Pith. sign in

REVIEW 3 major objections 6 minor 92 references

Latent Dynamics Graph Convolutional Networks for model order reduction of parameterized time-dependent PDEs

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A latent-dynamics graph convolutional network can reduce time-dependent parameterized PDEs to a few latent coordinates, reconstructing full fields on unstructured meshes with accuracy that beats a graph autoencoder at half the parameters.

desk verdict Solid empirical MOR paper combining LDNet and GCA-ROM; the UAT is proven for a simpler model than the one benchmarked, so the math claim is oversold but the benchmarks stand. read the letter →

arxiv 2601.11259 v1 pith:O4X2TWJU submitted 2026-01-16 cs.LG cs.NAmath.NA

Lorenzo Tomada , Federico Pichi , Gianluigi Rozza This is my paper
classification cs.LGcs.NAmath.NA
keywords modelorderreductionlatentdynamicsgraphconvolutionalnetworkneuralODEparameterizedPDEsunstructuredmeshesbifurcationdetectionzero-shotinterpolation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces LD-GCN, an encoder-free architecture that learns a low-dimensional latent dynamical system for a parameterized PDE and decodes it back onto a fixed unstructured mesh through a graph convolutional network. The central claim is that coupling a neural ODE in the latent space with a graph decoder yields accurate, temporally and geometrically consistent solutions, outperforming the graph convolutional autoencoder ROM while using roughly 50% fewer trainable parameters and matching the accuracy of the earlier meshless latent-dynamics network on a lid-driven cavity benchmark. The authors support the approach with a universal approximation theorem for encoder-free architectures, and they report experimental evidence that the latent trajectories are simple, interpretable, and capable of reproducing a Navier-Stokes pitchfork bifurcation diagram without decoding full fields. The upshot is a method that allows time extrapolation, zero-shot latent interpolation across parameters, and insight into the dynamics of the reduced system.

What carries the argument

The key mechanism is the LD-GCN architecture, consisting of (i) a Neural Ordinary Differential Equation (NODE) defined on a low-dimensional latent space, with the latent initial condition fixed to zero for every trajectory, that advances the latent state by explicit Euler time-stepping; and (ii) a graph convolutional decoder (a feedforward network followed by message-passing layers) that maps the latent state, together with time and signal, to the solution values on all mesh nodes. The theoretical backbone is Corollary 2.6, a universal approximation theorem for encoder-free architectures, whose proof reduces the graph decoder to a fully connected network by setting convolutional weights to z

What would settle it

Implement LD-GCN on a parameterized PDE with a genuinely parameter-dependent initial field and keep the latent start fixed at zero. If the model's error grows substantially, or if removing the decoder's explicit dependence on µ(t) destroys accuracy, then the encoder-free universal approximation claim does not actually apply to the varying-initial-condition setting the benchmarks use.

Watch

Extended reading notes

Core claim

The central discovery is that a latent dynamics network does not need an encoder to compress full-order states: if the initial condition is fixed and the dynamics is modeled by a neural ODE in a low-dimensional latent space, the latent trajectory itself is a global, geometry-independent representation of the system. A graph convolutional decoder then reconstructs the full solution on the mesh, and the paper proves (Corollary 2.6) that, under a 'perfect embedding' assumption, such an encoder-free model can approximate the parameter-to-solution map to arbitrary accuracy. Numerically, the model matches or beats the state-of-the-art graph autoencoder ROM GCA-ROM on advection-dominated problems w

Load-bearing premise

The load-bearing premise is that the initial condition of the PDE is fixed across all parameter instances, so the latent state can be initialized at zero for every trajectory; the numerical experiments in the paper discard early snapshots, so in practice the initial state varies with the parameter, and the theoretical guarantee proven for the encoder-free model does not directly apply to the model as run.

Editorial extensions

If this is right

  • Temporal extrapolation: because the latent dynamics is integrated in time, LD-GCN can predict solutions beyond the last training time step.
  • Zero-shot parameter predictions: interpolating latent trajectories via splines or Gaussian processes yields solutions for unseen parameter values without performing latent time integration.
  • Interpretable reduced dynamics: the latent coordinates are smooth and cluster by parameter value, and a scalar amplitude of a latent variable reproduces the pitchfork bifurcation diagram of a Coandă flow.
  • Parameter efficiency: the encoder-free design uses about 50% fewer trainable parameters than GCA-ROM while improving accuracy on the tested benchmarks.
  • Geometric flexibility: the graph decoder naturally supports unstructured meshes and geometrically parameterized domains, as shown in a moving-hole advection test case.

Reading between the lines

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

  • The fixed-initial-condition assumption in Corollary 2.6 is not met by the numerical benchmarks (which discard early snapshots), so the paper's universal approximation guarantee does not literally cover the model as run; the authors concede in Appendix A that the decoder then depends on the signal µ(t) to compensate. A direct extension would be to learn a parameter-dependent latent initial state.
  • The latent-space bifurcation detection could be turned into a practical screening tool: one could scan parameter space by interpolating latent trajectories and looking for amplitude changes or sign flips, without running any full-order solver.
  • The fact that spline and Gaussian-process interpolation nearly match the integrated LD-GCN suggests the regularizing effect of the latent dynamics may be doing much of the work; an ablation that removes the NODE and directly learns a map (t, µ) → s(t, µ) would isolate the benefit of causal time integration.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces LD-GCN, an encoder-free model-order-reduction architecture that couples a latent-space neural ODE with a graph-convolutional decoder. The method is tested on four parameterized time-dependent benchmarks (two advection–diffusion problems, lid-driven cavity flow, and the Coandă-effect Navier–Stokes problem), with reported improvements over GCA-ROM and a claimed performance comparable to LDNet. The manuscript also proves a universal approximation theorem for an encoder-free architecture, derives a Lipschitz-based interpolation error bound, and demonstrates that latent trajectories can reveal bifurcation structure. The paper is clearly written, includes a link to code, and provides concrete error numbers for each benchmark.

Significance. If the empirical results hold, the architecture is a genuinely useful contribution: it combines interpretable latent dynamics with geometric inductive biases, supports time extrapolation, and the latent bifurcation analysis is an insightful demonstration. The release of code and the explicit reporting of errors across four nontrivial problems are commendable and strengthen reproducibility. However, the theoretical certificate is substantially narrower than claimed, and the empirical comparisons rest on single runs, so the significance of the paper in its current form is reduced.

major comments (3)
  1. [§2.4.1, Corollary 2.6 and Eq. (8)] The UAT is proved for a model with a fixed initial condition u_h(t0) for all µ, a dynamics network N_Ndyn: I×R^{dµ}→R^n with no latent-state input, and a decoder N_Ndec: R^n→R^{Nh} with no t or µ dependence. The architecture actually trained in every benchmark has N_Ndyn(t,s(t),µ(t)) and, except for Coandă, N_Ndec(t,s(t),µ(t)). Appendix A states this dependence is introduced precisely because the retained initial condition varies with µ; moreover, in SA, MH, and cavity the first snapshot is removed, so the retained initial state at the first training time is parameter-dependent. Thus the fixed-initial-condition premise is violated in all experiments, and the proof does not extend to the stateful dynamics network. The abstract's claim that the methodology is 'mathematically validated via a UAT' therefore overstates what is proven: the theorem applies to a simpler, less expressive model th
  2. [Tables 1–2; Figs. 2a, 4] All reported errors are single-run point estimates. No seeds, error bars, or repeated-training statistics are provided. The headline comparative claims — LD-GCN outperforms GCA-ROM and is comparable to LDNet — are based on differences between single numbers (e.g., SA n=3 εmax 5.26e-2 vs 9.75e-2; cavity NRMSE 6.79e-3 vs 1.39e-3 for LDNet at a different latent dimension). Given the stochasticity of neural-network training, these differences may lie within run-to-run variability. Multi-seed means and standard deviations (or at least a sensitivity statement) are needed to support the comparative conclusions.
  3. [§3.2.2, Table 2] The claim that LD-GCN achieves 'performance comparable to LDNet' in the cavity benchmark is not strongly supported by the numbers: NRMSE=6.79e-3 for LD-GCN (n=3) versus 1.39e-3 for LDNet (n=10). While both are O(10^-3), a factor of roughly 5 is a meaningful gap in an MOR accuracy comparison. Since this is the only direct LDNet comparison in the paper, the text should either qualify the claim, compare at the same latent dimension, or provide additional context for why this gap is acceptable.
minor comments (6)
  1. [§2.5 and §3.1.4] In the definition of u_interp, 'N_Ndyn' should be 'N_Ndec' in both places. The current text writes u_interp(t,µ) = N_Ndyn(t, ˜s(t;µ),µ), but it is the decoder that maps the latent state to the physical field.
  2. [Remark 2.9] The phrase 'if the result of Proposition 9 holds' should refer to Proposition 2.8.
  3. [Algorithm 1, line 11] The integration uses N_int = floor(h/Δt). If h is not a multiple of Δt, the final partial step is discarded and the latent state is not advanced exactly to t_{k+1}. Please state that Δt divides h in all experiments, or describe how the remainder is handled.
  4. [Fig. 12b] The y-axis labels 'u2,h(T) 2' and 'u2,sim(T) 2' are unclear. Presumably this denotes the squared vertical-velocity component at final time; please clarify the notation.
  5. [§3.2, Eq. (16)] The regularization term L_ε is said to depend on a parameter ε, but its explicit form is not given in the text. A brief definition or a precise pointer to [79] would help.
  6. [§3.1.2] The first snapshot is removed from the SA trajectories 'for consistency with other test cases,' even though the true initial condition is identical for all µ. This removal is exactly what makes the retained initial state parameter-dependent and thus violates the fixed-initial-condition hypothesis of Corollary 2.6. The authors should comment on whether the fixed-initial-condition setting was also tested.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UAT is an explicit corollary of a stated external theorem, the interpolation bound is an explicit triangle inequality, and no fitted quantity is relabeled as a prediction.

full rationale

The claimed derivation chain is self-contained rather than circular. Corollary 2.6 is explicitly derived from the external theorem in [26] under the stated perfect-embedding and fixed-initial-condition assumptions; the only new part of the proof is the valid containment argument that setting all graph-convolutional weights to zero reduces the GCN decoder to a feedforward network (Section 2.4.1). This does not assume the conclusion. Proposition 2.8 is an explicit triangle-inequality decomposition, u_h - u_interp = (u_h - u_sim) + (u_sim - u_interp), with the Lipschitz constant and interpolation error as stated hypotheses, not as fitted outputs called predictions. The numerical benchmarks are external checks against finite-element data, and the reported improvements over GCA-ROM are empirical comparisons, not consequences of the theoretical statements. The paper's own caveats—Corollary 2.6 assumes a fixed initial condition and a dynamics network without latent-state input, while Appendix A concedes that the decoder is made to depend on (t, mu(t)) because 'the initial condition changes slightly for each instance of mu' and 'the only way to account for different initial conditions is to let the decoder depend on mu(t)'—identify a limitation in the scope of the theoretical guarantee relative to the deployed architecture. That is a validation-gap/overclaim issue, not a circular reduction: the theorem still has independent mathematical content for the fixed-initial-condition, constant-signal class it addresses. The self-citations used for the decoder template ([67]) and the Coandă benchmark setup ([85,69]) are published architectural/benchmark context, not load-bearing derivations that reduce the paper's claims to their own inputs. Therefore no prediction reduces by construction to its input, and no load-bearing argument reduces to an unverified self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or conserved quantities. Its theoretical claims rest on the standard perfect-embedding assumption, a fixed-initial-condition assumption that is at odds with all the numerical experiments, and the standard well-posedness/fixed-mesh modeling assumptions. The free parameters are the usual architectural and optimization hyperparameters, plus GPR kernel parameters for the auxiliary interpolation study.

free parameters (5)
  • latent dimension n = 3 (SA, cavity), 15 (MH), 2 (Coandă)
    Hyperparameter chosen per benchmark; results show robustness to n, but the central architecture depends on its value.
  • integration step Δt for latent Euler = 1e-2 (SA/MH), 5e-2 (cavity), 5e-1 (Coandă)
    Chosen manually for the explicit Euler integration of the latent ODE; affects time-stepping accuracy.
  • L1 regularization weight λ = 1e-5
    Set to 1e-5 in all experiments (Equation 7); standard weight-decay hyperparameter.
  • cavity loss hyperparameters δ, ε = δ=1e-1, ε=1e-4
    Used in the additional flow-direction regularization term (Equation 16) for the lid-driven cavity benchmark.
  • GPR kernel hyperparameters (length scales, noise variance) = ℓ initialized [0.1,0.5,0.5], σ²_n=1e-3, then optimized by marginal likelihood
    These are fitted to the learned latent trajectories in the auxiliary zero-shot interpolation study; not part of the core LD-GCN training.
assumptions (4)
  • domain assumption Perfect embedding assumption (Definition 2.5): there exist C^1 decoder Φ, Lipschitz latent dynamics f_n, and encoder Ψ such that the full-order solution equals Φ(s(t)) and the latent state obeys ṡ = f_n(t,s,µ) for all (t,µ).
    The UAT and the entire theoretical analysis depend on this unproved structural property of the data manifold, taken from [26,29]. If the data does not admit such a low-dimensional embedding with smooth decoder and Lipschitz latent field, the theorem gives no information.
  • domain assumption Fixed initial condition shared by all parameter instances (Corollary 2.6): u_h(t0) is the same for all µ, and latent initial state s(t0)=0.
    This premise is needed for the encoder-free construction. Appendix A explicitly acknowledges it is violated in the numerical experiments, where the decoder is made µ-dependent to absorb variable initial states.
  • domain assumption Well-posedness and causality of the signal-to-solution map S: P → U (Section 2.1).
    The learning problem is assumed well-posed: each signal µ maps to a unique solution, and the solution at time t depends only on the signal history up to t. This is standard for the target PDEs but is an assumption.
  • domain assumption Fixed mesh grid: number of nodes Nh and mesh connectivity are fixed across all parameter instances (Sections 2.1 and 2.4).
    The graph decoder operates on a single graph; the method cannot handle varying meshes. The authors acknowledge this as a limitation, but it is load-bearing for the geometric-parameter benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Dynamics Graph Convolutional Networks for model order reduction of parameterized time-dependent PDEs." pith.science (2026). https://pith.science/paper/O4X2TWJU

@misc{pith2026260111259,
  author       = {Pith},
  title        = {Pith review of: Latent Dynamics Graph Convolutional Networks for model order reduction of parameterized time-dependent PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4X2TWJU}},
  note         = {Machine review of arXiv:2601.11259}
}
read the original abstract

Graph Neural Networks (GNNs) are emerging as powerful tools for nonlinear Model Order Reduction (MOR) of time-dependent parameterized Partial Differential Equations (PDEs). However, existing methodologies struggle to combine geometric inductive biases with interpretable latent behavior, overlooking dynamics-driven features or disregarding spatial information. In this work, we address this gap by introducing Latent Dynamics Graph Convolutional Network (LD-GCN), a purely data-driven, encoder-free architecture that learns a global, low-dimensional representation of dynamical systems conditioned on external inputs and parameters. The temporal evolution is modeled in the latent space and advanced through time-stepping, allowing for time-extrapolation, and the trajectories are consistently decoded onto geometrically parameterized domains using a GNN. Our framework enhances interpretability by enabling the analysis of the reduced dynamics and supporting zero-shot prediction through latent interpolation. The methodology is mathematically validated via a universal approximation theorem for encoder-free architectures, and numerically tested on complex computational mechanics problems involving physical and geometric parameters, including the detection of bifurcating phenomena for Navier-Stokes equations. Code availability: https://github.com/lorenzotomada/ld-gcn-rom

Figures

Figures reproduced from arXiv: 2601.11259 by the authors.

Figure 1
Figure 1. Schematic representation of the LD-GCN architecture. the latent dynamics of the system according to Equation (3). The distinctive feature of LD-GCN lies [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Relative errors for all test trajectories and latent evolution w.r.t. time. To study the sensitivity of the LD-GCN architecture with respect to the latent dimension n, we repeated the simulations for increasing number of latent variables, the results of which are reported in [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Plot of the simulated solution (left), the full-order one (middle) and relative L 2 error field (right) for µ = (1, 1) at time t = 2. 3.1.3. Moving hole advection (MH). The second benchmark introduces the study of a geometrically parameterized domain. In fact, the main reason for introducing a graph-based model is the flexibility to consistently exploit geometric inductive bias, naturally embedding such important in… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Relative errors for all test trajectories w.r.t. time. 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 usim 0.0 0.2 0.5 0.8 1.0 1.2 1.5 1.8 2.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 uh 0.0 0.2 0.5 0.8 1.0 1.2 1.5 1.8 2.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.…
Figure 5
Figure 5. Figure 5: Plot of the simulated solution (left), the full-order one (middle) and relative L 2 error (right) for two different values of µ ∈ Ph. as LD-GCN. To this end, and in light of the discussion in Subsection 2.5, we study the latent state behavior of the SA system. A first …
Figure 6
Figure 6. Figure 6: Behavior in time of two different components of s(t; µ) as µ varies. Having assessed the interpretability of the resulting latent trajectories, following the argument from Subsection 2.5, we explore the possibility to perform zero-shot prediction, i.e., without inte￾gr…
Figure 7
Figure 7. Figure 7: Interpolated solution and errors for µtest = (0, 0). Moreover, we plot in [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Evolution of s3 for both latent trajectories with µtest = (0, 0). Finally, similar considerations can be made also in the more complex MH benchmark with geo￾metric parameters, where the trajectories exhibit regular patterns, as depicted in [PITH_FULL_IMAGE:figures/ful…
Figure 9
Figure 9. Figure 9: Evolution of s1 for two different fixed values of µ1 while varying µ2. of LD-GCN is lost and the feasibility of the method depends on the complexity of the problem and behavior of the latent space. We anticipate that, for instance, Subsection 3.3 provides an example in…
Figure 10
Figure 10. Figure 10: Realization of the signal v(t) and corresponding trajectory s(t; µ). 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 usim 2 0.0 0.2 0.5 0.8 1.0 1.2 1.5 1.8 2.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 uh 2 0.0 0.2 0.5 0.8 1.0 1.2 1.5 1.8 2.0 0.0 0.2 0.4 0.6 0.8…
Figure 11
Figure 11. Figure 11: Plot of the norm of the simulated solution (left), of the full-order one (middle) and of the relative L 2 error (right) for a given signal. ill-posed task, and to investigate whether the learned trajectories retain information about the underlying bifurcation, i.e., i…
Figure 12
Figure 12. Figure 12: Relative errors of LD-GCN for µ ∈ Ph and bifurcation diagram for FE and LD-GCN approximations. Finally, in [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Magnitude of the velocity field for pre- and post-bifurcating behaviors, top and bottom, respectively. 3.3.3. Bifurcation analysis. We conclude the analysis by investigating whether the interpretability and robustness of the LD-GCN architecture allow to detect the bif…
Figure 14
Figure 14. Figure 14: Phase portrait and amplitude of latent trajectories. In the left plot, the triangle corresponds to the initial condition s(t0; ·) = 0, while the dots denote the final values s(T; ·). varying geometries. In addition, we investigated the latent space identified by the n…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

92 extracted references · 14 canonical work pages

  1. [1]

    Flexible SE(2) graph neural networks with applications to PDE surrogates

    M. Bånkestad, O. Mogren, and A. Pirinen.Flexible SE(2) graph neural networks with applications to PDE surrogates. arXiv:2405.20287 [cs]. 2024.doi:10.48550/arXiv.2405.20287

    work page Pith review arXiv doi:10.48550/arxiv.2405.20287 2024

  2. [2]

    Regularized linear autoencoders recover the principal components, eventually

    X. Bao et al. “Regularized linear autoencoders recover the principal components, eventually”. In: Proceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20. Red Hook, NY, USA: Curran Associates Inc., 2020, pp. 6971–6981.url:https://dl.acm.org/ doi/10.5555/3495724.3496309

    arXiv 2020

  3. [3]

    Multiscale graph neural network autoencoders for interpretable scientific machine learning

    S. Barwey et al. “Multiscale graph neural network autoencoders for interpretable scientific machine learning”. In:Journal of Computational Physics495 (2023), p. 112537.doi:10.1016/j.jcp.2023. 112537

    doi:10.1016/j.jcp.2023 2023

  4. [4]

    Combining Differentiable PDE Solvers and Graph Neural Networks for Fluid Flow Prediction

    F. D. A. Belbute-Peres, T. Economon, and Z. Kolter. “Combining Differentiable PDE Solvers and Graph Neural Networks for Fluid Flow Prediction”. en. In:Proceedings of the 37th International Conference on Machine Learning. PMLR, 2020, pp. 2402–2411.url:https : / / proceedings . mlr . press/v119/de-avila-belbute-peres20a.html

    2020

  5. [5]

    Benner et al.Model Reduction and Approximation

    P. Benner et al.Model Reduction and Approximation. Computational Science & Engineering. Society for Industrial and Applied Mathematics, 2017.doi:10.1137/1.9781611974829

    doi:10.1137/1.9781611974829 2017

  6. [6]

    Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equa- tions

    R. Bergna et al. “Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equa- tions”. In:The Thirteenth International Conference on Learning Representations. 2025.url:https: //openreview.net/forum?id=TYSQYx9vwd

    2025

  7. [7]

    Error estimates for deep learning methods in fluid dynamics

    A. Biswas, J. Tian, and S. Ulusoy. “Error estimates for deep learning methods in fluid dynamics”. en. In:Numerische Mathematik151.3 (2022), pp. 753–777.doi:10.1007/s00211-022-01294-z

    work page doi:10.1007/s00211-022-01294-z 2022

  8. [8]

    Bonneville et al.A Comprehensive Review of Latent Space Dynamics Identification Algorithms for Intrusive and Non-Intrusive Reduced-Order-Modeling

    C. Bonneville et al.A Comprehensive Review of Latent Space Dynamics Identification Algorithms for Intrusive and Non-Intrusive Reduced-Order-Modeling. arXiv:2403.10748 [cs]. 2024.doi:10 . 48550 / arXiv.2403.10748

    arXiv doi:10.48550/arxiv.2403.10748 2024

Show all 92 references
  1. [9]

    Bonneville et al.Extrapolating Phase-Field Simulations in Space and Time with Purely Convolu- tional Architectures

    C. Bonneville et al.Extrapolating Phase-Field Simulations in Space and Time with Purely Convolu- tional Architectures. arXiv:2509.20770 [cs] version: 1. 2025.doi:10.48550/arXiv.2509.20770

    2025 doi

  2. [10]

    Bouziani and N

    N. Bouziani and N. Boullé.Structure-Preserving Operator Learning. arXiv:2410.01065 [cs]. 2024.doi: 10.48550/arXiv.2410.01065

    2024 doi arXiv

  3. [11]

    Message Passing Neural PDE Solvers

    J. Brandstetter, D. E. Worrall, and M. Welling. “Message Passing Neural PDE Solvers”. In:Inter- national Conference on Learning Representations. 2022.url:https://openreview.net/forum?id= vSix3HPYKSU

    2022

  4. [12]

    Geometric and Physical Quantities improve E(3) Equivariant Message Passing

    J. Brandstetter et al. “Geometric and Physical Quantities improve E(3) Equivariant Message Passing”. In:International Conference on Learning Representations. 2022.url:https : / / openreview . net / forum?id=_xwr8gOBeV1

    2022

  5. [13]

    Error estimates for POD-DL-ROMs: a deep learning framework for reduced order mod- eling of nonlinear parametrized PDEs enhanced by proper orthogonal decomposition

    S. Brivio et al. “Error estimates for POD-DL-ROMs: a deep learning framework for reduced order mod- eling of nonlinear parametrized PDEs enhanced by proper orthogonal decomposition”. In:Advances in Computational Mathematics50.3 (2024), p. 33.doi:10.1007/s10444-024-10110-1

    2024 doi

  6. [14]

    M. M. Bronstein et al.Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. arXiv:2104.13478 [cs]. 2021.doi:10.48550/arXiv.2104.13478

    2021 doi arXiv

  7. [15]

    S. L. Brunton and J. N. Kutz.Data-Driven Science and Engineering: Machine Learning, Dynam- ical Systems, and Control. Cambridge University Press, 2019.doi:https : / / doi . org / 10 . 1017 / 9781108380690

    2019

  8. [16]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    S. L. Brunton, J. L. Proctor, and J. N. Kutz. “Discovering governing equations from data by sparse identification of nonlinear dynamical systems”. In:Proceedings of the National Academy of Sciences 113.15 (2016). Publisher: Proceedings of the National Academy of Sciences, pp. ...

    2016

  9. [17]

    Proper Orthogonal Decomposition Extensions for Parametric Applications in Compressible Aerodynamics

    T. Bui-Thanh, M. Damodaran, and K. Willcox. “Proper Orthogonal Decomposition Extensions for Parametric Applications in Compressible Aerodynamics”. en. In:21st AIAA Applied Aerodynamics Conference. Orlando, Florida: American Institute of Aeronautics and Astronautics, 2003.doi:1...

    2003

  10. [18]

    Data-driven discovery of coordinates and governing equations

    K. Champion et al. “Data-driven discovery of coordinates and governing equations”. In:Proceedings of the National Academy of Sciences116.45 (2019). Publisher: Proceedings of the National Academy of Sciences, pp. 22445–22451.doi:10.1073/pnas.1906995116

    2019 doi

  11. [19]

    Neural Ordinary Differential Equations

    R. T. Q. Chen et al. “Neural Ordinary Differential Equations”. In:Advances in Neural Information Processing Systems. Vol. 31. Curran Associates, Inc., 2018.url:https://proceedings.neurips.cc/ paper_files/paper/2018/hash/69386f6bb1dfed68692a24c8686939b9-Abstract.html. REFERENCES 27

    2018

  12. [20]

    Chen et al.Time Extrapolation with Graph Convolutional Autoencoder and Tensor Train Decom- position

    Y. Chen et al.Time Extrapolation with Graph Convolutional Autoencoder and Tensor Train Decom- position. arXiv:2511.23037 [math]. 2025.doi:10.48550/arXiv.2511.23037

    2025 doi

  13. [21]

    Reduced order modeling of parametrized systems through autoencoders and SINDy approach: continuation of periodic solutions

    P. Conti et al. “Reduced order modeling of parametrized systems through autoencoders and SINDy approach: continuation of periodic solutions”. In:Computer Methods in Applied Mechanics and Engi- neering411 (2023). Publisher: Elsevier BV, p. 116072.doi:10.1016/j.cma.2023.116072

    2023

  14. [22]

    A non-intrusive approach for the reconstruction of POD modal coefficients through active subspaces

    N. Demo, M. Tezzele, and G. Rozza. “A non-intrusive approach for the reconstruction of POD modal coefficients through active subspaces”. In:Comptes Rendus Mécanique. Data-Based Engineering Sci- ence and Technology 347.11 (2019), pp. 873–881.doi:10.1016/j.crme.2019.11.012

    2019 doi

  15. [23]

    H. A. Dijkstra and F. W. Wubs.Bifurcation Analysis of Fluid Flows. Cambridge University Press, 2023.doi:DOI:https://doi.org/10.1017/9781108863148

    2023 doi

  16. [24]

    Non-intrusive data-driven reduced-order modeling for time-dependent parametrized problems

    J. Duan and J. S. Hesthaven. “Non-intrusive data-driven reduced-order modeling for time-dependent parametrized problems”. In:Journal of Computational Physics497 (2024), p. 112621.doi:10.1016/ j.jcp.2023.112621

    2024

  17. [25]

    Shallow recurrent decoder for reduced order modeling of E× B plasma dynamics

    F. Faraji, M. Reza, and J. N. Kutz. “Shallow recurrent decoder for reduced order modeling of E× B plasma dynamics”. en. In:Machine Learning: Science and Technology6.2 (2025), p. 025024.doi: 10.1088/2632-2153/adcd20

    2025 doi

  18. [26]

    On latent dynamics learning in nonlinear reduced order modeling

    N. Farenga et al. “On latent dynamics learning in nonlinear reduced order modeling”. In:Neural Networks185 (2025), p. 107146.doi:10.1016/j.neunet.2025.107146

    2025

  19. [27]

    A comprehensive and biophysically detailed computational model of the whole human heart electromechanics

    M. Fedele et al. “A comprehensive and biophysically detailed computational model of the whole human heart electromechanics”. In:Computer Methods in Applied Mechanics and Engineering410 (2023), p. 115983.doi:10.1016/j.cma.2023.115983

    2023

  20. [28]

    Fey et al.PyG 2.0: Scalable Learning on Real World Graphs

    M. Fey et al.PyG 2.0: Scalable Learning on Real World Graphs. arXiv:2507.16991 [cs]. 2025.doi: 10.48550/arXiv.2507.16991

    2025 doi arXiv

  21. [29]

    A deep learning approach to Reduced Order Modelling of parameter dependent partial differential equations

    N. Franco, A. Manzoni, and P. Zunino. “A deep learning approach to Reduced Order Modelling of parameter dependent partial differential equations”. In:Mathematics of Computation92.340 (2022), pp. 483–524.doi:10.1090/mcom/3781

    2022 doi

  22. [30]

    A practical existence theorem for reduced order models based on convolutional autoencoders

    N. R. Franco and S. Brugiapaglia. “A practical existence theorem for reduced order models based on convolutional autoencoders”. en. In:Foundations of Data Science7.1 (2025), pp. 72–98.doi:10.3934/ fods.2024030

    2025

  23. [31]

    Approximation bounds for convolutional neural networks in operator learning

    N. R. Franco et al. “Approximation bounds for convolutional neural networks in operator learning”. In:Neural Networks161 (2023), pp. 129–141.doi:10.1016/j.neunet.2023.01.029

    2023 doi

  24. [32]

    Deep learning-based surrogate models for parametrized PDEs: Handling geometric variability through graph neural networks

    N. R. Franco et al. “Deep learning-based surrogate models for parametrized PDEs: Handling geometric variability through graph neural networks”. In:Chaos: An Interdisciplinary Journal of Nonlinear Science33.12 (2023), p. 123121.doi:10.1063/5.0170101

    2023 doi

  25. [33]

    A Comprehensive Deep Learning-Based Approach to Reduced Order Modeling of Nonlinear Time-Dependent Parametrized PDEs

    S. Fresca, L. Dede’, and A. Manzoni. “A Comprehensive Deep Learning-Based Approach to Reduced Order Modeling of Nonlinear Time-Dependent Parametrized PDEs”. en. In:Journal of Scientific Com- puting87.2 (2021), p. 61.doi:10.1007/s10915-021-01462-7

    2021 doi

  26. [34]

    LaSDI: Parametric Latent Space Dynamics Identification

    W. D. Fries, X. He, and Y. Choi. “LaSDI: Parametric Latent Space Dynamics Identification”. In: Computer Methods in Applied Mechanics and Engineering399 (2022), p. 115436.doi:10.1016/j. cma.2022.115436

    2022

  27. [35]

    Physics-informed graph neural Galerkin networks: A unified framework for solving PDE-governed forward and inverse problems

    H. Gao, M. J. Zahr, and J.-X. Wang. “Physics-informed graph neural Galerkin networks: A unified framework for solving PDE-governed forward and inverse problems”. In:Computer Methods in Applied Mechanics and Engineering390 (2022), p. 114502.doi:10.1016/j.cma.2021.114502

    2022

  28. [36]

    M. L. Gao, J. P. Williams, and J. N. Kutz.Sparse identification of nonlinear dynamics and Koopman operators with Shallow Recurrent Decoder Networks. arXiv:2501.13329 [cs]. 2025.doi:10 . 48550 / arXiv.2501.13329. [37]GaussianProcessRegressor — scikit-learn 1.7.0 documentation.u...

    2025 doi arXiv

  29. [38]

    Approximation rates for neural networks with encodable weights in smoothness spaces

    I. Gühring and M. Raslan. “Approximation rates for neural networks with encodable weights in smoothness spaces”. In:Neural Networks134 (2021), pp. 107–130.doi:10.1016/j.neunet.2020.11. 010

    2021 doi

  30. [39]

    W. L. Hamilton.Graph Representation Learning. en. Synthesis Lectures on Artificial Intelligence and MachineLearning.Cham:SpringerInternationalPublishing,2020.doi:10.1007/978-3-031-01588-5

    2020 doi

  31. [40]

    Predicting Physics in Mesh-reduced Space with Temporal Attention

    X. Han et al. “Predicting Physics in Mesh-reduced Space with Temporal Attention”. In:Interna- tional Conference on Learning Representations. 2022.url:https : / / openreview . net / forum ? id = XctLdNfCmP. 28 REFERENCES

    2022

  32. [41]

    Thermodynamics-informed graph neural networks

    Q. Hernández et al. “Thermodynamics-informed graph neural networks”. In:IEEE Transactions on Artificial Intelligence5.3 (2024). arXiv:2203.01874 [cs], pp. 967–976.doi:10.1109/TAI.2022.3179681

    2024 arXiv

  33. [42]

    Non-intrusive reduced order modeling of nonlinear problems using neural networks

    J. S. Hesthaven and S. Ubbiali. “Non-intrusive reduced order modeling of nonlinear problems using neural networks”. In:Journal of Computational Physics363 (2018), pp. 55–78.doi:10.1016/j.jcp. 2018.02.037

    2018 doi

  34. [43]

    J. S. Hesthaven, G. Rozza, and B. Stamm.Certified Reduced Basis Methods for Parametrized Partial Differential Equations. en. SpringerBriefs in Mathematics. Cham: Springer International Publishing, 2016.doi:10.1007/978-3-319-22470-1

    2016 doi

  35. [44]

    Certified machine learning: A posteriori error estimation for physics- informed neural networks

    B. Hillebrecht and B. Unger. “Certified machine learning: A posteriori error estimation for physics- informed neural networks”. In:2022 International Joint Conference on Neural Networks (IJCNN). arXiv:2203.17055 [cs]. 2022, pp. 1–8.doi:10.1109/IJCNN55064.2022.9892569

    2022 arXiv

  36. [45]

    Graph neural PDE solvers with conservation and similarity-equivariance

    M. Horie and N. Mitsume. “Graph neural PDE solvers with conservation and similarity-equivariance”. In:Proceedings of the 41st International Conference on Machine Learning. Vol. 235. ICML’24. Vienna, Austria: JMLR.org, 2024, pp. 18785–18814.url:https : / / dl . acm . org / doi ...

    2024

  37. [46]

    T. Y. Hsieh and Y. J. Zhang.GALDS: A Graph-Autoencoder-based Latent Dynamics Surrogate model to predict neurite material transport. arXiv:2507.10871 [cs]. 2025.doi:10.48550/arXiv.2507.10871

    2025 doi arXiv

  38. [47]

    Learning continuous-time PDEs from sparse data with graph neural networks

    V. Iakovlev, M. Heinonen, and H. Lähdesmäki. “Learning continuous-time PDEs from sparse data with graph neural networks”. In:International Conference on Learning Representations. 2021.url: https://openreview.net/forum?id=aUX5Plaq7Oy

    2021

  39. [48]

    Optimal transport-based displacement interpolation with data augmentation for reduced order modeling of nonlinear dynamical systems

    M. Khamlich et al. “Optimal transport-based displacement interpolation with data augmentation for reduced order modeling of nonlinear dynamical systems”. In:Journal of Computational Physics531 (2025), p. 113938.doi:10.1016/j.jcp.2025.113938

    2025

  40. [49]

    Neural operator: learning maps between function spaces with applications to PDEs

    N. Kovachki et al. “Neural operator: learning maps between function spaces with applications to PDEs”. In:J. Mach. Learn. Res.24.1 (2023).url:https://dl.acm.org/doi/10.5555/3648699. 3648788

    2023 doi

  41. [50]

    Li et al.Geometric Operator Learning with Optimal Transport

    X. Li et al.Geometric Operator Learning with Optimal Transport. arXiv:2507.20065 [cs]. 2025.doi: 10.48550/arXiv.2507.20065

    2025 doi arXiv

  42. [51]

    Geometry-informed neural operator for large-scale 3D PDEs

    Z. Li et al. “Geometry-informed neural operator for large-scale 3D PDEs”. In:Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23. Red Hook, NY, USA: Curran Associates Inc., 2023, pp. 35836–35854.url:https : / / dl . acm . org / d...

    2023

  43. [52]

    2020.doi:10.48550/arXiv.2003.03485

    Z.Lietal.Neural Operator: Graph Kernel Network for Partial Differential Equations.arXiv:2003.03485 [cs]. 2020.doi:10.48550/arXiv.2003.03485

    2003 doi arXiv

  44. [53]

    Enabling Automatic Differentiation with Mollified Graph Neural Operators

    R. Y. Lin et al. “Enabling Automatic Differentiation with Mollified Graph Neural Operators”. In: Transactions on Machine Learning Research(2025).url:https : / / openreview . net / forum ? id = CGoR1hFAGr. [54]LinearNDInterpolator — SciPy v1.16.0 Manual.url:https : / / docs . s...

    2025

  45. [55]

    On the limited memory BFGS method for large scale optimization

    D. C. Liu and J. Nocedal. “On the limited memory BFGS method for large scale optimization”. en. In:Mathematical Programming45.1-3 (1989), pp. 503–528.doi:10.1007/BF01589116

    1989 doi

  46. [56]

    SEGNO: Generalizing Equivariant Graph Neural Networks with Physical Inductive Biases

    Y. Liu et al. “SEGNO: Generalizing Equivariant Graph Neural Networks with Physical Inductive Biases”. In:The Twelfth International Conference on Learning Representations. 2024.url:https: //openreview.net/forum?id=3oTPsORaDH

    2024

  47. [57]

    Hierarchical deep learning of multiscale differential equation time-steppers

    Y. Liu, J. N. Kutz, and S. L. Brunton. “Hierarchical deep learning of multiscale differential equation time-steppers”. en. In:Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences380.2229 (2022), p. 20210200.doi:10.1098/rsta.2021.0200

    2022

  48. [58]

    Graph ODEs and Beyond: A Comprehensive Survey on Integrating Differential Equations with Graph Neural Networks

    Z. Liu et al. “Graph ODEs and Beyond: A Comprehensive Survey on Integrating Differential Equations with Graph Neural Networks”. In:Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. KDD ’25. New York, NY, USA: Association for Computing Ma...

    2025

  49. [59]

    Logg, K.-A

    A. Logg, K.-A. Mardal, G. N. Wells, et al.Automated Solution of Differential Equations by the Finite Element Method. Springer, 2012.doi:10.1007/978-3-642-23099-8

    2012 doi

  50. [60]

    Latent space modeling of parametric and time-dependent PDEs using neural ODEs

    A. Longhi, D. Lathouwers, and Z. Perkó. “Latent space modeling of parametric and time-dependent PDEs using neural ODEs”. In:Computer Methods in Applied Mechanics and Engineering448 (2026), p. 118394.doi:10.1016/j.cma.2025.118394. REFERENCES 29

    2026

  51. [61]

    Learning Latent Graph Dynamics for Visual Manipulation of De- formable Objects

    X. Ma, D. Hsu, and W. S. Lee. “Learning Latent Graph Dynamics for Visual Manipulation of De- formable Objects”. In:2022 International Conference on Robotics and Automation (ICRA). 2022, pp. 8266–8273.doi:10.1109/ICRA46639.2022.9811597

    2022

  52. [62]

    Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs

    F. Monti et al. “Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs”. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017, pp. 5425–5434.doi:10.1109/CVPR.2017.576

    2017 doi

  53. [63]

    GFN: A graph feedforward network for resolution- invariant reduced operator learning in multifidelity applications

    O. M. Morrison, F. Pichi, and J. S. Hesthaven. “GFN: A graph feedforward network for resolution- invariant reduced operator learning in multifidelity applications”. In:Computer Methods in Applied Mechanics and Engineering432 (2024), p. 117458.doi:10.1016/j.cma.2024.117458

    2024

  54. [64]

    Ohlberger and S

    M. Ohlberger and S. Rave.Reduced Basis Methods: Success, Limitations and Future Challenges. 2016. doi:10.48550/arXiv.1511.02021. arXiv:1511.02021 [math]

    2016 doi arXiv

  55. [65]

    Learning two-phase microstructure evolution using neural operators and autoen- coder architectures

    V. Oommen et al. “Learning two-phase microstructure evolution using neural operators and autoen- coder architectures”. en. In:npj Computational Materials8.1 (2022), p. 190.doi:10.1038/s41524- 022-00876-7

    2022 doi

  56. [66]

    Learning Mesh-Based Simulation with Graph Networks

    T. Pfaff et al. “Learning Mesh-Based Simulation with Graph Networks”. In:International Conference on Learning Representations. 2021.url:https://openreview.net/forum?id=roNqYL0_XP

    2021

  57. [67]

    A graph convolutional autoencoder approach to model order reduction for parametrized PDEs

    F. Pichi, B. Moya, and J. S. Hesthaven. “A graph convolutional autoencoder approach to model order reduction for parametrized PDEs”. In:Journal of Computational Physics501 (2024), p. 112762.doi: 10.1016/j.jcp.2024.112762

    2024

  58. [68]

    Deflation-Based Certified Greedy Algorithm and Adaptivity for Bifur- cating Nonlinear PDEs

    F. Pichi and M. Strazzullo. “Deflation-Based Certified Greedy Algorithm and Adaptivity for Bifur- cating Nonlinear PDEs”. In:Communications in Nonlinear Science and Numerical Simulation149 (2025), p. 108941.doi:10.1016/j.cnsns.2025.108941

    2025

  59. [69]

    An artificial neural network approach to bifurcating phenomena in computational fluid dynamics

    F. Pichi et al. “An artificial neural network approach to bifurcating phenomena in computational fluid dynamics”. In:Computers & Fluids254 (2023), p. 105813.doi:10.1016/j.compfluid.2023.105813

    2023

  60. [70]

    DrivingbifurcatingparametrizednonlinearPDEsbyoptimalcontrolstrategies:applica- tion to Navier–Stokes equations with model order reduction

    F.Pichietal.“DrivingbifurcatingparametrizednonlinearPDEsbyoptimalcontrolstrategies:applica- tion to Navier–Stokes equations with model order reduction”. en. In:ESAIM: Mathematical Modelling and Numerical Analysis56.4 (2022), pp. 1361–1400.doi:10.1051/m2an/2022044

    2022

  61. [71]

    Poli et al.Graph Neural Ordinary Differential Equations

    M. Poli et al.Graph Neural Ordinary Differential Equations. arXiv:1911.07532 [cs]. 2021.doi:10. 48550/arXiv.1911.07532

    1911 doi arXiv

  62. [72]

    S. J. Prince.Understanding Deep Learning. MIT Press, 2023.url:http://udlbook.com

    2023

  63. [73]

    Symmetry breaking and preliminary results about a Hopf bifurcation for incompressible viscous flow in an expansion channel

    A. Quaini, R. Glowinski, and S. Čanić. “Symmetry breaking and preliminary results about a Hopf bifurcation for incompressible viscous flow in an expansion channel”. In:International Journal of Computational Fluid Dynamics(2016).doi:https://doi.org/10.1080/10618562.2016.1144877

    2016

  64. [74]

    Quarteroni.Numerical Models for Differential Problems

    A. Quarteroni.Numerical Models for Differential Problems. en. Vol. 16. MS&A. Cham: Springer In- ternational Publishing, 2017.doi:10.1007/978-3-319-49316-9

    2017 doi

  65. [75]

    Combining physics-based and data-driven models: advancing the frontiers of research with scientific machine learning

    A. Quarteroni, P. Gervasio, and F. Regazzoni. “Combining physics-based and data-driven models: advancing the frontiers of research with scientific machine learning”. In:Mathematical Models and Methods in Applied Sciences35.04 (2025), pp. 905–1071.doi:10.1142/S0218202525500125

    2025 doi

  66. [76]

    Quarteroni, A

    A. Quarteroni, A. Manzoni, and F. Negri.Reduced Basis Methods for Partial Differential Equations. Vol. 92. UNITEXT. Cham: Springer International Publishing, 2016.doi:10.1007/978-3-319-15431- 2

    2016 doi

  67. [77]

    New York, NY: Springer, 2007.doi:10.1007/b98885

    A.Quarteroni,R.Sacco,andF.Saleri.Numerical Mathematics.Vol.37.TextsinAppliedMathematics. New York, NY: Springer, 2007.doi:10.1007/b98885

    2007 doi

  68. [78]

    C. E. Rasmussen and C. K. I. Williams.Gaussian Processes for Machine Learning. en. The MIT Press, 2005.doi:10.7551/mitpress/3206.001.0001

    2005 doi

  69. [79]

    Learning the intrinsic dynamics of spatio-temporal processes through Latent Dynamics Networks

    F. Regazzoni et al. “Learning the intrinsic dynamics of spatio-temporal processes through Latent Dynamics Networks”. en. In:Nature Communications15.1 (2024), p. 1834.doi:10.1038/s41467- 024-45323-x

    2024 doi

  70. [80]

    Rozza et al.Real Time Reduced Order Computational Mechanics: Parametric PDEs Worked Out Problems

    G. Rozza et al.Real Time Reduced Order Computational Mechanics: Parametric PDEs Worked Out Problems. SISSA Springer Series. Springer Cham, 2024.doi:https://doi.org/10.1007/978-3-031- 49892-3

    2024 doi

  71. [81]

    Seydel.Practical Bifurcation and Stability Analysis

    R. Seydel.Practical Bifurcation and Stability Analysis. en. Vol. 5. Interdisciplinary Applied Mathe- matics. New York, NY: Springer, 2010.doi:10.1007/978-1-4419-1740-9

    2010 doi

  72. [82]

    GraphVAE: Towards Generation of Small Graphs Using Varia- tional Autoencoders

    M. Simonovsky and N. Komodakis. “GraphVAE: Towards Generation of Small Graphs Using Varia- tional Autoencoders”. en. In:Artificial Neural Networks and Machine Learning – ICANN 2018. Ed. by V. Kůrková et al. Cham: Springer International Publishing, 2018, pp. 412–422.doi:10.1007...

    2018 doi

  73. [83]

    Mesh neural networks for SE(3)-equivariant hemodynamics estimation on the artery wall

    J. Suk et al. “Mesh neural networks for SE(3)-equivariant hemodynamics estimation on the artery wall”. In:Computers in Biology and Medicine173 (2024), p. 108328.doi:10.1016/j.compbiomed. 2024.108328

    2024

  74. [84]

    A survey on universal approximation and its limits in soft computing techniques

    D. Tikk, L. T. Kóczy, and T. D. Gedeon. “A survey on universal approximation and its limits in soft computing techniques”. In:International Journal of Approximate Reasoning33.2 (2003), pp. 185–202. doi:10.1016/S0888-613X(03)00021-5

    2003 doi

  75. [85]

    Sparse identification for bifurcating phenomena in computational fluid dynamics

    L. Tomada et al. “Sparse identification for bifurcating phenomena in computational fluid dynamics”. In:Computers & Fluids(2025), p. 106841.doi:10.1016/j.compfluid.2025.106841

    2025

  76. [86]

    Reduced order modeling with shallow recurrent decoder networks

    M. Tomasetto et al. “Reduced order modeling with shallow recurrent decoder networks”. en. In:Nature Communications16.1 (2025), p. 10260.doi:10.1038/s41467-025-65126-y

    2025 doi

  77. [87]

    Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks

    M. L. Valencia, T. Pfaff, and N. Thuerey. “Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks”. In:The Thirteenth International Conference on Learning Representations. 2025.url:https://openreview.net/forum?id=uKZdlihDDn

    2025

  78. [88]

    Latent Neural Operator Pretraining for Solving Time-Dependent PDEs

    T. Wang and C. Wang. “Latent Neural Operator Pretraining for Solving Time-Dependent PDEs”. en. In:Neural Information Processing. Ed. by M. Mahmud et al. Singapore: Springer Nature, 2025, pp. 163–178.doi:10.1007/978-981-96-6948-6_12

    2025 doi

  79. [89]

    Model identification of reduced order fluid dynamics systems using deep learning

    Z. Wang et al. “Model identification of reduced order fluid dynamics systems using deep learning”. en. In:International Journal for Numerical Methods in Fluids86.4 (2018), pp. 255–268.doi:10.1002/ fld.4416

    2018

  80. [90]

    Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes

    T. Würth et al. “Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes”. In:Computer Methods in Applied Mechanics and Engineering429 (2024), p. 117102.doi:10.1016/j.cma.2024.117102

    2024

  81. [91]

    Equivariant graph neural operator for modeling 3D dynamics

    M. Xu et al. “Equivariant graph neural operator for modeling 3D dynamics”. In:Proceedings of the 41st International Conference on Machine Learning. Vol. 235. ICML’24. Vienna, Austria: JMLR.org, 2024, pp. 55015–55032.url:https://dl.acm.org/doi/10.5555/3692070.3694335

    2024

  82. [92]

    Yavich et al.Differentiable Implicit Solver on Graph Neural Networks for Forward and Inverse Problems

    N. Yavich et al.Differentiable Implicit Solver on Graph Neural Networks for Forward and Inverse Problems. 2025.url:https://openreview.net/forum?id=zuuhtmK1Ub

    2025

  83. [93]

    Combining physics-informed graph neural network and finite difference for solv- ing forward and inverse spatiotemporal PDEs

    H. Zhang et al. “Combining physics-informed graph neural network and finite difference for solv- ing forward and inverse spatiotemporal PDEs”. In:Computer Physics Communications308 (2025), p. 109462.doi:10.1016/j.cpc.2024.109462

    2025

  84. [94]

    LESnets (large-eddy simulation nets): Physics-informed neural operator for large- eddy simulation of turbulence

    S. Zhao et al. “LESnets (large-eddy simulation nets): Physics-informed neural operator for large- eddy simulation of turbulence”. In:Journal of Computational Physics537 (2025), p. 114125.doi: 10.1016/j.jcp.2025.114125. AppendixA.Architectures Here, we discuss some details of t...

    2025

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.