Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Automated discovery of finite volume schemes using Graph Neural Networks

T0 review · 3 major / 5 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 A GNN trained on two-node graphs alone provably implements the standard finite-volume scheme for the heat equation: training error ε means O(ε) error on any unseen mesh.

desk verdict Promising empirical idea undermined by a vacuous main theorem and inconsistent theory. read the letter →

arxiv 2508.19052 v1 pith:3SKGKOCY submitted 2025-08-26 cs.LG physics.comp-ph

Paul Garnier , Jonathan Viquerat , Elie Hachem This is my paper
classification cs.LGphysics.comp-ph MSC 65M0868T0765M12
keywords graphneuralnetworksfinitevolumemethodheatequationsymbolicregressionphysics-informedout-of-distributiongeneralizationnumericalschemediscoverymessagepassing
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 tries to establish that graph neural networks can do more than approximate physics: trained correctly, they can rediscover — and even extend — the numerical schemes that solvers use. The central result is a proof that a GNN trained only on graphs of two cells with one shared face, under a sparsity-regularized loss, implements the standard first-order finite-volume (FV) update for the heat equation on any unstructured mesh: training error ε on the minimal dataset implies error at most a constant times ε on every unseen mesh. Symbolic regression confirms the match is exact, recovering the textbook FV flux formula from the network's components, both when training is supervised and when it uses only the PDE residual (a physics-informed loss). Widening the network's receptive field then yields genuinely new schemes: a 2-hop variant learns a second-order correction term, and a two-layer variant rediscovers the classic explicit midpoint (second-order Runge–Kutta) scheme, both more accurate than the first-order baseline. If the claims hold, minimal data plus sparsity can turn a GNN into an automated discoverer of numerical methods with transferable guarantees.

What carries the argument

The argument turns on aligning the GNN's update structure with the finite-volume flux exchange. The message-passing architecture splits into an edge MLP that computes a pairwise flux and a node MLP that sums aggregated fluxes into the temperature update — the same shape as the FV rule T ← T + (Δt/V)Σ flux. A sparsity penalty (L0 in theory, L1 in practice) on the weights, together with a parameter-counting lemma, pins any low-loss network to the minimal sparse configuration, which turns out to be exactly the FV scheme; this pinning is what converts small training error into a global O(ε) bound. The two-node training graph, with geometry chosen so the face-area-to-distance factor cancels the c

What would settle it

Find any sparse ReLU GNN fitting the two-node training set to loss below ε with fewer non-zero weights than the claimed N_sparse = 2m(L+1)+4, or with weights outside the zero-bias diagonal family, and show its one-step error on a small unstructured mesh exceeds 2ε(1 + 2/Δt); a pruning or evolutionary search over general weight structures at the training-loss optimum would either confirm or refute the parameter-counting premise.

Watch

Extended reading notes

Core claim

Central claim (Theorem 4.3, appendix C.6): if an L-layer GNN with ReLU MLPs is trained on two-node graphs with a sparsity-regularized loss and achieves loss < ε (ε < η), then for every graph G and node r, ∥Mθ(G)_r − F(G)_r∥₁ < 2ε(1 + 2/Δt), with F the first-order finite-volume heat step. Small training error on the two-cell dataset therefore certifies O(ε) accuracy on any unseen mesh: sparsity forces the weights into the unique minimal configuration that fits the data, which is numerically the FV scheme. Symbolic regression confirms the match exactly — the learned edge function is α(A/δ)(T_N − T_P) and the node update is T_P + (Δt/V_P)Σfluxes + Δt S_P. With a pure PDE-residual (PINN) loss th

Load-bearing premise

The proof rests on the claim that the only way for a sparse ReLU network to fit the two-node training set is to take one specific parameter configuration — the finite-volume form — so a network with fewer or differently arranged parameters cannot match the data well.

Editorial extensions

If this is right

  • Learned schemes inherit a transfer guarantee: a small loss on a two-node dataset certifies O(ε) accuracy on any mesh satisfying the Fourier stability condition, so minimal datasets can replace large simulation corpora for training PDE solvers.
  • The physics-informed variant means schemes can be derived from the PDE alone, with no ground-truth solver data — opening the same pipeline to equations whose discretizations are not yet known.
  • Wider receptive fields produce higher-order methods: the discovered 2-hop correction and the explicit midpoint scheme beat the first-order baseline, and the midpoint scheme matches Crank–Nicolson accuracy at lower memory cost.
  • Because symbolic regression returns closed forms, discovered schemes can be compiled into classical solver code or inspected by numerical analysts, closing the loop between learning and analysis.
  • The theory suggests sparsity is not a practical nicety but the load-bearing ingredient for out-of-distribution generalization: without the L1 penalty, training loss near zero on the minimal set leads to large OOD and rollout errors.

Reading between the lines

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

  • The transfer bound depends on the parameter-counting lemma, which is proven for a restricted zero-bias diagonal family of ReLU configurations; whether a different low-parameter structure can fit the two-cell dataset and escape the O(ε) bound is left open — an exhaustive search over sparse architectures at the claimed threshold would settle it.
  • The discovered 2-hop term, (1/2)α(T_Q − T_P) summed over second neighbors, reads like a discrete biharmonic (Laplacian-squared) correction; testing it on anisotropic diffusion or time-varying meshes would show whether it generalizes beyond the heat equation studied here.
  • The same minimal-dataset argument should carry over to other conservation laws whose FV updates are sums of pairwise fluxes (advection, nonlinear diffusion), since nothing in the proof uses linearity of the heat equation specifically beyond the flux form.
  • The O(ε) constant grows as 1/Δt, so the certified regime is exactly the explicit-marching regime; pushing the method to implicit or larger-timestep schemes would need a different stability argument than the Fourier condition used here.
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 studies whether graph neural networks (GNNs) trained on trivial two-cell graphs can learn and extrapolate finite-volume (FV) discretizations of the heat equation. It states a theorem (Thm 4.3 / C.6) that if a sparse-regularized GNN attains loss ε < η on a two-node dataset, then its prediction error on any unseen mesh is O(ε). It reports experiments with supervised and PINN losses, uses symbolic regression to claim rediscovery of the first-order FV update, and further claims discovery of a 2-hop second-order correction and an explicit midpoint scheme. The empirical design is clean and the symbolic-regression checks are a useful interpretability device, but the theoretical guarantee is vacuous as stated and the PINN loss is inconsistent with the explicit schemes claimed to be discovered.

Significance. If the results were valid, the paper would provide an important bridge between neural algorithmic alignment and numerical analysis, with a concrete route to interpretable scheme discovery. The two-cell training setup, the out-of-distribution and rollout comparisons, and the use of symbolic regression as an inspection tool are all methodologically appealing; I credit the authors for making the target scheme explicit and for checking the learned update symbolically. However, the central theorem is vacuous because of the integer L0 regularizer, the sparsity lower-bound proof has unresolved parameter-count inconsistencies, and the PINN loss in Eq. (17), as written, cannot be minimized by the explicit schemes the paper reports. These are load-bearing issues for the paper's main claims, not presentation concerns.

major comments (3)
  1. [Theorem 4.3 / Appendix C.6, Eq. (21)] The premise of Theorem 4.3 is impossible for the stated loss. L(Mθ,Dtrain)=L_MAE+η||θ||0, and ||θ||0 is a nonnegative integer. If L<ε<η, then η||θ||0<η, hence ||θ||0=0. A zero-parameter network has zero weights and biases, so its prediction is constant zero, but Dtrain contains G(1,1,0,0) whose FV target is 1 (Sec. 4.1.3 and Remark 4.1). Thus no model satisfies the hypothesis, and the OOD bound (22) is vacuously true. The same integer-sparsity problem invalidates Theorem C.5's claim that ηN_sparse is the global minimum. To obtain a non-vacuous statement, the authors must decouple the MAE condition from the sparsity count, e.g., state 'if L_MAE<ε and ||θ||0=N_sparse, then ...', and then prove that version.
  2. [Lemmas C.1 and C.4, Theorem C.5] The parameter count underlying the sparsity argument is internally inconsistent. Lemma C.1 states that the sparse construction uses 2m(L+1)+4 non-zero parameters, but the proof's own construction (2m(L−1) for identity layers, 2m for the aggregation network, and 3+(m−1) for the update network) totals 2mL+m+2. Lemma C.4 then argues for a lower bound of 2mL+m+4 and later refers to 2mL+2. The proof of Lemma C.4 is a sign-preservation heuristic for zero-bias diagonal constructions, not a general lower bound over all ReLU MLP configurations, widths, and layer layouts. Since Theorem C.5 and C.6 both depend on Lemma C.4, the O(ε) transfer to arbitrary graphs is not established even after the vacuity issue is repaired.
  3. [Eq. (17), Remark 4.4, Sec. 5.2] The unsupervised/PINN claims are inconsistent with the loss as written. Eq. (17) computes the PDE residual using ∇²h_L, i.e., the Laplacian of the network's output (the next state). An explicit first-order FV update, or an explicit midpoint scheme, satisfies (h_L−h0)/Δt ≈ S+α∇²h0, not S+α∇²h_L; substituting the claimed schemes into Eq. (17) leaves a nonzero residual of order Δt or larger. Minimizing Eq. (17) should drive the network toward an implicit update, not toward the explicit schemes reported in Remark 4.4 and Eqs. (26)–(27). The authors need to specify how ∇² is discretized on graphs and on which temperature state it acts; otherwise the symbolic-regression results under PINN training do not support the claim that minimizing Eq. (17) rediscovers the explicit first-order or midpoint schemes.
minor comments (5)
  1. [Throughout] Typos and language issues: 'These findings follows', 'performancess', 'majoration constant'. The footnote about 'large modern transformers' is unsupported and not relevant to the paper's claims.
  2. [Eq. (17)] The symbol ∇²h_L is not defined for graphs. Please define the discrete graph Laplacian or the finite-volume operator used, including boundary treatment.
  3. [Lemma C.4] The statement 'cannot have a non-zero error' is ambiguous; presumably it means 'cannot achieve zero MAE' or 'must have non-zero MAE'. Please rephrase.
  4. [Sec. 5.2] The symbolic-regression step for the 2-hop aggregator and the midpoint network is described only through final formulas. Please report the Pareto fronts, coefficient values with uncertainties, and the number of training runs to show that the 1/2 coefficient and the midpoint structure are robust discoveries rather than post hoc selections.
  5. [Fig. 3] The numerical comparison in Section 4.4 appears to be a single-seed study. Given the small training set, error bars over multiple seeds would strengthen the empirical claim that sparsity, not chance, drives OOD generalization.

Circularity Check

2 steps flagged · score 6.0 of 10

The central OOD theorem is vacuous in the regime where it is used, and the supervised 'rediscovery' of the FV scheme is a fit to the label-generating formula.

  1. self definitional [Section 4.3, Eq. (21)-(22), Theorem 4.3; Appendix C Theorems C.5-C.6]
    "We define the training loss as: L(Mθ, Dtrain) = LMAE(Mθ, Dtrain) + η||θ||0 (21) ... Let ε > 0, with ε < η. If the trained model achieves a loss L(Mθ, Dtrain) < ε on the two-node graph dataset, then for any graph G, the model’s prediction error is bounded: ∥Mθ(G)_r − F(G)_r∥_1 < 2ε(1 + 2/Δt) (22)"

    For any model, ||θ||0 is a non-negative integer. If ||θ||0 ≥ 1 then L ≥ η, so L < ε < η is impossible. Hence any model satisfying the premise must have ||θ||0 = 0 and L = LMAE. A zero-parameter GNN outputs 0 everywhere, while Dtrain contains G(1,1,0,0) with F(G)_r = 1, so LMAE = 1; the premise then requires 1 < ε < η, i.e. η > 1. In the paper's own sparse regime (Sec. 4.4 uses η = 10^{-3}) no model satisfies the hypothesis at all. The O(ε) OOD bound is therefore vacuously true rather than a property of learned FV weights. The same self-defeating premise underlies Theorem C.5's 'ηNsparse is the global minimum' and Theorem C.6's bound.

  2. fitted input called prediction [Section 4.1.3 (dataset) and Section 4.6 (symbolic-regression results), Eq. (11)]
    "The ground truth for each graph G is the temperature at the next timestep, computed using the FV scheme from (11), denoted F(G). ... The best-fitting equations were: ... f_up_SR = T_r + Δt/V_r \bar{e}_r + Δt·S_r ... which is precisely the update defined in equation (11)."

    The supervised 'rediscovery' is a fit to the same formula that produced the training targets. The two-node dataset is labeled by the first-order FV update (11), the corner cases are explicitly chosen to constrain the learned function (Lemma B.1 forces the weights using G(0,0,0,0), G(1,1,0,0), G(0,0,1,1), G(0,1,0,0)), and symbolic regression is then run on the trained MLPs' input-output pairs and returns exactly (11). Recovering the generator of one's own training labels is expected, not an independent OOD confirmation; the paper's claim that the network 'did not merely approximate the function but learned the valid underlying physical equation' rests on the vacuous Theorem 4.3. The higher-order PINN experiments (Sec. 5) are not circular, because they use the PDE residual as an external obj

full rationale

The first-order supervised chain is the circular core. Labels are computed from Eq. (11); the corner-case graphs are engineered to force the FV form; symbolic regression returns Eq. (11); and the theorem that would convert this into an OOD guarantee (Theorem 4.3/C.6) has an unsatisfiable premise in the intended small-η regime: the L0 term makes L ≥ η for every nonzero-parameter model and L = 1 for the zero-parameter model, so with ε < η (and η = 10^{-3} in Sec. 4.4) no model can satisfy the hypothesis. Thus the claimed 'prediction' of the first-order FV scheme reduces by construction, and the formal bound is vacuous where invoked. The unsupervised higher-order section (GNN-2ndOrder and GNN-Midpoint, Eqs. (26)-(27)) is anchored to the PINN residual LPINN and is independently testable, which prevents a higher score. No load-bearing self-citation chain is present; refs. [14,15,25] are contextual, and the algorithmic-alignment citations [42,43,11,33] are not from the present authors. Lemma C.4's parameter count also appears inconsistent with Lemma C.1's construction (2mL+m+2 vs Nsparse), but that is a separate correctness concern rather than circularity. Overall: partial circularity, score 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

Everything the central claims rest on: the supervised claim rides on the corner-case training set and the Fourier stability bound; the theoretical guarantee rides on an unproven parameter-counting lower bound for ReLU MLPs; the higher-order discovery rides on symbolic-regression fits whose coefficients are not independently derived, and on a physics-informed loss whose printed form is inconsistent with the claimed recovery.

free parameters (3)
  • c1 in f_agg^SR (thermal diffusivity α recovered) = α (recovered as c1 by PySR)
    Symbolic regression fits a constant c1 multiplying A_PN/δ_PN (T_N - T_P); the paper identifies c1 with α. This is fitted to the trained MLP outputs, not predicted independently.
  • coefficient 1/2 in the 2-hop correction term (Eq. 26) = 1/2 α
    The symbolic regression extracts the term 1/2 α (T_Q - T_P) from f_agg2 of the trained GNN-2ndOrder; the coefficient is a fitted value, and no derivation justifies its magnitude or geometric independence.
  • coefficients in GNN-Midpoint scheme (1/2 and full step) = 1/2 and 1
    These are the standard explicit midpoint coefficients which emerged from training; they are fitted network weights extracted by symbolic regression rather than derived from a stability/consistency analysis in the paper.
assumptions (4)
  • domain assumption Fourier stability condition (Lemma A.3): Δt ≤ V_P / Σ_N α A_PN/δ_PN for every cell P
    Used throughout proofs (e.g., Appendix B, Eq. 33) to bound Σ|β_ij| ≤ 2V_P/Δt; the target problem is assumed to operate within the stable time-step regime.
  • ad hoc to paper Training set Dtrain includes the corner-case graphs G(0,0,0,0), G(1,1,0,0), G(0,0,1,1), G(0,1,0,0) (Sec. 4.1.3)
    The weight-constraining inequalities in Lemmas B.1/B.3 and Theorem C.5 require these exact deterministic graphs to be part of Dtrain with loss below ε.
  • ad hoc to paper The L0-regularized loss landscape satisfies η < min_{K<N_sparse} C(K)/(N_sparse - K) (Theorem C.5)
    The optimality proof assumes a separation between achievable MAE levels at different parameter counts; C(K) is never computed or bounded, so this assumption is unverified.
  • ad hoc to paper ReLU MLPs with zero biases require two non-zero weights per layer to propagate signed information (used in Lemma C.1 and C.4)
    The parameter-counting lower bound assumes this restriction; it is not a general property of ReLU networks, which can use biases, wider layers, or shared weights, so the lower bound is not established for arbitrary configurations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated discovery of finite volume schemes using Graph Neural Networks." pith.science (2026). https://pith.science/paper/3SKGKOCY

@misc{pith2026250819052,
  author       = {Pith},
  title        = {Pith review of: Automated discovery of finite volume schemes using Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3SKGKOCY}},
  note         = {Machine review of arXiv:2508.19052}
}
abstract

Graph Neural Networks (GNNs) have deeply modified the landscape of numerical simulations by demonstrating strong capabilities in approximating solutions of physical systems. However, their ability to extrapolate beyond their training domain (\textit{e.g.} larger or structurally different graphs) remains uncertain. In this work, we establish that GNNs can serve purposes beyond their traditional role, and be exploited to generate numerical schemes, in conjunction with symbolic regression. First, we show numerically and theoretically that a GNN trained on a dataset consisting solely of two-node graphs can extrapolate a first-order Finite Volume (FV) scheme for the heat equation on out-of-distribution, unstructured meshes. Specifically, if a GNN achieves a loss $\varepsilon$ on such a dataset, it implements the FV scheme with an error of $\mathcal{O}(\varepsilon)$. Using symbolic regression, we show that the network effectively rediscovers the exact analytical formulation of the standard first-order FV scheme. We then extend this approach to an unsupervised context: the GNN recovers the first-order FV scheme using only a residual loss similar to Physics-Informed Neural Networks (PINNs) with no access to ground-truth data. Finally, we push the methodology further by considering higher-order schemes: we train (i) a 2-hop and (ii) a 2-layers GNN using the same PINN loss, that autonomously discover (i) a second-order correction term to the initial scheme using a 2-hop stencil, and (ii) the classic second-order midpoint scheme. These findings follows a recent paradigm in scientific computing: GNNs are not only strong approximators, but can be active contributors to the development of novel numerical methods.

Figures

Figures reproduced from arXiv: 2508.19052 by the authors.

Figure 1
Figure 1. (Top left) Each edge is updated using its features and the features of its connected nodes. (Bottom left) Each node is updated using its features and the aggregated features from its incoming edges. (Right) A schematic of the information flow from a node’s perspective after one message passing step. 3 Graph neural networks We now consider a mesh as an undirected graph G “ pV, Eq. V “ tviui“1:N is the set of nodes, w… view at source ↗
Figure 2
Figure 2. A graph from the training set made from a 2-cells mesh. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. We display of the left the results with regularization and on the right the results without it. (Top) Performance metrics for models trained with and without L1 regularization. The regularized model (sparse GNN) shows significantly lower out-of-distribution MSE and 50-step rollout error compared to the non-regularized model, despite all models achieving near-zero training loss. (Bottom) Weight distributions for the … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Symbolic regression results for the f agg function on the left and f up function on the right. The plots show the trade-off between equation complexity (number of operations and variables) and accuracy. The optimal equations correspond exactly to the terms in the FVM, …
Figure 5
Figure 5. Figure 5: Sample meshes from the regular (left) and irregular (right) datasets. The colorscale indicates the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the 2-hop neighborhood used for the GNN architecture in the higher-order scheme [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Residual for the four studied schemes studied, computed for different mesh sizes and timestep [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: We display the memory consumption for each schemes, on different mesh sizes and timesteps. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages

  1. [1]

    Yohai Bar-Sinai, Stephan Hoyer, Jason Hickey, and Michael P. Brenner. Learning data-driven dis- cretizations for partial differential equations. Proceedings of the National Academy of Sciences (PNAS) , 116(31):15344–15349, 2019

    work page 2019

  2. [2]

    Battaglia, Jessica B

    Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gul- cehre, Francis Song, Andrew Ballard, Justin Gilmer, George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierst...

    arXiv 2018

  3. [3]

    Battaglia, Razvan Pascanu, Matthew Lai, David Rezende, and Koray Kavukcuoglu

    Peter W. Battaglia, Razvan Pascanu, Matthew Lai, David Rezende, and Koray Kavukcuoglu. Interaction networks for learning about objects, relations and physics. In Advances in Neural Information Processing Systems (NeurIPS), 2016

    work page 2016

  4. [4]

    Accurate medium-range global weather forecasting with 3d neural networks

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619:533–538, 2023

    work page 2023

  5. [5]

    Chang, Tomer D

    Michael B. Chang, Tomer D. Ullman, Antonio Torralba, and Joshua B. Tenenbaum. A neural physics engine for learning and reasoning. In International Conference on Machine Learning (ICML) , 2017

    work page 2017

  6. [6]

    Interpretable machine learning for science with pysr and symbolicregression.jl, 2023

    Miles Cranmer. Interpretable machine learning for science with pysr and symbolicregression.jl, 2023

    2023

  7. [7]

    Discovering symbolic models from deep learning with inductive biases

    Miles Cranmer, Alvaro Sanchez-Gonzalez, Peter Battaglia, Rui Xu, Kyle Cranmer, David Spergel, and Shirley Ho. Discovering symbolic models from deep learning with inductive biases. In Advances in Neural Information Processing Systems (NeurIPS) , 2020

    work page 2020

  8. [8]

    Dauphin, Angela Fan, Michael Auli, and David Grangier

    Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks, 2017

    work page 2017

Show all 52 references
  1. [9]

    Smith, Kelsey R

    Filipe de Avila Belbute-Peres, Kevin A. Smith, Kelsey R. Allen, J. B. Tenenbaum, and J. Zico Kolter. End- to-end differentiable physics for learning and control. In Advances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  2. [10]

    Finite volume informed graph neural network for myocardial perfusion simulation

    Raoul Sall´e de Chou, Matthew Sinclair, Sabrina Lynch, Nan Xiao, Laurent Najman, Ir`ene Vignon- Clementel, and Hugues Talbot. Finite volume informed graph neural network for myocardial perfusion simulation. In Medical Imaging with Deep Learning (MIDL) , 2024

    2024

  3. [11]

    Graph neural networks are dynamic programmers

    Andrew Dudzik and Petar Veliˇckovi´c. Graph neural networks are dynamic programmers. InAdvances in Neural Information Processing Systems (NeurIPS) , 2022

    2022

  4. [12]

    Deep fdm: Enhanced finite difference methods by deep learning

    Matthias Ehrhardt and Michael G¨unther. Deep fdm: Enhanced finite difference methods by deep learning. Franklin Open, 2(1):100039, 2023

    2023

  5. [13]

    How powerful are k-hop message passing graph neural networks, 2023

    Jiarui Feng, Yixin Chen, Fuhai Li, Anindya Sarkar, and Muhan Zhang. How powerful are k-hop message passing graph neural networks, 2023

    2023

  6. [14]

    Meshmask: Physics-based simulations with masked graph neural networks

    Paul Garnier, Vincent Lannelongue, Jonathan Viquerat, and Elie Hachem. Meshmask: Physics-based simulations with masked graph neural networks. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  7. [15]

    Multi-grid graph neural networks with self- attention for computational mechanics

    Paul Garnier, Jonathan Viquerat, and Elie Hachem. Multi-grid graph neural networks with self- attention for computational mechanics. arXiv preprint arXiv:2409.11899, 2024

    2024 arXiv

  8. [16]

    Gmsh: A 3-d finite element mesh generator with built-in pre- and post-processing facilities

    Christophe Geuzaine and Jean-Franc ¸ois Remacle. Gmsh: A 3-d finite element mesh generator with built-in pre- and post-processing facilities. International Journal for Numerical Methods in Engineering, 79(11):1309–1331, May 2009. 16

    2009

  9. [17]

    Hachem, B

    E. Hachem, B. Rivaux, T. Kloczko, H. Digonnet, and T. Coupez. Stabilized finite element method for incompressible flows with high reynolds number. Journal of Computational Physics , 229(23):8643– 8665, 2010

    2010

  10. [18]

    Gaussian error linear units (gelus), 2023

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus), 2023

    2023

  11. [19]

    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

  12. [20]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

    2017

  13. [21]

    Neural relational inference for interacting systems

    Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning (ICML) , 2018

    2018

  14. [22]

    Stuart, and Anima Anandkumar

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew M. Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. Journal of Machine Learning Research , 24(23):1–55, 2023

    2023

  15. [23]

    Investigating the ability of pinns to solve burgers’ pde near finite-time blowup, 2024

    Dibyakanti Kumar and Anirbit Mukherjee. Investigating the ability of pinns to solve burgers’ pde near finite-time blowup, 2024

    2024

  16. [24]

    Graphcast: Learning skillful medium-range global weather forecasting

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and P...

    2023

  17. [25]

    Physics- informed machine learning for enhanced 4D flow

    Vincent Lannelongue, Paul Garnier, Aur´elien Larcher, Philippe Meliga, and Elie Hachem. Physics- informed machine learning for enhanced 4D flow. In 29th Congress of the European Society of Biomechanics, Edinburgh, United Kingdom, June 2024

    2024

  18. [26]

    Rediscovering orbital mechanics with machine learning.Monthly Notices of the Royal Astronomical Society, 514(3):3965–3977, 2022

    Pablo Lemos, Niall Jeffrey, Miles Cranmer, Shirley Ho, and Peter Battaglia. Rediscovering orbital mechanics with machine learning.Monthly Notices of the Royal Astronomical Society, 514(3):3965–3977, 2022

    2022

  19. [27]

    Finite volume graph network (fvgn): Predicting unsteady incompressible fluid dynamics with finite volume informed neural network

    Tianyu Li, Shufan Zou, Xinghua Chang, Laiping Zhang, and Xiaogang Deng. Finite volume graph network (fvgn): Predicting unsteady incompressible fluid dynamics with finite volume informed neural network. Physics of Fluids, 36(4):043601, 2024

    2024

  20. [28]

    Learning to solve pdes with finite volume-informed neural networks in a data-free approach

    Tianyu Li, Yiye Zou, Shufan Zou, Xinghua Chang, and Laiping Zhang. Learning to solve pdes with finite volume-informed neural networks in a data-free approach. Journal of Computational Physics , 530:113919, 2025

    2025

  21. [29]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations (ICLR) , 2021

    2021

  22. [30]

    Multipole graph neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Andrew Stuart, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations. In Advances in Neural Information Processing Systems (NeurIPS) , 2020

    2020

  23. [31]

    Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218–229, 2021

    2021

  24. [32]

    Tenenbaum, and Daniel L

    Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Li Fei-Fei, Joshua B. Tenenbaum, and Daniel L. K. Yamins. Flexible neural representation for physics prediction. In Advances in Neural Information Processing Systems (NeurIPS) , 2018

    2018

  25. [33]

    Nerem, Samantha Chen, Sanjoy Dasgupta, and Yusu Wang

    Robert R. Nerem, Samantha Chen, Sanjoy Dasgupta, and Yusu Wang. Graph neural networks extrapolate out-of-distribution for shortest paths. arXiv preprint arXiv:2503.19173, 2025. 17

    2025 arXiv

  26. [34]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. In Advances in Neural Information Processing Systems (NeurIPS) , 2022

    2022

  27. [35]

    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 (ICLR) , 2021

    2021

  28. [36]

    Karniadakis

    Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving pdes. Journal of Computational Physics, 378:686–707, 2019

    2019

  29. [37]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning (ICML), 2020

    2020

  30. [38]

    Graph networks as learnable physics engines for inference and control

    Alvaro Sanchez-Gonzalez, Nicolas Heess, Jost Tobias Springenberg, Josh Merel, Martin Riedmiller, Raia Hadsell, and Peter Battaglia. Graph networks as learnable physics engines for inference and control. arXiv preprint arXiv:1806.01242, 2018

    2018 arXiv

  31. [39]

    Distilling free-form natural laws from experimental data

    Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data. Science, 324(5923):81–85, 2009

    2009

  32. [40]

    Ai feynman: A physics-inspired method for symbolic regression

    Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020

    2020

  33. [41]

    Chang, Ashesh Rambachan, and Sendhil Mullainathan

    Keyon Vafa, Peter G. Chang, Ashesh Rambachan, and Sendhil Mullainathan. What has a foundation model found? using inductive bias to probe for world models, 2025

    2025

  34. [42]

    Neural algorithmic reasoning

    Petar Veli ˇckovi´c and Charles Blundell. Neural algorithmic reasoning. Patterns, 2(7):100273, 2021

    2021

  35. [43]

    Neural execution of graph algorithms

    Petar Veliˇckovi´c, Rex Ying, Matilde Padovano, Raia Hadsell, and Charles Blundell. Neural execution of graph algorithms. In International Conference on Learning Representations (ICLR) , 2020

    2020

  36. [44]

    Yixuan Wang, Ziming Liu, Zongyi Li, Anima Anandkumar, and Thomas Y. Hou. High precision pinns in unbounded domains: application to singularity formulation in pdes, 2025

    2025

  37. [45]

    Asymptotic self-similar blow-up profile for three-dimensional axisymmetric euler equations using neural networks, 2023

    Yongji Wang, Ching-Yao Lai, Javier G´omez-Serrano, and Tristan Buckmaster. Asymptotic self-similar blow-up profile for three-dimensional axisymmetric euler equations using neural networks, 2023. 18 Contents 1 Introduction 1 2 Numerical framework 3 2.1 The heat equation . . . ....

    2023

  38. [46]

    We start by explicitely writing out a GNN update using the different features and trainables weights in the network

  39. [47]

    We apply our main hypothesis on the training set on specific graphs, such as one with a unit temperature in one cell, and all source terms and the other cell temperature set at 0

  40. [48]

    Finally, we use those inequalities to bound the error on a general graph

    This gives us several inequalities on the networks weights. Finally, we use those inequalities to bound the error on a general graph. 21 Proof. We expand hp1q i and simplify it for a 2-cells graph from Dtrain: hp1q i “ TiW u 1 ` ∆tSiW u 3 ` ∆t Vi W u 2 ÿ j pwaggβij` baggq` bu ...

  41. [49]

    We start by showing that we can implement a perfect first order scheme using 2mL` 2m` 4 parameters (Lemma C.1)

  42. [50]

    This leads to Theorem C.5 where we demonstrate that a small training error requires at least 2mL` 2m` 4 parameters

    We then show that any solution with stricly less than 2mL` 2m` 4 parameters will achieve a higher loss (Lemma C.4). This leads to Theorem C.5 where we demonstrate that a small training error requires at least 2mL` 2m` 4 parameters

  43. [51]

    C.1 Sparsity consideration Lemma C.1

    Finally, we apply the previous results to vastly simplify the weights in the network, leading to a final proof very similar to the previous proof B.2. C.1 Sparsity consideration Lemma C.1. An L-layers GNN, with m-layers ReLU MLPs of width d can reproduce perfectly F on any gra...

  44. [52]

    W a p “ » ———– 1 0 0

    The key thing here is to notice that pTj´ Tiq can be negative as well and needs 2 non-null weights with opposite signs to produce: σpW a 1 βijq“r βij, 0, ...,0s ifpTj´ Tiqą 0 (61) “r 0, ∥βij∥1, ...,0s ifpTj´ Tiqă 0 (62) The remaining m´ 2 layers (all hidden layers) can then im...

Pith tools

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