Pith. sign in

REVIEW 3 major objections 5 minor 47 references

On the under-reaching phenomenon in message-passing neural PDE solvers: revisiting the CFL condition

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Neural PDE solvers fail when their message-passing count is below a physics-determined lower bound; this paper derives that bound for hyperbolic, parabolic, and elliptic equations and validates its sharpness.

desk verdict A useful, parameter-free rule of thumb for setting message-passing iterations in GNN PDE solvers, with a CFL-style bound that is genuinely new, but the parabolic/elliptic sharpness claim overreaches and the hyperbolic formula has an unresolved time-step ambiguity. read the letter →

arxiv 2507.08861 v1 pith:3ZMGMAS7 submitted 2025-07-09 cs.LG stat.ML

Lucas Tesan , Mikel M. Iparraguirre , David Gonzalez , Pedro Martins , Elias Cueto This is my paper · ORCID
classification cs.LGstat.ML
keywords under-reachingmessage-passingiterationsgraphneuralnetworksPDEsolversCFLconditionhyperbolicparabolicelliptic
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

Under-reaching—a node being unable to see nodes farther away than the number of message-passing hops—is treated here as a physics problem rather than an architecture problem. The paper claims that the minimum number of message-passing iterations in a GNN PDE solver is fixed by the equation class, the discretization, and the domain geometry: hyperbolic equations require $M \geq \lceil \sqrt{2} c \Delta t / \Delta x \rceil$ so that messages stay ahead of the wave, while parabolic and elliptic equations require $M = L / \Delta x$ so that every node receives a global view of the domain in each inference step. If these bounds hold, a hyperparameter that is normally tuned by trial and error can be computed in advance, and models that meet the bound are stable, accurate, and able to extrapolate to domains within reach. Four experiments—wave, heat, Poisson, and incremental sheet forming—show errors dropping sharply at the predicted thresholds.

What carries the argument

The load-bearing object is a distance-per-hop identity: each message-passing iteration propagates information across one cell of the spatial discretization, so the hop count $M$ measures the reach of the network in units of $\Delta x$. Combined with the CFL stability condition for hyperbolic schemes, this identity yields the wave-speed bound; combined with the infinite-propagation-speed property of diffusive and static problems, it yields the geometry bound $M = L / \Delta x$. The experimental design isolates the mechanism by keeping the number of trainable parameters fixed while varying only $M$, so changes in accuracy are attributable to information propagation rather than model capacity.

What would settle it

Run a heat or Poisson problem on a square grid with a source at one corner and the quantity of interest measured at the opposite corner, using a GNN with $M = L / \Delta x$ message-passing iterations; if the far-corner error remains large and only drops when $M$ approaches $2L / \Delta x$, then the geometric bound is not sharp. An even simpler check is to compute the graph diameter of the four-connected mesh and test whether the bound equals diameter hops rather than side length.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that under-reaching in message-passing neural PDE solvers is not a neural-network pathology but a failure to satisfy a discrete information-propagation constraint inherited from the governing equation. One hop is identified with crossing one spatial cell $\Delta x$, and the required number of hops is derived from first principles: for hyperbolic PDEs the bound $M > \sqrt{2} c \Delta t / \Delta x$, rounded up to an integer, reproduces the CFL condition in the language of message passing; for parabolic and elliptic PDEs, whose perturbations formally propagate at infinite speed, the bound $M = L / \Delta x$ ensures the message reaches the far boundary of the domain. The numerical results show a sharp threshold: below the bound the model cannot integrate rollouts and can diverge; at or above it, error saturates, with extra passes yielding little and eventually worsening predictions through over-smoothing. The paper further shows that success in extrapolating to unseen domains is governed by the same condition: models extrapolate to larger domains exactly when the MPI count still satisfies the bound for the new domain.

Load-bearing premise

The load-bearing premise for the parabolic and elliptic bounds is that messages only need to travel from one side of the square domain to the opposite side ($M = L / \Delta x$) to give every node the global information it needs; if corner-to-corner interactions matter, the true minimum would be about twice as large, and the claimed bound would under-predict.

Editorial extensions

If this is right

  • The number of message-passing iterations stops being a free hyperparameter: for a given PDE class, $\Delta x$, $\Delta t$, and domain size, the minimum $M$ is known before training.
  • Below the bound, no amount of depth or latent capacity compensates for under-reaching, because the missing information never reaches the nodes.
  • At or above the bound, rollout error saturates; increasing $M$ beyond the bound is wasted computation and can degrade accuracy through over-smoothing.
  • Hyperbolic GNN solvers generalize to unseen geometries and longer rollouts because their information need is local, while parabolic and elliptic solvers generalize only when the new domain lies within the message-passing reach.
  • The hyperbolic bound connects GNN solver design to classical numerical analysis, giving learned simulators a stability condition analogous to CFL.

Reading between the lines

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

  • Beyond the paper: on irregular or non-uniform meshes, the same reasoning suggests the hyperbolic bound should be evaluated per edge, using the local spacing and possibly directional wave speeds, rather than a single global $\Delta x$.
  • Beyond the paper: for parabolic and elliptic problems with strong corner-to-corner coupling, the geometric bound may need to be the graph diameter in hops, about $2L / \Delta x$ on a square grid, rather than the side-to-side $L / \Delta x$; this is a testable sharpening of the claimed bound.
  • Beyond the paper: because the bound is derived from physics alone, other sources of error—over-squashing on irregular graphs, over-smoothing at high $M$, and numerical diffusion in the learned scheme—can shift or widen the optimal operating range in practice.
  • Beyond the paper: the hop-count logic could be inverted to choose the temporal sampling: selecting $\Delta t$ so that the required $M$ matches a computational budget trades rollout accuracy against inference cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a priori lower bounds on the number of message-passing iterations (MPI) for graph-neural-network PDE solvers, for the three main PDE classes. For hyperbolic equations it derives M > ceil(sqrt(2) c dt/dx), which it connects to the CFL condition; for parabolic and elliptic equations it proposes M = L/dx, requiring messages to cross the domain. The claims are tested on the wave equation, the heat equation, Poisson's equation, and an incremental metal-forming problem, using shared-weight GNN architectures and multiple random seeds. The paper reports that accuracy saturates once MPI reaches the proposed bounds and that under-reaching below the bounds causes large errors.

Significance. If established, the proposed bounds would give practical, physics-based guidance for choosing MPI in GNN-based PDE solvers, reducing hyperparameter search. The paper has genuine strengths: the hyperbolic bound is derived from physical constants rather than fitted to error curves, the experiments use a shared-weight design to keep parameter count fixed, and the validation includes a realistic MeshGraphNet example and extrapolation tests. However, the parabolic/elliptic bound M = L/dx is not consistent with the graph topology used in the experiments, and the claimed sharpness is therefore not established. The central idea is useful, but the load-bearing bound for two of the three PDE classes needs substantial revision or qualification.

major comments (3)
  1. [§3.2–3.3, Eq. (8)] The bound M = L/dx is not a sufficient number of message-passing iterations for global information flow on the four-connected square grids implied by the uniform-grid finite-difference setup. On an N x N grid with N = L/dx, the graph diameter is 2(N-1) hops, so after M = N hops a node in one corner has received no information from the opposite corner. The authors' own criterion in §3.2 is that 'the state at every node … is influenced by all other nodes', and parabolic and elliptic Green's functions are nonzero everywhere, so a point source at a corner must influence the opposite corner. Consequently, Eq. (8) underestimates the all-to-all requirement by roughly a factor of two, and the saturation curves in Figs. 6, 8, and 10 cannot establish sharpness unless the tested problems have no significant corner-to-corner coupling. The authors should either replace Eq. (8) with the graph diameter (about 2L/dx), or explicitly restrict the claim to crossing the domain side-to-side and provide evidence that the omitted corner-to-corner dependence is negligible for their benchmarks.
  2. [§3.1, Eq. (5), Table 1] The time step dt entering Eq. (5) is inconsistent with the value reported in Table 1. For Wave-LowRes, substituting the tabulated values c = 0.5, dx = 0.04, dt = 0.001 gives ceil(sqrt(2)*0.5*0.001/0.04) = 1, not the claimed M = 4. The value M = 4 is obtained only if dt is the pruned snapshot interval, 0.2 s (2.0 s divided by the 10 training snapshots). The text never defines this effective prediction interval, and Table 1 labels dt as the finite-difference time step used to generate the data. The hyperbolic bound is therefore not reproducible as written; the authors must define the dt that enters Eq. (5) (e.g., an effective rollout interval) and reconcile it with the entries in Table 1.
  3. [§4.1, Table 1; §4.4.1; §4.4.3] The parabolic validation does not actually vary spatial resolution: both Fourier datasets have dx = 0.1, and the difference between M = 10 and M = 20 arises from changing the domain size from 1 x 1 to 2 x 2, yet §4.4.1 describes them as 'low and high resolution'. The table also lists both rows under the same name 'Fourier-LowRes'. In addition, for the plastic plate (a 3D tetrahedral mesh on a 1 x 1 x 0.1 domain), the values of L and dx that would enter Eq. (8) are not defined, and M = 15 is asserted without a clear calculation. These inconsistencies make the claimed universality of the geometric bound difficult to assess.
minor comments (5)
  1. [§4.4.2, Fig. 8 caption] The caption of Fig. 8 refers to 'n-step parabolic rollouts for the electrostatic potential problem', but Poisson's equation is elliptic and has no time variable or rollout; the wording should be corrected.
  2. [§4.1, Table 1] The table row naming should distinguish the two parabolic datasets by domain size (e.g., Fourier-1x1 and Fourier-2x2) rather than listing both as 'Fourier-LowRes'.
  3. [Eq. (10), §4.4.3] The normalized latent magnitude U_i^m divides by ||xi_i^0||_2, which could be near or exactly zero for some inputs; the authors should state how this case is handled or show that zero initial encodings do not occur.
  4. [Fig. 12] The color maps in Fig. 12 are independently normalized for each panel, which makes the claimed difference in latent-magnitude coverage harder to judge; a shared color scale would be more informative.
  5. [§2] The sentence about the result of using extended neighborhoods translating to a 'much higher order scheme and a highly populated stiffness matrix' is asserted without support or a reference; it should either be justified or removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MPI bounds are derived from CFL stability and mesh geometry, and the experiments test the a priori bounds rather than fit them.

full rationale

The claimed lower bounds are derived from physical constants and mesh geometry rather than from the error curves. For hyperbolic equations, Eq. (6) re-expresses the CFL condition as a message-propagation requirement (M Δx > √2 c Δt), and the experiments vary MPI and compare against this a priori bound. For parabolic and elliptic equations, Eq. (8) sets M = L/Δx as the hop distance needed to cross the domain; although this is a geometric definition of 'global reach', the paper then tests whether smaller MPI fails and whether performance saturates at the bound, so the sharpness claim is an empirical check, not a consequence of the definition. The thermodynamic-GNN citations ([14], [28]-[35]) are background and do not enter the bound derivation. No fitted parameter is relabeled as a prediction, and no load-bearing uniqueness theorem is imported from the authors' prior work. A possible issue is that Eq. (6) as written does not numerically reproduce the stated wave examples in Table 1, but that is a correctness or arithmetic concern, not circularity. No significant circularity found.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The bounds are parameter-free, using physical constants from the problem setup. The key postulates are the global-view requirement for parabolic/elliptic problems and the diamond/disk containment for hyperbolic problems. No new physical entities are introduced.

assumptions (2)
  • domain assumption For parabolic and elliptic PDEs, every node must acquire a global view of the system state, formalized as receiving a message from the opposite boundary.
    Section 3.2 and 3.3: the paper states 'the entire domain must be informed about the origin of the perturbation at each temporal step' and equates this with M = L/dx.
  • domain assumption Message passing on a 4-connected uniform grid expands the receptive field as an L1 diamond of radius M dx, and the hyperbolic wave's domain of influence is a disk of radius c dt; the bound M > sqrt(2) c dt/dx follows from containing the disk in the diamond.
    Section 3.1, Eq (5)-(6): the CFL analogy is used to relate M to the physical propagation distance, but the geometric model of the receptive field is assumed, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the under-reaching phenomenon in message-passing neural PDE solvers: revisiting the CFL condition." pith.science (2026). https://pith.science/paper/3ZMGMAS7

@misc{pith2026250708861,
  author       = {Pith},
  title        = {Pith review of: On the under-reaching phenomenon in message-passing neural PDE solvers: revisiting the CFL condition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZMGMAS7}},
  note         = {Machine review of arXiv:2507.08861}
}
read the original abstract

This paper proposes sharp lower bounds for the number of message passing iterations required in graph neural networks (GNNs) when solving partial differential equations (PDE). This significantly reduces the need for exhaustive hyperparameter tuning. Bounds are derived for the three fundamental classes of PDEs (hyperbolic, parabolic and elliptic) by relating the physical characteristics of the problem in question to the message-passing requirement of GNNs. In particular, we investigate the relationship between the physical constants of the equations governing the problem, the spatial and temporal discretisation and the message passing mechanisms in GNNs. When the number of message passing iterations is below these proposed limits, information does not propagate efficiently through the network, resulting in poor solutions, even for deep GNN architectures. In contrast, when the suggested lower bound is satisfied, the GNN parameterisation allows the model to accurately capture the underlying phenomenology, resulting in solvers of adequate accuracy. Examples are provided for four different examples of equations that show the sharpness of the proposed lower bounds.

Figures

Figures reproduced from arXiv: 2507.08861 by the authors.

Figure 1
Figure 1. Illustration of a GNN with an encoder-processor-decoder approach and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Temporal evolution of one of the wave problems computed using finite difference methods with two distinct [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Evolution of the von Mises stress for the incremental forming problem simulated with finite element methods. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Relative Root Mean Square Error (RRMSE) over three random seeds for [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Comparison of GNN prediction performance across different time steps and message passing iterations for the [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Error over five random seeds for n-step parabolic rollouts. Dashed lines represent the suggested lower bound. Left: 2×2 domain, giving an optimal number of message passing iterations of M = 20. Right: 1×1 domain, thus giving M = 10. Errors remain virtually stable from …
Figure 7
Figure 7. Figure 7: Comparison of GNN performance for the Fourier problem. (a) Optimal design with [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Relative root mean square error over four random seeds for [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Comparison of GNN performance as dynamic parabolic integrators, with lower bound set to 20. (a) [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Error over five random seeds for n-step predictions in the incremental forming problem, showing cumulative error across rollout. The dashed vertical line represents the suggested lower bound for the number of message passing iterations, M. When the model’s design is b…
Figure 11
Figure 11. Figure 11: Comparison of GNN performance as as a function of the number of passes for the incremental forming [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: The color map represents the normalized latent vector magnitude [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Relative Root Mean Square Error (RRMSE) over a test extrapolation out of the training domain for the [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: A comparison between the dynamic rollout predicted by the Graph Neural Network (GNN) and the numerical [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: A comparison between the dynamic rollout predicted by the Graph Neural Network (GNN) and the numerical [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages

  1. [1]

    Machine learning in computer aided engineering

    Francisco J Montáns, Elías Cueto, and Klaus-Jürgen Bathe. Machine learning in computer aided engineering. In Machine Learning in Modeling and Simulation: Methods and Applications, pages 1–83. Springer, 2023

    work page 2023

  2. [2]

    Learning mesh-based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2021

    work page 2021

  3. [3]

    Physics-based machine learning for computational fracture mechanics

    Fadi Aldakheel, Elsayed S Elsayed, Yousef Heider, and Oliver Weeger. Physics-based machine learning for computational fracture mechanics. Machine Learning for Computational Science and Engineering, 1(1):18, 2025

    work page 2025

  4. [4]

    Flowgnn: A dataflow architecture for real-time workload-agnostic graph neural network inference

    Rishov Sarkar, Stefan Abi-Karam, Yuqi He, Lakshmi Sathidevi, and Cong Hao. Flowgnn: A dataflow architecture for real-time workload-agnostic graph neural network inference. In 2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pages 1099–1112, 2023

    work page 2023

  5. [5]

    Zahr, and Jian-Xun Wang

    Han Gao, Matthew J. Zahr, and Jian-Xun Wang. Physics-informed graph neural galerkin networks: A unified framework for solving pde-governed forward and inverse problems. Computer Methods in Applied Mechanics and Engineering, 390:114502, 2022

    work page 2022

  6. [6]

    R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. Fortunato, F. Alet, and P. Battaglia. Graphcast: Learning skillful medium-range global weather forecasting. arXiv preprint arXiv:2212.12794, 2022

    arXiv 2022

  7. [7]

    Calculus of Finite Differences, volume 33

    Károly Jordán. Calculus of Finite Differences, volume 33. American Mathematical Society, 1965. 18 On the under-reaching phenomenon in Graph Neural Networks for physical simulations

    work page 1965

  8. [8]

    Finite element method

    Klaus Jurgen Bathe. Finite element method. Wiley encyclopedia of computer science and engineering, pages 1–12, 2007

    work page 2007

Show all 47 references
  1. [9]

    Finite volume methods

    Robert Eymard, Thierry Gallouët, and Raphaèle Herbin. Finite volume methods. Handbook of numerical analysis, 7:713–1018, 2000

    2000

  2. [10]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics- informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021

    2021

  3. [11]

    A review of graph neural network applications in mechanics-related domains

    Yingxue Zhao, Haoran Li, Haosu Zhou, Hamid Reza Attar, Tobias Pfaff, and Nan Li. A review of graph neural network applications in mechanics-related domains. Artificial Intelligence Review, 57(11):315, 10 2024

    2024

  4. [12]

    Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff

    Kelsey R. Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff. Physical design using differentiable learned simulators, 2022

    2022

  5. [13]

    Allen, Yulia Rubanova, Tatiana Lopez-Guevara, William Whitney, Alvaro Sanchez-Gonzalez, Peter Battaglia, and Tobias Pfaff

    Kelsey R. Allen, Yulia Rubanova, Tatiana Lopez-Guevara, William Whitney, Alvaro Sanchez-Gonzalez, Peter Battaglia, and Tobias Pfaff. Learning rigid dynamics with face interaction graph networks, 2022

    2022

  6. [14]

    Iparraguirre, Icíar Alfaro, David González, Francisco Chinesta, and Elías Cueto

    Alicia Tierz, Mikel M. Iparraguirre, Icíar Alfaro, David González, Francisco Chinesta, and Elías Cueto. On the feasibility of foundational models for the simulation of physical phenomena. International Journal for Numerical Methods in Engineering, 126(6):e70027, 2025

    2025

  7. [15]

    Scientific machine learning through physics–informed neural networks: Where we are and what’s next

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3):88, 2022

    2022

  8. [16]

    Physics-informed neural networks (pinns) for fluid mechanics: a review

    Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (pinns) for fluid mechanics: a review. Acta Mechanica Sinica, 37(12):1727–1738, 2021

    2021

  9. [17]

    Dpm: A novel training method for physics-informed neural networks in extrapolation

    Jungeun Kim, Kookjin Lee, Dongeun Lee, Sheo Yon Jhin, and Noseong Park. Dpm: A novel training method for physics-informed neural networks in extrapolation. 35:8146–8154, May 2021

    2021

  10. [18]

    Understanding and mitigating extrapolation failures in physics-informed neural networks, 2023

    Lukas Fesser, Luca D’Amico-Wong, and Richard Qiu. Understanding and mitigating extrapolation failures in physics-informed neural networks, 2023

    2023

  11. [19]

    A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes

    Stefania Fresca, Luca Dede’, and Andrea Manzoni. A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes. Journal of Scientific Computing, 87(2):61, 2021

    2021

  12. [20]

    Pod-dl-rom: Enhancing deep learning-based reduced order models for nonlinear parametrized pdes by proper orthogonal decomposition

    Stefania Fresca and Andrea Manzoni. Pod-dl-rom: Enhancing deep learning-based reduced order models for nonlinear parametrized pdes by proper orthogonal decomposition. Computer Methods in Applied Mechanics and Engineering, 388:114181, 2022

    2022

  13. [21]

    Handling geometrical variability in nonlinear reduced order modeling through continuous geometry-aware dl-roms

    Simone Brivio, Stefania Fresca, and Andrea Manzoni. Handling geometrical variability in nonlinear reduced order modeling through continuous geometry-aware dl-roms. Computer Methods in Applied Mechanics and Engineering, 442:117989, 2025

    2025

  14. [22]

    Gradient-based learning applied to document recognition

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  15. [23]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges, chapter GeometricDomains: the 5Gs, page 50. arXiv, 2021

    2021

  16. [24]

    Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst

    Michael M. Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: Going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017

    2017

  17. [25]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S. Yu. A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 32(1):4–24, 2021

    2021

  18. [26]

    Unravelling the performance of physics-informed graph neural networks for dynamical systems

    Abishek Thangamuthu, Gunjan Kumar, Suresh Bishnoi, Ravinder Bhattoo, N M Anoop Krishnan, and Sayan Ranu. Unravelling the performance of physics-informed graph neural networks for dynamical systems. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, ...

    2022

  19. [27]

    Wendi Liu and Michael J. Pyrcz. Physics-informed graph neural network for spatial-temporal production forecasting. Geoenergy Science and Engineering, 223:211486, 2023

    2023

  20. [28]

    Thermodynamics-informed graph neural networks

    Quercus Hernandez, Alberto Badias, Francisco Chinesta, and Elias Cueto. Thermodynamics-informed graph neural networks. IEEE Transactions on Artificial Intelligence, 2022

    2022

  21. [29]

    Structure-preserving neural networks

    Quercus Hernández, Alberto Badías, David González, Francisco Chinesta, and Elías Cueto. Structure-preserving neural networks. Journal of Computational Physics, 426:109950, 2021. 19 On the under-reaching phenomenon in Graph Neural Networks for physical simulations

    2021

  22. [30]

    Port-metriplectic neural networks: thermodynamics-informed machine learning of complex physical systems

    Quercus Hernández, Alberto Badías, Francisco Chinesta, and Elías Cueto. Port-metriplectic neural networks: thermodynamics-informed machine learning of complex physical systems. Computational Mechanics, 72(3):553– 561, 2023

    2023

  23. [31]

    Thermodynamics-informed neural networks for physically realistic mixed reality

    Quercus Hernández, Alberto Badías, Francisco Chinesta, and Elías Cueto. Thermodynamics-informed neural networks for physically realistic mixed reality. Computer Methods in Applied Mechanics and Engineering , 407:115912, 2023

    2023

  24. [32]

    Graph neural networks informed locally by thermodynamics

    Alicia Tierz, Iciar Alfaro, David González, Francisco Chinesta, and Elías Cueto. Graph neural networks informed locally by thermodynamics. Engineering Applications of Artificial Intelligence, 144:110108, 2025

    2025

  25. [33]

    Reversible and irreversible bracket-based dynamics for deep graph neural networks

    Anthony Gruber, Kookjin Lee, and Nathaniel Trask. Reversible and irreversible bracket-based dynamics for deep graph neural networks. Advances in Neural Information Processing Systems, 36:38454–38484, 2023

    2023

  26. [34]

    A thermodynamics-informed active learning approach to perception and reasoning about fluids

    Beatriz Moya, Alberto Badías, David González, Francisco Chinesta, and Elías Cueto. A thermodynamics-informed active learning approach to perception and reasoning about fluids. Computational Mechanics, 72(3):577–591, 2023

    2023

  27. [35]

    Thermodynamics- informed super-resolution of scarce temporal dynamics data

    Carlos Bermejo-Barbanoj, Beatriz Moya, Alberto Badías, Francisco Chinesta, and Elías Cueto. Thermodynamics- informed super-resolution of scarce temporal dynamics data. Computer Methods in Applied Mechanics and Engineering, 430:117210, 2024

    2024

  28. [36]

    Hyperparameter Optimization, pages 3–33

    Matthias Feurer and Frank Hutter. Hyperparameter Optimization, pages 3–33. 05 2019

    2019

  29. [37]

    Breaking the limits of message passing graph neural networks

    Muhammet Balcilar, Pierre Heroux, Benoit Gauzere, Pascal Vasseur, Sebastien Adam, and Paul Honeine. Breaking the limits of message passing graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volu...

    2021

  30. [38]

    Peter D. Lax. Hyperbolic Systems of Conservation Laws and the Mathematical Theory of Shock Waves. Number 11 in Conference Board of the Mathematical Sciences Regional Conference Series in Applied Mathematics. Society for Industrial and Applied Mathematics, Philadelphia, PA, 1973

    1973

  31. [39]

    The paradox of fourier heat equation: A theoretical refutation

    J-L Auriault. The paradox of fourier heat equation: A theoretical refutation. International Journal of Engineering Science, 118:82–88, 2017

    2017

  32. [40]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Message Passing Neural Networks, pages 199–214. Springer International Publishing, Cham, 2020

    2020

  33. [41]

    Nodemixup: Tackling under-reaching for graph neural networks

    Weigang Lu, Ziyu Guan, Wei Zhao, Yaming Yang, and Long Jin. Nodemixup: Tackling under-reaching for graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 14175–14183, 2024

    2024

  34. [42]

    Message passing neural pde solvers

    Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022

    2022 arXiv

  35. [43]

    The Courant–Friedrichs–Lewy (CFL) condition

    Carlos A De Moura and Carlos S Kubrusly. The Courant–Friedrichs–Lewy (CFL) condition. AMC, 10(12):45–90, 2013

    2013

  36. [44]

    On the speed of heat

    V Bertola and E Cafaro. On the speed of heat. Physics Letters A, 372(1):1–4, 2007

    2007

  37. [45]

    Understanding over-squashing and bottlenecks on graphs via curvature

    Jake Topping, Francesco Di Giovanni, Benjamin Paul Chamberlain, Xiaowen Dong, and Michael M Bronstein. Understanding over-squashing and bottlenecks on graphs via curvature. In International Conference on Learning Representations, 2021

    2021

  38. [46]

    Limits of depth: Over-smoothing and over-squashing in gnns

    Shaima Qureshi et al. Limits of depth: Over-smoothing and over-squashing in gnns. Big Data Mining and Analytics, 7(1):205–216, 2023

    2023

  39. [47]

    Limiting over-smoothing and over- squashing of graph message passing by deep scattering transforms

    Yuanhong Jiang, Dongmian Zou, Xiaoqun Zhang, and Yu Guang Wang. Limiting over-smoothing and over- squashing of graph message passing by deep scattering transforms. arXiv preprint arXiv:2407.06988, 2024. 20

    2024 arXiv

Pith tools

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