, and the composite $\sigma = H \circ G \circ \iota$ is the exact symplectic embedding used for the autoencoder. The same HénonNet construction is reused as the latent flow map, so the entire ROM is a composition of symplectic maps and hence exactly s

What would settle it

Train the composite symplectic autoencoder to approximate a known non-surjective symplectic embedding on a compact domain and measure the uniform error $\sup_z \|\Phi(z) - \sigma(z)\|$ as network depth and width increase; if the error levels off above zero, the universality claim fails. Alternatively, if a Hamiltonian ROM is built for a system whose solution manifold is not tangent to any symplectic subspace, persistently large reconstruction error would show the symplectic-embedding premise is violated.

Watch

Extended reading notes

Core claim

The central object is the composite symplectic embedding $\sigma = H \circ G \circ \iota$ from latent coordinates in $\mathbb{R}^{2k}$ into the full phase space $\mathbb{R}^{2n}$, where $\iota$ is the canonical inclusion, $G$ is a product of G-reflectors ($G = I_{2n} + \beta u u^T J_{2n}$), and $H$ is a HénonNet. The encoder is $f_{\mathrm{enc}} = \tau \circ G_{\mathrm{full}} \circ H_{\mathrm{full}}$ with a symplectic truncation $\tau$, and the decoder is the inverse composition $f_{\mathrm{dec}} = H_{\mathrm{full}}^{-1} \circ G_{\mathrm{full}}^{-1} \circ \iota$; the latent flow map is itself a HénonNet. The paper's claim is that $f_{\mathrm{dec}} \circ f_{\mathrm{flow}} \circ f_{\mathrm{enc

Load-bearing premise

The method depends on the assumption that the low-dimensional solution manifold can be embedded in the full phase space by an exactly symplectic map that the HénonNet-plus-G-reflector network can approximate arbitrarily well; the paper states this as a theorem but does not prove it.

Editorial extensions

If this is right

  • Because the encoder-decoder and latent flow are exactly symplectic by construction, the reduced model inherits phase-space volume preservation and the non-dissipative character of Hamiltonian flows, supporting stable predictions beyond the training horizon.
  • The reported two-to-three order-of-magnitude drop in reconstruction error versus linear symplectic embeddings indicates that nonlinear symplectic manifolds, not linear subspaces, are needed to represent Hamiltonian solution manifolds.
  • The Hamiltonian term in the unified loss keeps the discrete Hamiltonian almost constant, with deviations near machine precision in the nonlinear Schrödinger example.
  • The framework handles parametric dependence and nonlinearity in a single architecture, so the same recipe applies across linear wave, parametric wave, and nonlinear Schrödinger equations.
  • Exact symplecticity is guaranteed independently of the trained weights, so the structure-preservation claim does not depend on how well the loss is optimized.

Reading between the lines

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

  • If the latent map is exactly symplectic, the reduced model is itself Hamiltonian with respect to some latent Hamiltonian; identifying that latent Hamiltonian from the trained flow would give an interpretable energy function for the ROM, which the paper does not do.
  • Theorem 2, the universal approximation statement for non-surjective symplectic embeddings, is asserted without proof; settling it with a rigorous proof or a counterexample would determine whether the architecture can always reach the true latent manifold.
  • A direct way to isolate the benefit of symplecticity would be to compare this architecture with an identical one in which one component is deliberately made non-symplectic; the performance gap would quantify how much of the gain is geometric rather than due to model capacity.
  • The composite embedding construction could plausibly extend to non-canonical or stochastic Hamiltonian systems by replacing $J_{2n}$ with the appropriate structure matrix, a direction the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a unified symplectic reduced-order modeling (ROM) framework for Hamiltonian systems. The architecture combines a HenonNet-based autoencoder, optionally augmented with G-reflector layers, to construct an exactly invertible encoder-decoder pair whose decoding map is a symplectic embedding, together with a HenonNet flow map for the latent dynamics. Since symplectic maps are closed under composition, the reduced-order model is exactly symplectic by construction. The paper claims a universal approximation theorem for symplectic embeddings, a unified multi-objective loss for reconstruction, prediction, and Hamiltonian conservation, and numerical validation on the linear wave equation, a parametric linear wave equation, and the nonlinear Schrodinger equation, reporting accurate reconstruction, long-horizon extrapolation, and Hamiltonian preservation.

Significance. The architectural idea is natural and the exact symplecticity-by-composition argument is sound and independent of training. The explicit invertibility of HenonNet layers and the G-reflector parameterization of symplectic linear maps are useful ingredients. If the numerical validation were correct, this would be a valuable contribution to structure-preserving ROM. However, the printed semi-discrete equations in Section 5 contain mismatches that make the experiments unreproducible, and the stated universal approximation theorem is unproved. In addition, the reported Hamiltonian preservation is partly a fitted property because the loss explicitly penalizes Hamiltonian drift. These issues are load-bearing for the empirical claims and need to be fixed before the paper can be accepted.

major comments (4)
  1. [Section 5.1, Eqs. (42)-(44)] Eq. (42) states dz/dt = J grad H = J L z with L = diag(I_N, omega^2 D_xx). For H in (43), direct differentiation gives grad_p H = Delta x p and grad_q H = -omega^2 Delta x D_xx q (boundary terms aside), hence J grad H = (Delta x p, omega^2 Delta x D_xx q). This is not equal to J L z = (omega^2 D_xx p, -q), nor to the intended discretization q_t = p, p_t = omega^2 q_xx. The same L is repeated in Example 5.2, Eq. (53). Since the Stormer-Verlet reference data and Hamiltonian-preservation metrics are generated from these equations, the experiments cannot be reproduced as printed.
  2. [Section 5.3, Eqs. (57)-(60)] In Example 5.3, Eq. (57) writes dz/dt = J grad H = J L z + J g(z), with L = diag(D_xx, D_xx) and g defined in (59). The function g contains (q_i^2+p_i^2)q_i and (q_i^2+p_i^2)p_i without the parameter epsilon, so the nonlinearity does not depend on epsilon. The discrete Hamiltonian (60) contains terms (q_{i-1}q_i - q_i^2)^2/Delta x^2 and epsilon/4(p_i^2+q_i^2), neither of which discretizes the quartic potential epsilon/2(q^2+p^2)^2 in Eq. (56). The reference data and Hamiltonian-preservation results are therefore for a different system than the stated nonlinear Schrodinger equation.
  3. [Section 3.4, Theorem 2] Theorem 2 is stated without proof. The cited Turaev result concerns approximation of symplectic diffeomorphisms by Henon-type maps; the extension to non-surjective embeddings Phi: R^{2k}->R^{2n} on a compact domain is not automatic. This gap is load-bearing because Theorem 2 underpins the claim that the architecture can represent the latent submanifold.
  4. [Section 4.3, Eqs. (28)-(29), (33)] The Hamiltonian-conservation results in Figures 3, 5, and 7 are not independent checks of structure preservation. The total loss (29) includes L_ham, which explicitly penalizes deviations of H(f_dec(f_flow(f_enc(x)))) from H(x0) during training. The observed near-constancy of the discrete Hamiltonian is therefore partly a fitted property, not a consequence of symplecticity alone. To support the structure-preservation claim, report Hamiltonian error for a model trained without L_ham, or otherwise separate the symplecticity guarantee from the fitted conservation.
minor comments (4)
  1. [Section 4.3, Eqs. (28) and (33)] The notation for the Hamiltonian is inconsistent: Eq. (28) uses a tilde over H for the original Hamiltonian, while Eq. (33) uses H without redefinition. Please standardize.
  2. [Section 3.1, Eqs. (10)-(12)] The symbol H is overloaded: it denotes both the basic Henon mapping and the fourth-power Henon layer in Eq. (11). Use distinct symbols, for example h for the basic map and curly H for the layer.
  3. [Section 4.1, Eqs. (19)-(21)] The encoder f_enc = tau circle G circle H is a left inverse of the symplectic embedding f_dec, not a symplectic map in the usual same-dimension sense. The statement that the encoder preserves symplectic structure should be qualified; tau is co-symplectic on the embedded submanifold.
  4. [Section 5, reference data] The label Exact for Stormer-Verlet reference solutions is misleading, since these are numerical solutions. Use Reference or SV instead.

Circularity Check

1 steps flagged · score 6.0 of 10

Hamiltonian conservation is partly enforced by the training loss (Lham), so the reported conservation evidence is a fitted-input-called-prediction; the exact-symplectic construction and trajectory extrapolation are not circular.

  1. fitted input called prediction [Section 4.3, Eqs. (28)-(29) and (33); validation presented in Section 5 (Figs. 3, 5, 7)]
    "To explicitly maintain Hamiltonian invariance, particularly crucial for complex nonlinear systems, we introduce an additional Hamiltonian conservation constraint: Lham = ... where ˜H denotes the original Hamiltonian of the system state x. This constraint ensures Hamiltonian conservation during both reconstruction and latent space dynamics evolution. The complete loss function is therefore expressed as: Ltotal = λ1Lrom + λ2Lham ... while the Hamiltonian conservation loss takes the form: Lham = ... ||H(fdec ◦ fenc(xi))−H(x0)||2 + ... ||H(fdec ◦ fflow ◦ fenc(ˆxi+j))−H(x0)||2."

    Eqs. (28)-(29) and (33) directly add ||H(fdec(fenc(xi)))−H(x0)||² and ||H(fdec(fflow(fenc(xi))))−H(x0)||² to the objective that is minimized. The paper then presents flat discrete-Hamiltonian curves in Figures 3, 5, and 7 as evidence that the method 'preserves the Hamiltonian structure'. That evidence is not an independent prediction: the optimizer was explicitly trained to drive these deviations to zero on the training trajectories. The Hamiltonian-conservation headline is therefore partly a restatement of the loss function, not a free result. This is distinct from the trajectory reconstructions and the beyond-training-horizon extrapolation, which are not directly minimized by Lham and retain independent content.

full rationale

The exact-symplectic mechanism is not circular: Hénon layers, G-reflectors, and their compositions are symplectic by construction, independent of learned weights. Linear symplectic lifts and G-reflector parameterizations follow from external references [34,38], not from the paper's conclusions. Self-citations to prior HénonNet work supply the architecture, but its symplecticity is verified by definition, so no load-bearing self-citation circle appears. Theorem 2 (Section 3.4) is an omitted proof rather than a circularity: universality for symplectic embeddings is asserted from Turaev's symplectic-diffeomorphism result without proving the non-surjective embedding case. Section 5's discrete systems (Eqs. (42)-(44), (51)-(53), (57)-(60)) appear internally inconsistent with the stated Hamiltonians and the continuous PDEs; these are reproducibility/correctness concerns, not derivation-circle concerns. The only construction-based circularity is the Hamiltonian-conservation loss: 'accurate Hamiltonian preservation' is a stated central result but is directly minimized in Lham, so its reported accuracy is partly self-fulfilling. Score 6 reflects this one central fitted-input-called-prediction element while the remaining claims (exact symplecticity, latent extrapolation) are not circular.

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

The method introduces no new physical entities. The load-bearing assumptions are the cited universal-approximation and factorization theorems, the low-dimensional symplectic-manifold hypothesis, and the implicit extension step in Theorem 2. Free parameters are the hand-chosen hyperparameters; neural weights are fit to data but not enumerated.

free parameters (6)
  • latent dimension 2k = 10 (Ex. 5.1, 5.2), 12 (Ex. 5.3)
    Chosen by hand as the reduced dimension; the paper provides no systematic selection criterion.
  • loss weight lambda2 = 0.01 (Ex. 5.1, 5.2), 0.05 (Ex. 5.3)
    Hand-tuned to balance the Hamiltonian-conservation penalty against reconstruction/prediction loss; no ablation is reported.
  • noise amplitude sigma = 1e-3
    Hand-set amplitude for Gaussian noise injection in training.
  • G-reflector layer count = 10 (autoencoder); 100 (linear baseline)
    Hand-chosen; the paper notes more layers offered limited improvement.
  • HénonNet width/depth = [300]x2, [512]x2, [128]x2, [30]x2, [32]x2
    Hand-selected per example; no sensitivity analysis.
  • multi-step unrolling length M = not reported
    Training uses M-step auto-regressive unrolling (Eq. 32) but M is never specified, so the exact training objective is underdetermined.
assumptions (6)
  • domain assumption HénonNets can approximate arbitrary symplectic diffeomorphisms on compact sets (Turaev 2002).
    Invoked in Sections 3.2 and 3.4 as the basis for Theorem 2 and for replacing learned embeddings with the fixed inclusion iota; if this theorem does not extend to the composition sigma = H ∘ G ∘ iota, the expressive-power guarantee fails.
  • standard math Every symplectic matrix factors into at most 4n G-reflectors (Mackey & Mackey 2003).
    Used in Section 3.3 to justify the completeness of the G-reflector parameterization of Sp(2n).
  • standard math The map S -> S iota from Sp(2n) to SpSt(2k,2n) is surjective (Bendokat & Zimmermann 2021).
    Used in Section 3.3 to assert every symplectic Stiefel embedding can be represented by a symplectic matrix acting on iota.
  • domain assumption The solution manifold of each parametric Hamiltonian system lies near a low-dimensional symplectic submanifold.
    Core ROM premise stated in Section 1; if false, the autoencoder cannot compress the dynamics without large error.
  • ad hoc to paper A symplectic embedding of a compact domain extends to a symplectic diffeomorphism of R^2n in the HénonNet-representable class.
    Theorem 2 requires any target embedding Phi to be approximable by H ∘ G ∘ iota; the paper states this without proof and does not cite a specific extension theorem, making this the weak link in the universality claim.
  • domain assumption The trained neural network generalizes from the sampled parameter and trajectory data to unseen parameters and times.
    Standard ML generalization assumption; the representative test parameters (e.g., a0=8.939 in Ex. 5.1) are claimed to lie outside the training range, but no statistical generalization guarantee is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reduced-order modeling of Hamiltonian dynamics based on symplectic neural networks." pith.science (2026). https://pith.science/paper/O7E3ULPW

@misc{pith2026250811911,
  author       = {Pith},
  title        = {Pith review of: Reduced-order modeling of Hamiltonian dynamics based on symplectic neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O7E3ULPW}},
  note         = {Machine review of arXiv:2508.11911}
}
read the original abstract

We introduce a novel data-driven symplectic induced-order modeling (ROM) framework for high-dimensional Hamiltonian systems that unifies latent-space discovery and dynamics learning within a single, end-to-end neural architecture. The encoder-decoder is built from Henon neural networks (HenonNets) and may be augmented with linear SGS-reflector layers. This yields an exact symplectic map between full and latent phase spaces. Latent dynamics are advanced by a symplectic flow map implemented as a HenonNet. This unified neural architecture ensures exact preservation of the underlying symplectic structure at the reduced-order level, significantly enhancing the fidelity and long-term stability of the resulting ROM. We validate our method through comprehensive numerical experiments on canonical Hamiltonian systems. The results demonstrate the method's capability for accurate trajectory reconstruction, robust predictive performance beyond the training horizon, and accurate Hamiltonian preservation. These promising outcomes underscore the effectiveness and potential applicability of our symplectic ROM framework for complex dynamical systems across a broad range of scientific and engineering disciplines.

Figures

Figures reproduced from arXiv: 2508.11911 by the authors.

Figure 1
Figure 1. Schematic illustration of the symplectic neural network architecture for ROM of Hamiltonian systems. We [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. ROM prediction in Example 5.1 for (a0, x0) = (8.939, 0.079): (a) latent space dynamics; (b)–(d) recon￾structed wave solutions at t = 0, t = 12, and t = 24. (a) Hamiltonian (b) Hamiltonian deviation [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Hamiltonian conservation results in Example [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: ROM prediction in Example 5.2 with κ(ω) = 0.01736: (a) latent space dynamics; reconstructed solutions at (b) t = 0, (c) t = 1.5, and (d) t = 3 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Hamiltonian conservation results for a single trajectory in Example [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: ROM prediction in Example 5.3 with ϵ = 1.007858: (a) latent space dynamics, and reconstructed solutions at times (b) t = 0, (c) t = 10, and (d) t = 20. (a) Hamiltonian (b) Hamiltonian deviation [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Hamiltonian conservation results for a single trajectory in Example [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Hyperbolicity-Preserving Stochastic Galerkin Methods for Conservation Laws Based on Associative Truncated Products on Polynomial Spaces

    math.NA 2026-06 unverdicted novelty 8.0 of 10

    Develops associative truncated products on polynomial spaces to construct hyperbolicity-preserving stochastic Galerkin discretizations for conservation laws, with consistency results and applications to Euler equations.

  2. Stability threshold of Couette flow for Boussinesq equations in $\mathbb{R}^2$

    math.AP 2025-08 unverdicted novelty 6.0 of 10

    The stability threshold for the 2D Boussinesq Couette flow on the whole plane is at most {1/3+, 2/3+}, extending the known periodic-channel result.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages · cited by 2 Pith papers

  1. [1]

    L. Peng, K. Mohseni, Symplectic model reduction of hamiltonian systems, SIAM Journal on Scientific Computing 38 (1) (2016) A1–A27

    work page 2016

  2. [2]

    Leimkuhler, S

    B. Leimkuhler, S. Reich, Simulating hamiltonian dynamics, Cambridge university press, 2004. 22

    work page 2004

  3. [3]

    Hairer, M

    E. Hairer, M. Hochbruck, A. Iserles, C. Lubich, Geometric numerical integration, Oberwolfach Reports 3 (1) (2006) 805–882

    work page 2006

  4. [4]

    P. J. Morrison, Hamiltonian description of the ideal fluid, Reviews of modern physics 70 (2) (1998) 467

    work page 1998

  5. [5]

    Lubich, C

    C. Lubich, C. Lubich, C. Lubich, C. Lubich, From quantum to classical molecular dynamics: reduced models and numerical analysis, Vol. 12, European Mathematical Society Z¨ urich, 2008

    work page 2008

  6. [6]

    Benner, S

    P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction methods for parametric dynamical systems, SIAM review 57 (4) (2015) 483–531

    2015

  7. [7]

    Quarteroni, A

    A. Quarteroni, A. Manzoni, F. Negri, Reduced basis methods for partial differential equations: an introduction, Vol. 92, Springer, 2015

    work page 2015

  8. [8]

    Rouleau, J.-F

    L. Rouleau, J.-F. De¨ u, A. Legay, A comparison of model reduction techniques based on modal projection for structures with frequency-dependent damping, Mechanical Systems and Signal Processing 90 (2017) 110–125

    work page 2017

Show all 48 references
  1. [9]

    Bonvin, D

    D. Bonvin, D. Mellichamp, A unified derivation and critical review of modal approaches to model reduction, International Journal of Control 35 (5) (1982) 829–848

    1982

  2. [10]

    Moore, Principal component analysis in linear systems: Controllability, observability, and model reduction, IEEE transactions on automatic control 26 (1) (1981) 17–32

    B. Moore, Principal component analysis in linear systems: Controllability, observability, and model reduction, IEEE transactions on automatic control 26 (1) (1981) 17–32

    1981

  3. [11]

    P. A. Parrilo, S. Lall, F. Paganini, G. C. Verghese, B. C. Lesieutre, J. E. Marsden, Model reduction for analysis of cascading failures in power systems, in: Proceedings of the 1999 American Control Conference (Cat. No. 99CH36251), Vol. 6, IEEE, 1999, pp. 4208–4212

    1999

  4. [12]

    Kunisch, S

    K. Kunisch, S. Volkwein, Galerkin proper orthogonal decomposition methods for a general equation in fluid dynamics, SIAM Journal on Numerical analysis 40 (2) (2002) 492–515

    2002

  5. [13]

    Odabasioglu, M

    A. Odabasioglu, M. Celik, L. T. Pileggi, Prima: Passive reduced-order interconnect macro- modeling algorithm, The Best of ICCAD: 20 Years of Excellence in Computer-Aided Design (2003) 433–450

    2003

  6. [14]

    R. W. Freund, Model reduction methods based on krylov subspaces, Acta Numerica 12 (2003) 267–319

    2003

  7. [15]

    Bai, Krylov subspace techniques for reduced-order modeling of large-scale dynamical sys- tems, Applied numerical mathematics 43 (1-2) (2002) 9–44

    Z. Bai, Krylov subspace techniques for reduced-order modeling of large-scale dynamical sys- tems, Applied numerical mathematics 43 (1-2) (2002) 9–44

    2002

  8. [16]

    Chaturantabut, D

    S. Chaturantabut, D. C. Sorensen, Nonlinear model reduction via discrete empirical interpo- lation, SIAM Journal on Scientific Computing 32 (5) (2010) 2737–2764

    2010

  9. [17]

    Drohmann, B

    M. Drohmann, B. Haasdonk, M. Ohlberger, Reduced basis approximation for nonlinear parametrized evolution equations based on empirical operator interpolation, SIAM Journal on Scientific Computing 34 (2) (2012) A937–A969

    2012

  10. [18]

    Drmac, S

    Z. Drmac, S. Gugercin, A new selection operator for the discrete empirical interpolation method—improved a priori error bound and extensions, SIAM Journal on Scientific Com- puting 38 (2) (2016) A631–A648. 23

    2016

  11. [19]

    L. Peng, K. Mohseni, An online manifold learning approach for model reduction of dynamical systems, SIAM Journal on Numerical Analysis 52 (4) (2014) 1928–1952

    2014

  12. [20]

    Mill´ an, M

    D. Mill´ an, M. Arroyo, Nonlinear manifold learning for model reduction in finite elastodynamics, Computer Methods in Applied Mechanics and Engineering 261 (2013) 118–131

    2013

  13. [21]

    F. J. Gonzalez, M. Balajewicz, Deep convolutional recurrent autoencoders for learning low- dimensional feature dynamics of fluid systems, arXiv preprint arXiv:1808.01346 (2018)

    2018 arXiv

  14. [22]

    Pagani, A

    S. Pagani, A. Manzoni, K. Carlberg, Statistical closure modeling for reduced-order models of stationary systems by the romes method, International Journal for Uncertainty Quantification 12 (1)

  15. [23]

    P. Pant, R. Doshi, P. Bahl, A. Barati Farimani, Deep learning for reduced order modelling and efficient temporal evolution of fluid simulations, Physics of Fluids 33 (10)

  16. [24]

    E. Qian, B. Kramer, B. Peherstorfer, K. Willcox, Lift & learn: Physics-informed machine learning for large-scale nonlinear dynamical systems, Physica D: Nonlinear Phenomena 406 (2020) 132401

    2020

  17. [25]

    W. Chen, Q. Wang, J. S. Hesthaven, C. Zhang, Physics-informed machine learning for reduced- order modeling of nonlinear problems, Journal of computational physics 446 (2021) 110666

    2021

  18. [26]

    A. J. Linot, J. W. Burby, Q. Tang, P. Balaprakash, M. D. Graham, R. Maulik, Stabilized neural ordinary differential equations for long-time forecasting of dynamical systems, Journal of Computational Physics 474 (2023) 111838

    2023

  19. [27]

    D. A. Serino, A. Alvarez Loya, J. W. Burby, I. G. Kevrekidis, Q. Tang, Fast-slow neural net- works for learning singularly perturbed dynamical systems, Journal of Computational Physics (2025) 114090

    2025

  20. [28]

    Alvarez Loya, D

    A. Alvarez Loya, D. A. Serino, J. Burby, Q. Tang, Structure-preserving neural ordinary differ- ential equations for stiff systems, arXiv preprint arXiv:2503.01775 (2025)

    2025

  21. [29]

    Buchfink, S

    P. Buchfink, S. Glas, B. Haasdonk, Symplectic model reduction of hamiltonian systems on nonlinear manifolds and approximation with weakly symplectic autoencoder, SIAM Journal on Scientific Computing 45 (2) (2023) A289–A311

    2023

  22. [30]

    Sharma, H

    H. Sharma, H. Mu, P. Buchfink, R. Geelen, S. Glas, B. Kramer, Symplectic model reduction of hamiltonian systems using data-driven quadratic manifolds, Computer Methods in Applied Mechanics and Engineering 417 (2023) 116402

    2023

  23. [31]

    J. W. Burby, Q. Tang, R. Maulik, Fast neural poincar´ e maps for toroidal magnetic fields, Plasma Physics and Controlled Fusion 63 (2) (2020) 024001

    2020

  24. [32]

    Duruisseaux, J

    V. Duruisseaux, J. W. Burby, Q. Tang, Approximation of nearly-periodic symplectic maps via structure-preserving neural networks, Scientific reports 13 (1) (2023) 8351

    2023

  25. [33]

    D. S. Mackey, N. Mackey, F. Tisseur, G-reflectors: Analogues of householder transformations in scalar product spaces, Linear algebra and its applications 385 (2004) 187–213

    2004

  26. [34]

    D. S. Mackey, N. Mackey, On the determinant of symplectic matrices, Manchester Centre for Computational Mathematics Manchester, England, 2003. 24

    2003

  27. [35]

    Turaev, Polynomial approximations of symplectic dynamics and richness of chaos in non- hyperbolic area-preserving maps, Nonlinearity 16 (1) (2002) 123

    D. Turaev, Polynomial approximations of symplectic dynamics and richness of chaos in non- hyperbolic area-preserving maps, Nonlinearity 16 (1) (2002) 123

    2002

  28. [36]

    Kobyzev, S

    I. Kobyzev, S. J. Prince, M. A. Brubaker, Normalizing flows: An introduction and review of current methods, IEEE transactions on pattern analysis and machine intelligence 43 (11) (2020) 3964–3979

    2020

  29. [37]

    E. G. Drimalas, F. Fraschetti, C.-K. Huang, Q. Tang, Symplectic neural network and its applications to charged particle dynamics in electromagnetic fields, Physics of Plasmas, in revision (2025)

    2025

  30. [38]

    Bendokat, R

    T. Bendokat, R. Zimmermann, The real symplectic stiefel and grassmann manifolds: Metrics, geodesics and applications, arXiv preprint arXiv:2108.12447 (2021)

    2021 arXiv

  31. [39]

    Simpson, N

    T. Simpson, N. Dervilis, E. Chatzi, Machine learning approach to model order reduction of nonlinear systems via autoencoder and lstm networks, Journal of Engineering Mechanics 147 (10) (2021) 04021061

    2021

  32. [40]

    S. B. Reddy, A. R. Magee, R. K. Jaiman, J. Liu, W. Xu, A. Choudhary, A. Hussain, Reduced order model for unsteady fluid flows via recurrent neural networks, in: International Conference on Offshore Mechanics and Arctic Engineering, Vol. 58776, American Society of Mechanical En...

    2019

  33. [41]

    Brandstetter, D

    J. Brandstetter, D. E. Worrall, M. Welling, Message passing neural pde solvers, in: Interna- tional Conference on Learning Representations, 2021

    2021

  34. [42]

    Stachenfeld, D

    K. Stachenfeld, D. B. Fielding, D. Kochkov, M. Cranmer, T. Pfaff, J. Godwin, C. Cui, S. Ho, P. Battaglia, A. Sanchez-Gonzalez, Learned coarse models for efficient turbulence simulation, in: International Conference on Learning Representations, 2022

    2022

  35. [43]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, P. Battaglia, Learning mesh-based simulation with graph networks, in: International Conference on Learning Representations, 2021

    2021

  36. [44]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, P. Battaglia, Learning to sim- ulate complex physics with graph networks, in: International conference on machine learning, PMLR, 2020, pp. 8459–8468

    2020

  37. [45]

    Clevert, T

    D.-A. Clevert, T. Unterthiner, S. Hochreiter, Fast and accurate deep network learning by exponential linear units (elus), in: Proceedings of the International Conference on Learning Representations (ICLR), 2016

    2016

  38. [46]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: International Conference on Learning Representations, 2018

    2018

  39. [47]

    Sulem, P.-L

    C. Sulem, P.-L. Sulem, The nonlinear Schr¨ odinger equation: self-focusing and wave collapse, Vol. 139, Springer Science & Business Media, 2007

    2007

  40. [48]

    Huang, Q

    C. Huang, Q. Tang, Y. Batygin, O. Beznosov, J. Burby, A. Kim, S. Kurennoy, T. Kwan, H. Rakotoarivelo, Symplectic neural surrogate models for beam dynamics, in: Journal of Physics: Conference Series, Vol. 2687, IOP Publishing, 2024, p. 062026. 25

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.