Pith. sign in

REVIEW 4 major objections 8 minor 44 references

A data-free GNN trained on weak-form residuals solves 3D elasticity and generalises to unseen geometries.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:19 UTC pith:LKVMTZTR

load-bearing objection The weak-form GNN solver is a coherent and useful contribution, but the 'unseen geometry' claim currently rests on one geometry that only works after test-time refinement against the training loss — that needs fixing before publication. the 4 major comments →

arxiv 2607.27901 v1 pith:LKVMTZTR submitted 2026-07-30 cs.CE

Data-free neural PDE solvers based on Graph Neural Networks and weak forms

Mikel M. Iparraguirre , Iciar Alfaro , David Gonzalez , Elias Cueto This is my paper
classification cs.CE
keywords physics-informed neural networksgraph neural networksweak formfinite element methodhyperelasticityneural operatortest-time refinementMeshGraphNet-Transformer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a physics-informed graph-transformer network can act as a continuous neural operator for 3D hyperelastic solids without ever seeing simulation data. Instead of matching output to labels, it minimises the finite-element weak-form residual computed from the network's own displacement predictions, using shape-function gradients rather than automatic differentiation. The authors argue this sidesteps the data-generation bottleneck of learned solvers and the optimisation pathologies of strong-form PINNs, while the graph-transformer hybrid provides the geometric bias needed for complex meshes. They report residual convergence below 1% on unseen loads and, after a residual-driven test-time refinement, displacement error below 3% on a geometry not seen during training.

Core claim

The central claim is that the equilibrium residual of the weak form, evaluated with finite-element shape functions (polynomials) instead of autograd, is a sufficient training signal for a MeshGraphNet-Transformer to learn the solution operator of compressible neo-Hookean elasticity. Training minimises a loss that combines the force residual, a penalty for negative Jacobians, and a boundary-condition term; at inference the same residual serves as an error indicator, and if it exceeds a user-set tolerance the network is refined by further backpropagation. The paper reports that on a cube patch test the physics-informed loss reaches force-residual errors comparable to supervised training, that

What carries the argument

The key object is the nodal force residual R_i = sum over connected elements of internal nodal forces, computed from the predicted displacement field via FEM shape-function gradients and the neo-Hookean constitutive law. This residual is both the training loss (log(1+L_R) plus penalties) and the test-time error indicator. The architecture is MGN-T, which combines message-passing blocks for local geometry with a physics-attention transformer for long-range coupling; node and edge attributes encode only relative mesh geometry and prescribed displacements, so the representation is intended to be invariant to absolute positions and thus transferable to new meshes.

Load-bearing premise

Generalization to unseen geometries rests on the premise that the training meshes (from two fixed topologies) sample the space of possible geometries densely enough that relative graph features transfer to new meshes at test time.

What would settle it

Train on the two documented topologies using the reported hyperparameters, then evaluate zero-shot on a distinctly different geometry (e.g., a bracket with different holes or an L-shaped domain) and measure the force residual; if r_max stays above the 1% threshold and test-time refinement cannot reduce it within the reported ≈10% of training iterations, the claim of geometry generalisation is contradicted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training requires no finite-element simulations, eliminating the expensive data generation, curation, and storage step in current learned solvers.
  • Because the residual is a physics-based error indicator, the model can detect when a query lies outside its regression domain and refine itself at test time.
  • The method produces a continuous neural operator: after training, arbitrary prescribed displacements inside the training range are evaluated in a single forward pass.
  • The combination of local message passing and global attention allows the architecture to scale to meshes of thousands of nodes with complex topologies.
  • The weak-form, shape-function approach avoids the high-order autodiff instability that plagues strong-form PINNs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to treat the force residual as a calibrated uncertainty or out-of-distribution score: it could flag when a geometry or load is too far from training and warn the user to refine or regenerate.
  • The shape-function gradient trick should transfer to other constitutive models and to higher-order elements, though the paper tests only linear tetrahedra with one integration point.
  • If the geometry-manifold assumption holds, the method could be combined with online refinement in many-query industrial settings, such as inverse design or optimisation loops that require thousands of near-identical simulations.
  • Comparing zero-shot versus refined performance on a continuum of deformed geometries—beyond one hand-modified cube—would quantify how much of the generalization is due to the architecture's relational bias and how much to test-time optimisation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The manuscript presents a data-free, physics-informed neural PDE solver for 3D nonlinear hyperelasticity. A MeshGraphNet-Transformer (MGN-T) predicts nodal displacement fields from a mesh graph and prescribed boundary conditions; the training loss is built from the weak-form force residual computed with finite-element shape-function gradients, together with a boundary-consistency penalty and a negative-Jacobian penalty. The residual is also proposed as an a posteriori error indicator, enabling test-time refinement when a predicted solution does not meet a user-set residual tolerance. Experiments cover a homogeneous traction cube (156 nodes), a perforated cube and a plate (1K and 2.5K nodes), and a single hand-modified version of the perforated cube used to claim geometry generalization. The paper reports residual-based convergence below 1% after refinement and displacement RRMSE below 3% on the unseen geometry, and compares physics-based training with data-driven training on the cube.

Significance. If substantiated, the contribution is significant: it combines a weak-form, autodifferentiation-free residual computation with a graph/transformer architecture that carries geometric bias, and it proposes a concrete residual-based mechanism for test-time correction. The pseudo-code for the residual assembly is a useful strength, and the comparison of a physics-informed loss against a data-driven loss on the same architecture is informative. However, the headline claim of generalization to unseen geometries rests on thin and partly circular evidence: the residual used as the error indicator is also the training objective, and the only unseen-geometry example fails the zero-shot residual tolerance. The paper needs additional experiments and clearer reporting to separate what is learned as an operator from what is achieved by per-query residual minimization.

major comments (4)
  1. [§6.2, §6.3 and Eqs. (3)–(4)] The central claim of generalization to 'previously unseen ... geometries' is supported by exactly one hand-modified perforated cube. Fig. 10(b) shows that the zero-shot prediction on this geometry fails the rmax ≤ 0.01 tolerance, and the reported sub-3% RRMSE is obtained after test-time refinement that backpropagates the same L_phys used in training. Because rmax is computed from the same nodal residuals R_i that enter L_R in Eq. (4), a decrease in rmax after refinement is partly by construction. The relation between the 'below 5%' statement in §6.2 and the 'below 3%' statement in §6.3 is not clarified, and no zero-shot RRMSE is reported separately. Moreover, the paper does not compare refinement from the pretrained weights with refinement from a randomly initialized network on the same query. Without these numbers, the result is consistent with a per-query residual-minimization solver w
  2. [§5.2, §6.2] The geometry-manifold premise of the generalization claim is not quantified. Training uses only two fixed topologies (perforated cube, plate), and the single unseen geometry is a modification of one training geometry. There is no systematic variation of topology (hole size/position, number of holes, aspect ratio, curvature) and no measure of geometric distance between training and test meshes. Consequently the reader cannot assess how far a geometry can deviate before zero-shot prediction fails and test-time refinement can no longer recover. A geometry-transfer study with several modified geometries of increasing deviation, reporting both zero-shot and refined errors, is needed to make the claim credible.
  3. [§5.1, §5.2, Abstract] The abstract claims that the method 'scal[es] up to models of considerable size and arbitrary geometries,' but the largest mesh in the experiments has 2.5K nodes, and no scaling study of node count, training cost, or inference time is presented. The claim is therefore unsupported. Please either add scaling results (even a simple node-count sweep) or temper the abstract to the demonstrated range.
  4. [§4.3.1, Algorithm 1] Reproducibility is hampered by unspecified hyperparameters: λ1 and λ2 in Eq. (4), the threshold ε in Eq. (3), the number of message-passing steps, the physics-attention token count P, and the learnable temperature τ are not given. The sensitivity to λ1/λ2 is not analyzed. Since these weights balance the residual, boundary, and inversion terms, the reader cannot tell whether the reported results are robust or tuned for each example. Please report all hyperparameters and include an ablation for at least λ1/λ2.
minor comments (8)
  1. [Abstract] The phrase 'achieving easily convergence errors' is garbled; it should read something like 'achieving convergence errors in the residual easily below 1%' or be rephrased.
  2. [§4.3.1] The label 'autodifferentiation-free scalable method' is misleading: network training and test-time refinement still use backpropagation through the network and loss. Only the residual assembly avoids automatic differentiation. Suggest 'autodifferentiation-free residual computation'.
  3. [§4.2] Typo: 'Guble-Softmax' should be 'Gumbel-Softmax'.
  4. [§5.1] The homogeneous traction cube is called a 'patch test,' but the text does not describe the standard patch-test verification (e.g., constant-strain states and mesh-independence). If this is not a standard patch test, use a different term such as 'manufactured-solution check'.
  5. [Eq. (1)] The notation 'P:∇N' in Eq. (1) is ambiguous; specify the contraction over reference coordinates and indicate the correct expression for the nodal internal force in a total-Lagrangian setting.
  6. [Figures 6 and 8] The boxplots and curves lack numerical values and explicit captions. State what each boxplot covers (number of queries), which metric is shown, and whether error bars are over seeds.
  7. [Conclusion] The phrase 'without requiring form data' should read 'without requiring ground-truth data' or 'without requiring simulation data.'
  8. [Figure 10] The caption 'Unseen geometry: test-time and refinement' is unclear; the boxes should be explicitly labeled 'zero-shot' and 'after test-time refinement'.

Circularity Check

2 steps flagged

Residual-based convergence and test-time refinement are partly by construction; independent displacement RMSE anchors keep the central claim from being fully circular.

specific steps
  1. fitted input called prediction [Abstract; §4.3.1 Eq. (4); §6.3]
    "whenever a candidate solution yields a residual that exceeds our residual tolerance, the model undergoes back-propagation to minimise the physics loss until the convergence criteria are met (see Figure 9). Initially, the zero-shot prediction exhibits a force residual above the convergence threshold."

    The convergence criterion (Eq. 3, rmax = Rmax/Ravg ≤ 0.01) is defined on the same nodal residual vector R that is minimized as L_R = Σ||R_i||² in Eq. (4). Test-time refinement is continued backpropagation of Lphys until Eq. (3) is satisfied, so the reported post-refinement residual (<1%) is a stopping condition, not an independently predicted outcome. Because the unseen-geometry case fails zero-shot and is rescued only by this refinement, the headline 'generalises ... achieving convergence errors in the residuals of less than 1%' is at least partly an artifact of the optimizer. Independent RRMSE-vs-FEM values keep it from being wholly circular.

  2. self definitional [Abstract; §6.3]
    "To ensure compliance with the laws of physics and provide guarantees regarding the inference, it is possible to use the residual itself as an error indicator for the inference, and thus perform a refinement at the testing stage if the residual tolerance set in advance by the user is not met."

    This defines physical acceptability by rmax ≤ 0.01 and then obtains that state by minimizing precisely that residual. Thus 'physical consistency' after refinement is true by construction rather than by an independent check. The paper's non-circular evidence is the displacement RMSE against a finite-element reference (below 5% zero-shot and below 3% after refinement), which is a different observable from the minimized loss.

full rationale

The physics-loss derivation itself is self-contained: L_R is assembled from FE shape-function gradients and a neo-Hookean constitutive model (§3–§4), not imported from the authors' prior work. Held-out load cases on the training geometry show zero-shot residual improvement, and the unseen-geometry displacement error is checked against an independent FEM solution, so there is real non-circular content. The circular element is the double use of R: Eq. (3)'s convergence indicator is exactly what Eq. (4)'s L_R minimizes, and the unseen-geometry <1% residual claim is met only after refining that same residual. This makes part of the reported 'guarantee' and the headline unseen-geometry convergence a consequence of the optimizer rather than a prediction. No load-bearing self-citation chain or imported uniqueness theorem was found; MGN-T [3] is used as an architecture choice, not as a proof. Score 4 reflects partial circularity with independent displacement anchors.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central method uses only standard finite-element machinery plus the MGN-T architecture. The ledger captures the hand-set hyperparameters and tolerances (λ1, λ2, ε, rmax≤0.01) that the empirical claims depend on, and the domain assumptions about mesh adequacy, graph-transferability, and material model. No new physical entities are proposed.

free parameters (4)
  • λ1, λ2 (loss weights) = not reported
    Scalar weights in Eq. (4) balancing residual, negative-Jacobian penalty, and boundary-condition loss; hand-chosen and not specified, so experiments cannot be exactly reproduced.
  • Residual normalization threshold ε = 1e-5
    Threshold in Eq. (3) selecting 'active' nodes for computing Ravg; hand-set and directly affects whether rmax ≤ 0.01 is achieved.
  • Convergence tolerance rmax ≤ 0.01 = 0.01
    User-defined acceptance criterion in Eq. (3); the headline 'less than 1% residual error' is defined by this chosen threshold, not by an independent error measure.
  • Learnable temperature τ in physics-attention = learnable
    MGN-T's slicing temperature is optimized during training; it is an architecture-level fitted parameter inherited from the authors' prior architecture [3].
axioms (4)
  • domain assumption Minimum total potential energy is the correct equilibrium principle for quasi-static hyperelasticity.
    The weak-form residual and loss are derived from this principle (§3), which is standard but not verified for the specific loading conditions used.
  • domain assumption Linear tetrahedral elements with a single integration point are an adequate discretization for the meshes and deformation ranges used.
    The authors acknowledge this limitation in §7: only 3D linear tets with one integration point are used, a modeling choice that can affect accuracy in large-deformation or near-incompressible settings.
  • domain assumption MGN-T can represent the solution operator for the family of PDE problems considered.
    The architecture is taken from the authors' separate preprint [3]; no approximation/expressivity guarantee is given here, so the method's success depends on that architecture's empirical capacity.
  • domain assumption Graph node/edge features preserve enough geometric information for cross-topology generalization.
    The claim that unseen geometries can be solved relies on relative-distance edge features transferring to new mesh connectivity (§4.2), but the paper tests only one geometry modification.

pith-pipeline@v1.3.0-daily-deepseek · 11951 in / 13702 out tokens · 139254 ms · 2026-07-31T23:19:40.563894+00:00 · methodology

0 comments
read the original abstract

We present a physics-informed, data-free neural solver for partial differential equations, built on a graph neural network architecture that utilises message passing. By relying on the weak form of the problem, we use gradients of finite-element shape functions (which are therefore polynomials) rather than automatic differentiation operators to compute the residuals of the equation from the displacements predicted by the network itself. Our approach generalises to previously unseen load cases and geometries, achieving easily convergence errors in the residuals of less than 1% and being capable of scaling up to models of considerable size and arbitrary geometries. To ensure compliance with the laws of physics and provide guarantees regarding the inference, it is possible to use the residual itself as an error indicator for the inference, and thus perform a refinement at the testing stage if the residual tolerance set in advance by the user is not met. Examples are provided to demonstrate the performance of the proposed method. This results in a method that avoids the costly process of obtaining, curating and storing high-fidelity synthetic data for training the neural network. Whilst this is not unique to our method, it is the first time it has been combined with a geometric machine learning technique capable of providing the necessary geometric bias to overcome the well-known difficulties of physics-informed neural networks.

Figures

Figures reproduced from arXiv: 2607.27901 by David Gonzalez, Elias Cueto, Iciar Alfaro, Mikel M. Iparraguirre.

Figure 1
Figure 1. Figure 1: To ensure consistency, strain and stress tensors are directly computed from the predicted displacement using the equations described in Section 3. 4.2 Spatial Representation: MGN-T The MeshGraphnet-Transformer (MGN-T) architecture [3] follows an encoder-processor-decoder framework, com￾bining graph based inductive biases with global attention to efficiently model complex solid mechanics problems (see [PIT… view at source ↗
Figure 1
Figure 1. Figure 1: Sketch of the proposed method. Given a finite element mesh (top row, left), this is modelled as a graph [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of MGN-T architecture and physics post-processing. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Test prediction for unseen loading case for the homogeneous traction on a cube. a) problem statement, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Test predictions for unseen loading cases for perforated cube geometry. Top: Prescribed displacements. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Test predictions for unseen loading cases for the plate geometry. (a) Colour maps represent von Mises stress [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of the obtained residual norms and RRMSE displacement errors for unseen test loading cases. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of the physics-informed strategy (red) vs the data-driven (blue) strategy, monitoring evaluation [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison at inference for continuous input [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Test-time refinement example for an unseen geometry. The first two figures, top left, show the displacement [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Convergence residuals rmax for the perforated cube example. The dashed line represents convergence tolerance. a) Seen geometry: train and test queries b) Unseen geometry: test-time and refinement. 7 Limitations The current implementation has been accomplished by employing 3D tetrahedral linear elements with a single inte￾gration point. Although the formulation is completely general, the incorporation of h… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 15 linked inside Pith

  1. [1]

    Machine learning in computer aided engineering

    Francisco J Mont ´ans, El´ıas Cueto, and Klaus-J¨urgen Bathe. Machine learning in computer aided engineering. In Machine Learning in Modeling and Simulation: Methods and Applications, pages 1–83. Springer, 2023

    2023

  2. [2]

    Raissi, P

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Compu- tational Physics, 378:686–707, 2019

    2019

  3. [3]

    Meshgraphnet-transformer: Scalable mesh-based learned simulation for solid mechanics.arXiv preprint arXiv:2601.23177, 2026

    Mikel M Iparraguirre, Iciar Alfaro, David Gonzalez, and Elias Cueto. Meshgraphnet-transformer: Scalable mesh-based learned simulation for solid mechanics.arXiv preprint arXiv:2601.23177, 2026

    arXiv 2026

  4. [4]

    Learn- ing to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learn- ing to simulate complex physics with graph networks. In Hal Daume III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 8459–8468. PMLR, 13–18...

    2020

  5. [5]

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

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

    Pith/arXiv arXiv 2010

  6. [6]

    Learning nonlinear op- erators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear op- erators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, 2021

    2021

  7. [7]

    Gnot: A general neural operator transformer for operator learning

    Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. InInternational conference on machine learning, pages 12556–12569. PMLR, 2023

    2023

  8. [8]

    Latent mamba operator for partial differ- ential equations, 2025

    Karn Tiwari, Niladri Dutta, N M Anoop Krishnan, and Prathosh A P. Latent mamba operator for partial differ- ential equations, 2025. arXiv:2505.19105

    Pith/arXiv arXiv 2025

  9. [9]

    Learning the solution operator of parametric partial differential equations with physics-informed deeponets, 2021

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed deeponets, 2021. 13 M.M. Iparraguirre et al

    2021

  10. [10]

    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 Gulcehre, Francis Song, Andrew Ballard, Justin Gilmer, George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra...

    Pith/arXiv arXiv 2018

  11. [11]

    Physics-informed and graph neural networks for enhanced inverse analysis.Engineering Computations, 42(7):2427–2455, 2025

    Daniele Di Lorenzo, Victor Champaney, Chady Ghnatios, Elias Cueto, and Francisco Chinesta. Physics-informed and graph neural networks for enhanced inverse analysis.Engineering Computations, 42(7):2427–2455, 2025

    2025

  12. [12]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simula- tion with graph networks, 2021. arXiv:2010.03409

    Pith/arXiv arXiv 2021

  13. [13]

    On the under-reaching phenomenon in message passing neural pde solvers: Revisiting the cfl condition.Computer Methods in Applied Mechanics and Engineering, 449:118476, 2026

    Lucas Tes ´an, Mikel M Iparraguirre, David Gonz ´alez, Pedro Martins, and El ´ıas Cueto. On the under-reaching phenomenon in message passing neural pde solvers: Revisiting the cfl condition.Computer Methods in Applied Mechanics and Engineering, 449:118476, 2026

    2026

  14. [14]

    Recurrent u-net-based graph neural network (rugnn) for accurate deformation predictions in sheet material forming.Advanced Engineering Informatics, 69:104021, 2026

    Yingxue Zhao, Qianyi Chen, Haoran Li, Haosu Zhou, Hamid Reza Attar, Tobias Pfaff, Tailin Wu, and Nan Li. Recurrent u-net-based graph neural network (rugnn) for accurate deformation predictions in sheet material forming.Advanced Engineering Informatics, 69:104021, 2026

    2026

  15. [15]

    M4gn: Mesh-based multi-segment hierarchical graph network for dynamic simulations.arXiv preprint arXiv:2509.10659, 2025

    Bo Lei, Victor M Castillo, and Yeping Hu. M4gn: Mesh-based multi-segment hierarchical graph network for dynamic simulations.arXiv preprint arXiv:2509.10659, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Efficient learning of mesh-based physical simulation with bsms-gnn, 2026

    Yadi Cao, Menglei Chai, Minchen Li, and Chenfanfu Jiang. Efficient learning of mesh-based physical simulation with bsms-gnn, 2026. arXiv:2210.02573

    Pith/arXiv arXiv 2026

  17. [17]

    Predicting physics in mesh-reduced space with temporal attention, 2022

    Xu Han, Han Gao, Tobias Pfaff, Jian-Xun Wang, and Li-Ping Liu. Predicting physics in mesh-reduced space with temporal attention, 2022. arXiv:2201.091133

    Pith/arXiv arXiv 2022

  18. [18]

    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

    2022

  19. [19]

    Transolver: A fast transformer solver for pdes on general geometries

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries. InInternational Conference on Machine Learning, 2024

    2024

  20. [20]

    Tran- solver++: An accurate neural solver for pdes on million-scale geometries, 2025

    Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Tran- solver++: An accurate neural solver for pdes on million-scale geometries, 2025. arXiv:2502.02414

    Pith/arXiv arXiv 2025

  21. [21]

    Geotransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer, 2026

    Corey Adams, Rishikesh Ranade, Ram Cherukuri, and Sanjay Choudhry. Geotransolver: Learning physics on irregular domains using multi-scale geometry aware physics attention transformer, 2026. arXiv:2512.20399

    Pith/arXiv arXiv 2026

  22. [22]

    Ab-upt: Scaling neural cfd surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers, 2025

    Benedikt Alkin, Maurits Bleeker, Richard Kurle, Tobias Kronlachner, Reinhard Sonnleitner, Matthias Dorfer, and Johannes Brandstetter. Ab-upt: Scaling neural cfd surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers, 2025. arXiv:2502.09692

    arXiv 2025

  23. [23]

    A mechanistic analysis of transformers for dynamical systems.arXiv preprint arXiv:2512.21113, 2025

    Gregory Duth ´e, Nikolaos Evangelou, Wei Liu, Ioannis G Kevrekidis, and Eleni Chatzi. A mechanistic analysis of transformers for dynamical systems.arXiv preprint arXiv:2512.21113, 2025

    arXiv 2025

  24. [24]

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

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

    2019

  25. [25]

    Lagrangian neural networks.arXiv preprint arXiv:2003.04630, 2020

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

    Pith/arXiv arXiv 2003

  26. [26]

    Structure-preserving neural networks.Journal of Computational Physics, 426:109950, 2021

    Quercus Hern ´andez, Alberto Bad´ıas, David Gonz´alez, Francisco Chinesta, and El´ıas Cueto. Structure-preserving neural networks.Journal of Computational Physics, 426:109950, 2021

    2021

  27. [27]

    Thermodynamics of learning physical phenomena.Archives of Computa- tional Methods in Engineering, 30(8):4653–4666, 2023

    Elias Cueto and Francisco Chinesta. Thermodynamics of learning physical phenomena.Archives of Computa- tional Methods in Engineering, 30(8):4653–4666, 2023

    2023

  28. [28]

    Carlos Bermejo, Alberto Bad ´ıas, David Gonz´alez, and El´ıas Cueto. Meshgraphnets informed locally by thermo- dynamics for the simulation of flows around arbitrarily shaped objects.Advanced Modeling and Simulation in Engineering Sciences, 12(1):27, 2025

    2025

  29. [29]

    Elsayed, Abedulgader Baktheer, and Fadi Aldakheel

    Elsayed S. Elsayed, Abedulgader Baktheer, and Fadi Aldakheel. Physics-based machine learning for modeling cyclic damage evolution.Proceedings in Applied Mathematics and Mechanics, 26(1):e70071, 2026

    2026

  30. [30]

    J. S. Hesthaven, C. Pagliantini, and N. Ripamonti. Structure-preserving model order reduction of hamiltonian systems, 2021. arXiv:2109.12367

    Pith/arXiv arXiv 2021

  31. [31]

    Symplectic learning for hamiltonian neural networks.Journal of Computa- tional Physics, 494:112495, December 2023

    Marco David and Florian Mehats. Symplectic learning for hamiltonian neural networks.Journal of Computa- tional Physics, 494:112495, December 2023. 14 M.M. Iparraguirre et al

    2023

  32. [32]

    Physics-informed neural networks for mod- eling rate- and temperature-dependent plasticity, 2022

    Rajat Arora, Pratik Kakkar, Biswadip Dey, and Amit Chakraborty. Physics-informed neural networks for mod- eling rate- and temperature-dependent plasticity, 2022. arXiv:2201.08363

    Pith/arXiv arXiv 2022

  33. [33]

    Challenges in training pinns: A loss landscape perspective, 2024

    Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in training pinns: A loss landscape perspective, 2024. arXiv:2402.0186

    arXiv 2024

  34. [34]

    Haolin Li, Yuyang Miao, Zahra Sharif Khodaei, and M. H. Aliabadi. Finite-pinn: A physics-informed neural network with finite geometric encoding for solid mechanics, 2025. arXi:2412.0945

    arXiv 2025

  35. [35]

    The deep ritz method: A deep learning-based numerical algorithm for solving variational problems, 2017

    Weinan E and Bing Yu. The deep ritz method: A deep learning-based numerical algorithm for solving variational problems, 2017

    2017

  36. [36]

    A deep energy method for finite deformation hyperelasticity.European Journal of Mechanics-A/Solids, 80:103874, 2020

    Vien Minh Nguyen-Thanh, Xiaoying Zhuang, and Timon Rabczuk. A deep energy method for finite deformation hyperelasticity.European Journal of Mechanics-A/Solids, 80:103874, 2020

    2020

  37. [37]

    Hamel, Kyle Johnson, Reese Jones, and Nikolaos Bouklas

    Jan Niklas Fuhg, Craig M. Hamel, Kyle Johnson, Reese Jones, and Nikolaos Bouklas. Modular machine learning- based elastoplasticity: Generalization in the context of limited data.Computer Methods in Applied Mechanics and Engineering, 407:115930, 2023

    2023

  38. [38]

    Tobias Wuerth, Niklas Freymuth, Clemens Zimmerling, Gerhard Neumann, and Luise Kaerger. Physics-informed meshgraphnets (pi-mgns): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes.Computer Methods in Applied Mechanics and Engineering, 429:117102, 2024

    2024

  39. [39]

    Learning, solving and optimizing pdes with tensorgalerkin: an efficient high-performance galerkin assembly algorithm, 2026

    Shizheng Wen, Mingyuan Chi, Tianwei Yu, Ben Moseley, Mike Yan Michelis, Pu Ren, Hao Sun, and Siddhartha Mishra. Learning, solving and optimizing pdes with tensorgalerkin: an efficient high-performance galerkin assembly algorithm, 2026

    2026

  40. [40]

    Wino: A weak-form physics informed neural operator for hyperelasticity on variable domains, 2026

    Bokai Zhu, Qinghui Zhang, and Timon Rabczuk. Wino: A weak-form physics informed neural operator for hyperelasticity on variable domains, 2026. arXiv:2605.24651

    Pith/arXiv arXiv 2026

  41. [41]

    Elsevier, 2005

    Olgierd Cecil Zienkiewicz and Robert Leroy Taylor.The finite element method for solid and structural mechan- ics. Elsevier, 2005

    2005

  42. [42]

    John wiley & sons, 2014

    Ted Belytschko, Wing Kam Liu, Brian Moran, and Khalil Elkhodary.Nonlinear finite elements for continua and structures. John wiley & sons, 2014

    2014

  43. [43]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. arXiv:1412.6980

    Pith/arXiv arXiv 2017

  44. [44]

    Sgdr: Stochastic gradient descent with warm restarts, 2017

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts, 2017. arXiv:1608.03983. 15 M.M. Iparraguirre et al. A Pseudo-code algorithm Algorithm 1physics-informed loss via Tensor Contraction and Scatter-Add Input:node displacement ˆu, elementsE, prescribedu bc, gradients∇N, volumesV, Dirichlet maskΓ D Output:nodal force residu...

    Pith/arXiv arXiv 2017