Pith. sign in

REVIEW 2 major objections 2 minor 55 references

Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees

T0 review · 2 major / 2 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that a neural operator becomes a reusable, assemblable element only when it exports a scalar energy that is convex in its boundary degrees of freedom — and that this convexity, not field-prediction accuracy, is

desk verdict A genuinely useful construction with careful experiments, but Theorem 4 as written does not prove the geometry-generalization claim, and the code is still placeholder. read the letter →

arxiv 2608.02036 v1 pith:HGSNWOE2 submitted 2026-08-03 cs.LG

Hongyue Jiang , Jianjiang Zhan , Chenzhuo Zhang , Fan Wang This is my paper
classification cs.LG MSC 65N3068T0765N55
keywords neuraloperatorsoperatorlearningfiniteelementassemblyinput-convexnetworksstaticcondensationhypernetworkerrorguaranteesgeometryparameterization
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 argues that training a neural operator to predict fields is the wrong target when the operator is meant to serve as a reusable, finite-element-like building block: a field-accurate element (0.7–1.3% prediction error) induces an energy whose assembled Hessian is indefinite, and Newton converges to a spurious state that is 247% wrong. The repair is to make the element export a scalar energy that is convex by construction in its boundary degrees of freedom, parameterized by geometry through a hypernetwork, and regularized so the regularizer's nullspace contains the physics nullspace. The paper proves that if this learned energy is epsilon-accurate and mu-strongly convex, assembled solutions are within 8 epsilon / mu of the true solution, and singular element stiffnesses assemble into a positive-definite global system exactly as in classical finite elements. Experiments with one trained element type reach 0.6–1.0% error on 2×2 to 8×8 grids and L-shaped layouts at unseen geometries, 0.2% in a plane-strain elasticity instantiation, and 0.23% in a 3D cube assembly. A sympathetic reader would care because the result identifies the structural condition — convexity of the exported energy, not accuracy of field predictions — that makes learned surrogates composable into solvers.

What carries the argument

The central object is the convex neural energy element: E(g,U) = 1/2 ||L(g)U||^2 + PICNN(g,U) + lambda/2 ||(I-P_c)U||^2, where U holds the element's boundary degrees of freedom, g is the geometry parameter, L(g) is a hypernetwork-generated matrix making the quadratic part positive semidefinite for every g, and PICNN is an input-convex neural network for non-quadratic physics. The gauge term's projector (I-P_c) enforces the nullspace principle: the regularizer's nullspace must contain the physics nullspace, here the constant mode. Theorems 1–4 carry the argument: energy accuracy controls solution accuracy, singular element stiffnesses assemble into a strictly convex global system when connect

What would settle it

Assemble four instances of a value-trained field operator with about 1% single-tile field accuracy and evaluate the eigenvalues of the induced energy's Hessian at the true solution; if any independently seeded network yields no negative eigenvalues and Newton converges to the true solution rather than a 200%-error fixed point, the paper's structural-failure diagnosis is wrong. Alternatively, compute epsilon and mu at a held-out geometry and check whether the assembled solution error respects 8 epsilon / mu; one verified violation would falsify the transfer bound.

Watch

Extended reading notes

Core claim

The paper's central claim is that neural operators can be turned from single-use surrogates into reusable, geometry-parameterized elements only if the learned object is the element energy and that energy is convex in the boundary degrees of freedom. Value-trained field prediction is shown to fail structurally: even with 0.7–1.3% field accuracy, the induced energy's assembled Hessian has tens of negative eigenvalues and Newton converges to a spurious critical point (247% error). The proposed convex neural energy element, E(g,U) = 1/2 ||L(g)U||^2 + PICNN(g,U) + lambda/2 ||(I-P_c)U||^2, where L(g) comes from a hypernetwork and PICNN is an input-convex network, is architecturally positive semide

Load-bearing premise

All error guarantees are conditional on the trained element energy being epsilon-accurate and mu-strongly convex at the test state, and that accuracy is not proven in advance for unseen geometries — it is only measured — so if the hypernetwork's smooth interpolation across the geometry family silently degrades, the assembly's positive-definiteness guarantee can fail.

Editorial extensions

If this is right

  • A single trained element type can be instantiated at unseen geometries and assembled into grids, L-shapes, and mixed-type layouts, reaching roughly 0.4–1.0% boundary-solution error with Newton converging in two iterations.
  • Energy accuracy epsilon bounds solution error by 8 epsilon / mu, and all reported model checkpoints stay inside that bound; the observed error rates are faster (roughly epsilon^1.9 and N^0.13), so the bound is safe but loose.
  • Boundary-data robustness is structural: because the element stores a quadratic operator rather than a response surface, out-of-distribution boundary patterns such as ramps and plateaus cause no measurable degradation, unlike field-predicting elements.
  • The library mechanism is compositional: adding an element type is one training run, mixed assemblies obey the same per-element hypotheses, and the 3D instantiation transfers the recipe unchanged.
  • For nonlinear reaction-diffusion physics, the exact condensed energy assembles to machine precision, and a training-free differentiable-refinement ladder reaches 0.04% error; fully learned nonlinear corrections remain an explicitly open gap.

Reading between the lines

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

  • The paper leaves implicit that any coupling strategy that extracts forces and tangents from a learned surrogate — domain decomposition, co-simulation, or substructuring — should check the induced energy's convexity and nullspace; value-regression surrogates can carry the same hidden instability even when their predictions look accurate.
  • The nullspace-compatible regularization principle is transferable: learned reduced-order models or constitutive laws regularized with a term whose nullspace misses rigid modes would exhibit floors analogous to the measured 0.85–6.7% floors reported here for gauge2 and full Tikhonov regularizers.
  • A testable extension suggested by the paper's own scaling analysis is to replace the dense hypernetwork output with low-rank or spectral output parameterizations of L(g); if successful, the 30–175x per-geometry setup savings would survive at fine 3D interface resolutions, where the current O(n_b^2) output cost becomes prohibitive.
  • The failure diagnosis implies a practical diagnostic for any learned element: evaluate the Hessian of the induced energy at the true solution before assembly; finding negative eigenvalues predicts the 247%-style failure regardless of field accuracy.
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

2 major / 2 minor

Summary. The paper proposes convex neural energy elements: geometry-parameterized element energies of the form E(g,U) = (1/2)||L(g)U||^2 + PICNN(g,U) + (λ/2)||(I-P_c)U||^2, where L(g) is produced by a hypernetwork so the quadratic head is PSD, PICNN is input-convex in U, and the gauge term obeys a nullspace principle. The authors argue that value-trained field-predicting neural operators fail as assembly elements (Section 2), that convex energy elements repair this failure (Theorems 1–3), and that geometry generalization follows from Lipschitz dependence of the energies on g (Theorem 4). They validate the construction on 2D heat conduction with elliptic holes, heterogeneous element types, plane-strain elasticity, and a 3D cube element, with carefully measured error bounds, ablations, and cost amortization.

Significance. If the construction and claims hold, this is a valuable step toward reusable, geometry-parameterized neural elements that inherit classical FEM assembly guarantees. The paper's strengths are substantial: the linear-assembly theorems (Theorems 1–3) are elementary and appear internally correct; the experimental protocol is unusually careful, with exact-condensation controls, five-seed stability checks, held-out geometries, training-free regularization floors, and ablations that isolate the target variable; the cost and amortization statements are honest and include measured crossovers. The main weakness is that the geometry-generalization guarantee, Theorem 4, is not derived from its stated hypotheses as written, and Section 2's failure diagnosis lacks a precise definition of how a field-predicting network induces an element energy. These issues are load-bearing because the abstract highlights geometry generalization and the structural-failure diagnosis motivates the whole construction. The conditional nature of the error bounds is disclosed clearly in Section 7, and I do not fault the paper for not having sample-complexity bounds; I do fault Theorem 4 for overclaiming a derivation t

major comments (2)
  1. [Appendix A.4 / Theorem 4] As written, Theorem 4 does not follow from its stated hypotheses. Theorem 1's ε is the energy gap |E_θ−E_FEM| at the two minimizers of the geometry in question. The minimizers at g* and g_test are different states. The proof bounds |E(g_test,U)−E(g*,U)| for a fixed U using Lipschitz continuity in g, but it does not bound the gap at g_test's minimizers by the gap at g*'s minimizers unless one has a uniform-in-U energy-gap bound at g* (for example, an operator-norm error of the quadratic forms). No such uniformity is stated or verified. Section 7 explicitly lists formal generalization over geometry families as open. Consequently, Theorem 4 as written either restates a measured ε at the test geometry or requires an additional unstated hypothesis; the abstract's 'geometry generalization' guarantee is not derived.
  2. [Section 2.3] The failure diagnosis is load-bearing for the paper's motivation, but the 'induced energy' of a value-trained field-predicting NOE is never defined. The text reports 36 negative Hessian eigenvalues and 247% error without specifying how a field-predicting network induces an element energy E(U), and hence internal forces q=∂E/∂U and tangents K=∂^2E/∂U^2. Please give the exact construction (e.g., a variational integral over the predicted field, or a Schur complement of a network-induced interior energy) and confirm that this is the energy used by NOEM-style assembly. Without this definition, the claimed structural failure is not reproducible from the manuscript.
minor comments (2)
  1. [Data and code availability] The code/data URLs are placeholders ('GITHUB-REPOSITORY-URL-TO-BE-ADDED', 'ZENODO-DOI-TO-BE-ADDED'). These must be resolved before publication; currently the empirical claims cannot be independently checked.
  2. [Section 4.7 / Table 2] The polynomial ridge baseline slightly outperforms the hypernetwork (0.44% vs. 0.66%). The text interprets this as evidence for the target choice, which is reasonable, but it would strengthen the paper to state more directly that the hypernetwork is not claimed to be the optimal regressor and to explain any architectural choices in that light.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned energy is a regression target with conditional error bounds, not a prediction forced by its own fit.

full rationale

The paper's central derivation is not circular. The learned element energy E(g,U) is trained by operator regression to M_g = sqrtm(S_g - lambda(I-P_c)), where S_g is the exact condensed stiffness; therefore the 'learned energy' is a fitted surrogate of the exact energy, and the main error bound (Theorem 1) is a conditional inequality: if |E_theta - E_FEM| <= epsilon at the relevant minimizers, then ||U*_theta - U*|| <= 8 epsilon / mu. The epsilon used in the experimental checks is measured on held-out geometries and boundary conditions; it is not a parameter fitted to the reported solution errors. Theorem 2 restates the classical FEM fact that singular element stiffnesses with compatible nullspaces assemble into a positive-definite global system; Theorem 3 is an architectural convexity verification; Remark 1 is an analytic floor. There is no load-bearing self-citation and no uniqueness theorem imported from the authors' prior work. The genuine weaknesses are not circularity: Theorem 4 / Appendix A.4 transfers epsilon(g*) to epsilon(g_test) by Lipschitz dependence on g, but Theorem 1 needs the gap at the test-geometry minimizers, so as written the proof requires a uniform-in-U (operator-norm) gap that is neither stated nor verified; Section 7 explicitly leaves formal generalization bounds over geometry families open ('Formal generalization bounds over geometry families (sample complexity) remain open'). The code/data URLs are placeholders, which hampers reproducibility but does not make the derivation circular. The 'all nine points within the bound' checks are post-hoc verification of a conditional theorem, not a prediction forced by construction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method's guarantees are conditional on the trained energy being a good approximation of the exact condensed energy (epsilon small) and on the hypernetwork preserving the PSD and nullspace structure; the paper verifies these on benchmarks but provides no formal sample-complexity bound.

free parameters (3)
  • lambda (gauge regularization strength) = 0.005 (heat), 5e-4 (plane-strain), 0.002 (3D)
    Chosen per physics to satisfy the admissibility condition lambda <= lambda_2(S_g) or lambda_4(S_g) measured at the pool minimum; targets M_g = sqrtm(S_g - lambda(I-P)) depend on it. Transparent hyperparameter; the paper shows full-Tikhonov variants sit exactly on the predicted floor.
  • r (reduced-basis rank for nonlinear correction) = 18
    Selected to capture 99.9% of the variance of the remainder-gradient covariance in the nonlinear benchmark (Section 5).
  • eta (Dirichlet penalty weight) = 1e6
    Classical penalty parameter for boundary conditions in Algorithm 1; the paper notes elimination or Lagrange multipliers work equally, so it is not load-bearing.
assumptions (6)
  • domain assumption The discrete solution minimizes a convex energy functional Pi(U) = sum_e E_e(U|e) - F^T U.
    Section 2.1: the framework presupposes a self-adjoint variational problem; validated scope is linear self-adjoint physics and one convex nonlinear benchmark.
  • standard math Static condensation is exact and energy-preserving: assembling the condensed Schur complements S_g reproduces the monolithic fine solve.
    Section 2.1, Eq. (2); classical FEM/condensation result (Guyan, Craig-Bampton), verified by the exact-condensation control at 0.067% / 0.047% / 0.69% in 2D / elasticity / 3D.
  • domain assumption The hypernetwork can represent sqrtm(S_g - lambda(I-P)) accurately across the geometry family with 400 training geometries.
    Section 3.2: fit achieves matrix MSE 5.5e-6, but this is an empirical capacity and interpolation assumption; no sample-complexity bound is provided (acknowledged open in Section 7).
  • domain assumption E_theta and E_FEM are Lipschitz in the smooth geometry encoding with finite constants L_theta, L_F.
    Theorem 4 and Appendix A.4: needed for the linear-in-d geometry generalization bound; the paper measures the resulting exponent about 0.9.
  • standard math PICNN architecture with nonnegative weights and softplus is convex in U.
    Appendix A.3, Theorem 3, citing Amos et al. [3]; used only in the nonlinear correction.
  • domain assumption The regression preserves the physics nullspace: L(g)1 about 0 for heat, and L(g)r about 0 for rigid modes in elasticity, at unseen geometries.
    Theorem 2 needs each element's nullspace to be exactly the admitted physics nullspace; observed via one near-null assembled mode, but this is an empirical property of the fitted L(g), not an architectural guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees." pith.science (2026). https://pith.science/paper/HGSNWOE2

@misc{pith2026260802036,
  author       = {Pith},
  title        = {Pith review of: Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HGSNWOE2}},
  note         = {Machine review of arXiv:2608.02036}
}
read the original abstract

Extending the neural-operator element method from individually trained, fixed-geometry neural elements to a library of reusable, geometry-parameterized element types fails structurally: a field-predicting operator trained by value regression induces an energy whose assembled Hessian is indefinite, and Newton converges to spurious minima (247% error) even with 1%-accurate field predictions. We introduce convex neural energy elements: each element exports a scalar energy E(g,U), architecturally convex in its boundary degrees of freedom U and smoothly parameterized by its geometry g, realized as a hypernetwork-generated positive-semidefinite quadratic form (an input-convex correction is reserved for non-quadratic physics). A regularization-nullspace principle--the regularizer's nullspace must contain the physics nullspace--removes an otherwise irreducible bias, and assembled elements inherit the classical guarantee that singular element stiffnesses yield a positive-definite global system. We prove conditional error bounds (energy-to-solution accuracy, element-count scaling, geometry generalization) and verify each experimentally. On heat conduction with elliptic holes, one trained element assembles into 2x2 to 8x8 grids and an L-shaped layout of unseen geometries at 0.6-1.0% relative L2 error, with 175x faster per-geometry setup for boundary-quantity workloads. A second trained element type mixes freely with the first in one monolithic assembly, and a three-dimensional instantiation reaches 0.23% on eight-element assemblies--the guarantees are type- and dimension-agnostic. A plane-strain elasticity element, whose physics nullspace is three-dimensional, lands on the analytically predicted regularization floors. Making the energy the learned object turns neural operators from single-use surrogates into reusable elements that inherit the assembly guarantees of the method they extend.

Figures

Figures reproduced from arXiv: 2608.02036 by the authors.

Figure 1
Figure 1. Convex neural energy elements at a glance. (a) Each element carries a scalar energy [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Four convex neural energy elements with unseen geometries assembled into a 1.6×1.6 domain. (a) Reference solution; (b) assembled solution (0.31% relative L2 , 2 Newton iterations), with each element’s held-out geometry (a, b, θ) annotated; (c) pointwise error TNIEL − Tref (absolute, same units as T; the field spans [−1.6, 0.8]), concentrated at element interfaces where the learned energies meet. White lines mark ele… view at source ↗
Figure 3
Figure 3. Theory validated point-by-point. (a) Every model-quality checkpoint satisfies the Theorem 1 bound (gray line; shaded region is [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Out-of-distribution behavior. (a) Geometry extrapolation beyond the training box degrades gracefully (worst case [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Non-rectangular topology. An L-shaped domain (a [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Heterogeneous element types. An L-shaped 12-element assembly mixing two trained element types (labels in panel b): type A, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Elasticity fields. A four-element plane-strain assembly of unseen geometries with the gauge [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Plane-strain elasticity: the nullspace principle with a three-dimensional physics nullspace (rigid modes). (a) Assembly error of [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Three-dimensional instantiation. A 2 × 2 × 2 assembly of eight cube elements with unseen ellipsoidal-void geometries (a, b, c) ∈ [0.15, 0.3]3 , solved with the trained 3D element (λ = 0.002). Cutaway view on the three internal shared tile-face planes x = 0.8, y = 0.8 a…
Figure 10
Figure 10. Figure 10: Measured amortization. Cumulative wall-clock of a design [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 5 linked inside Pith

  1. [1]

    Ouyang, Y

    W. Ouyang, Y. Shin, S.-W. Liu, and L. Lu. NOEM: efficient and scalable finite element method enabled by reusable neural operators.Nature Computational Science, 6(4):417–429, 2026

    2026

  2. [2]

    M. Yin, N. Charon, R. Brody, L. Lu, N. Trayanova, and M. Maggioni. DIMON: Learning solution oper- ators of partial differential equations on a diffeomor- phic family of domains.Nature Computational Sci- ence, 4:928–940, 2024

    2024

  3. [3]

    B. Amos, L. Xu, and J. Z. Kolter. Input convex neural networks. InICML, 2017

    2017

  4. [4]

    Amsallem and C

    D. Amsallem and C. Farhat. An online method for interpolating linear parametric reduced-order models. SIAM Journal on Scientific Computing, 33(5):2169– 2198, 2011

    2011

  5. [5]

    W. E and B. Yu. The deep Ritz method: a deep learning-based numerical algorithm for solving vari- ational problems.Communications in Mathematics and Statistics, 6:1–12, 2018

    2018

  6. [6]

    Dondl, J

    P. Dondl, J. Müller, and M. Zeinhofer. Uniform con- vergenceguaranteesforthedeep Ritzmethodfornon- linear problems.Advances in Continuous and Dis- crete Models, 2022

    2022

  7. [7]

    M. S. Eshaghi, C. Anitescu, M. Thombre, Y. Wang, X. Zhuang, and T. Rabczuk. Variational physics- informed neural operator (VINO) for solving partial differential equations.Computer Methods in Applied Mechanics and Engineering, 437:117785, 2025. 16

    2025

  8. [8]

    W. Wang, A. Gupta, H. Ruan, and S. Goswami. A non-overlapping Schwarz hybrid finite element–neural operator framework for solid mechanics on irregular domains.arXiv:2606.08796, 2026

    arXiv 2026

Show all 55 references
  1. [9]

    S. W. Chung, Y. Choi, C. Miller, H. K. Springer, and K. T. Sullivan. Latent space element method. arXiv:2601.01741, 2026

    2026

  2. [10]

    Song and G

    Y. Song and G. H. Yoon. Efficient acoustic finite el- ement simulation and optimization through inverse matrix prediction by neural network.Structural and Multidisciplinary Optimization, 69:29, 2026

    2026

  3. [11]

    R. J. Guyan. Reduction of stiffness and mass matri- ces.AIAA Journal, 3(2):380, 1965

    1965

  4. [12]

    R. R. Craig and M. C. C. Bampton. Coupling of substructures for dynamic analyses.AIAA Journal, 6(7):1313–1319, 1968

    1968

  5. [13]

    Škardová, A

    K. Škardová, A. Daby-Seesaram, and M. Genet. Fi- nite element neural network interpolation: inter- pretableandadaptivediscretizationforsolvingPDEs. arXiv:2412.05719, 2024

    2024 arXiv

  6. [14]

    CompNO:afoun- dation model approach for solving partial differential equations.arXiv:2601.07384, 2026

    H.Hmida, H.C.Joly, andY.Mesri. CompNO:afoun- dation model approach for solving partial differential equations.arXiv:2601.07384, 2026

    2026 arXiv

  7. [15]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karni- adakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of op- erators.Nature Machine Intelligence, 3(3):218–229, 2021

    2021

  8. [16]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. Fourier neural operator for parametric partial dif- ferential equations. InInternational Conference on Learning Representations, 2021

    2021

  9. [17]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar. Neural operator: learning maps between function spaces with applications to PDEs.Journal of Ma- chine Learning Research, 24(89):1–97, 2023

    2023

  10. [18]

    Lanthaler, S

    S. Lanthaler, S. Mishra, and G. E. Karniadakis. Error estimates for DeepONets: a deep learning framework in infinite dimensions.Transactions of Mathematics and Its Applications, 6(1):tnac001, 2022

    2022

  11. [19]

    L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, and G. E. Karniadakis. A comprehensive and fair comparison of two neural operators (with practical extensions) based on FAIR data.Com- puter Methods in Applied Mechanics and Engineering, 393:114778, 2022

    2022

  12. [20]

    Z. Li, D. Z. Huang, B. Liu, and A. Anandkumar. Fourier neural operator with learned deformations for PDEs on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    2023

  13. [21]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, and P. W. Battaglia. Learning mesh-based simulation with graph networks. InInternational Conference on Learning Representations, 2021

    2021

  14. [22]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: a deep learn- ing framework for solving forward and inverse prob- lems involving nonlinear partial differential equa- tions.Journal of Computational Physics, 378:686– 707, 2019

    2019

  15. [23]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang. Physics- informed machine learning.Nature Reviews Physics, 3:422–440, 2021

    2021

  16. [24]

    Samaniego, C

    E. Samaniego, C. Anitescu, S. Goswami, V. M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang, and T. Rabczuk. An energy approach to the solution of partial differential equations in computational me- chanics via machine learning: concepts, implementa- tion and applications.Computer...

    2020

  17. [25]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, and G. E. Karniadakis. hp-VPINNs: variational physics-informed neural networks with domain decomposition.Computer Methods in Applied Mechanics and Engineering, 374:113547, 2021

    2021

  18. [26]

    Learningtheso- lution operator of parametric partial differential equa- tions with physics-informed DeepONets.Science Ad- vances, 7(40):eabi8605, 2021

    S.Wang, H.Wang, andP.Perdikaris. Learningtheso- lution operator of parametric partial differential equa- tions with physics-informed DeepONets.Science Ad- vances, 7(40):eabi8605, 2021

    2021

  19. [27]

    Goswami, M

    S. Goswami, M. Yin, Y. Yu, and G. E. Karniadakis. A physics-informed variational DeepONet for pre- dicting crack path in quasi-brittle materials.Com- puter Methods in Applied Mechanics and Engineering, 391:114587, 2022

    2022

  20. [28]

    M. Yin, E. Zhang, Y. Yu, and G. E. Karniadakis. In- terfacing finite elements with deep neural operators for fast multiscale modeling of mechanics problems. Computer Methods in Applied Mechanics and Engi- neering, 402:115027, 2022

    2022

  21. [29]

    Time-marching neural operator–FE coupling: AI- accelerated physics modeling.Computer Methods in Applied Mechanics and Engineering, 446:118319, 2025

    W.Wang, M.Hakimzadeh, H.Ruan, andS.Goswami. Time-marching neural operator–FE coupling: AI- accelerated physics modeling.Computer Methods in Applied Mechanics and Engineering, 446:118319, 2025. 17

    2025

  22. [30]

    Capuano and J

    G. Capuano and J. J. Rimoli. Smart finite elements: a novel machine learning application.Computer Meth- ods in Applied Mechanics and Engineering, 345:363– 381, 2019

    2019

  23. [31]

    Koeppe, F

    A. Koeppe, F. Bamer, and B. Markert. An intelligent nonlinear meta element for elastoplastic continua: deep learning using a new time-distributed residual U-Net architecture.Computer Methods in Applied Mechanics and Engineering, 366:113088, 2020

    2020

  24. [32]

    Oishi and G

    A. Oishi and G. Yagawa. Computational mechanics enhanced by deep learning.Computer Methods in Ap- plied Mechanics and Engineering, 327:462–480, 2017

    2017

  25. [33]

    S. Saha, Z. Gan, L. Cheng, J. Gao, O. L. Kafka, X. Xie, H. Li, M. Tajdari, H. A. Kim, and W. K. Liu. Hierarchical deep learning neural network (HiDeNN): an artificial intelligence (AI) framework for compu- tational science and engineering.Computer Methods in Applied Mechanics ...

    2021

  26. [34]

    Parish, P

    E. Parish, P. Lindsay, T. Shelton, and J. Mersch. Embedded symmetric positive semi-definite machine- learned elements for reduced-order modeling in finite- element simulations with application to threaded fas- teners.Computational Mechanics, 74(6), 2024

    2024

  27. [35]

    Benner, S

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

    2015

  28. [36]

    J. S. Hesthaven and S. Ubbiali. Non-intrusive reduced order modeling of nonlinear problems using neural networks.Journal of Computational Physics, 363:55– 78, 2018

    2018

  29. [37]

    Lee and K

    K. Lee and K. T. Carlberg. Model reduction of dy- namical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computa- tional Physics, 404:108973, 2020

    2020

  30. [38]

    F.FeyelandJ.-L.Chaboche. FE 2 multiscaleapproach for modelling the elastoviscoplastic behaviour of long fibre SiC/Ti composite materials.Computer Methods in Applied Mechanics and Engineering, 183(3–4):309– 330, 2000

    2000

  31. [39]

    Ghaboussi, J

    J. Ghaboussi, J. H. Garrett, and X. Wu. Knowledge- based modeling of material behavior with neu- ral networks.Journal of Engineering Mechanics, 117(1):132–153, 1991

    1991

  32. [40]

    B. A. Le, J. Yvonnet, and Q.-C. He. Computational homogenization of nonlinear elastic materials using neural networks.International Journal for Numerical Methods in Engineering, 104(12):1061–1084, 2015

    2015

  33. [41]

    Kirchdoerfer and M

    T. Kirchdoerfer and M. Ortiz. Data-driven computa- tional mechanics.Computer Methods in Applied Me- chanics and Engineering, 304:81–101, 2016

    2016

  34. [42]

    A. D. Jagtap and G. E. Karniadakis. Extended physics-informed neural networks (XPINNs): a gen- eralized space-time domain decomposition based deep learning framework for nonlinear partial differen- tial equations.Communications in Computational Physics, 28(5):2002–2041, 2020

    2002

  35. [43]

    Moseley, A

    B. Moseley, A. Markham, and T. Nissen-Meyer. Finite basis physics-informed neural networks (FBPINNs): a scalable domain decomposition ap- proach for solving differential equations.Advances in Computational Mathematics, 49:62, 2023

    2023

  36. [44]

    K. Li, K. Tang, T. Wu, and Q. Liao. D3M: a deep domain decomposition method for partial differential equations.IEEE Access, 8:5283–5294, 2020

    2020

  37. [45]

    Heinlein, A

    A. Heinlein, A. Klawonn, M. Lanser, and J. We- ber. Combining machine learning and domain de- composition methods for the solution of partial dif- ferential equations—a review.GAMM-Mitteilungen, 44(1):e202100001, 2021

    2021

  38. [46]

    D. K. Klein, M. Fernández, R. J. Martin, P. Neff, and O. Weeger. Polyconvex anisotropic hyperelastic- ity with neural networks.Journal of the Mechanics and Physics of Solids, 159:104703, 2022

    2022

  39. [47]

    As’ad, P

    F. As’ad, P. Avery, and C. Farhat. A mechanics- informed artificial neural network approach in data- driven constitutive modeling.International Journal for Numerical Methods in Engineering, 123(12):2738– 2759, 2022

    2022

  40. [48]

    T. J. R. Hughes.The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Dover Publications, Mineola, NY, 2000

    2000

  41. [49]

    O. C. Zienkiewicz, R. L. Taylor, and J. Z. Zhu.The Finite Element Method: Its Basis and Fundamentals. Butterworth-Heinemann, Oxford, 7th edition, 2013

    2013

  42. [50]

    Ouyang, S.-W

    W. Ouyang, S.-W. Liu, and S.-L. Chan. Second-order analysis method for pile-supported structures using neural-operator element method.Engineering Struc- tures, 2026. doi:10.1016/j.engstruct.2026.122385

    2026

  43. [51]

    Secchi, D

    P. Secchi, D. S. Balint, and M. Maurizi. Neural- Schwarz tiling for geometry-universal PDE solving at scale. arXiv:2605.12343, 2026

    2026 arXiv

  44. [52]

    C. Wu, C. Liu, Y. Guo, and X. Guo. DFENN: a penalty-free variational framework coupling finite el- ements and neural networks via interface condensa- tion.Journal of the Mechanics and Physics of Solids,

  45. [53]

    Pundir, F

    M. Pundir, F. Lorez, and D. S. Kammer. A versa- tile FEM framework with native GPU scalability via globally-applied AD (tatva). arXiv:2602.12365, 2026

    2026

  46. [54]

    B. D. Shaffer, S. Koohy, B. Kinch, M. A. Hsieh, and N. Trask. Structure-preserving learning improves ge- ometry generalization in neural PDEs (Geo-NeW). arXiv:2602.02788, 2026. 19

    2026 arXiv

  47. [2026]

    doi:10.1016/j.jmps.2026.106703. 18

    2026

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.