Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Solving engineering eigenvalue problems with neural networks using the Rayleigh quotient

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By minimizing the Rayleigh quotient in Gram-Schmidt-orthogonalized subspaces, a neural network can recover ordered eigenpairs of continuous engineering eigenvalue problems, with demonstrated accuracy on linear, parametric, nonlinear, and…

desk verdict A clean, honest methods paper pairing the Rayleigh quotient with Gram-Schmidt for neural eigenvalue problems; the core works on the tests, but the robustness claim is conditional and the p-Laplace comparison overreaches. read the letter →

arxiv 2506.04375 v1 pith:MCVGPGUT submitted 2025-06-04 math.NA cs.LGcs.NA

Conor Rowan , John Evans , Kurt Maute , Alireza Doostan This is my paper
classification math.NAcs.LGcs.NA MSC 65N2565N3568T07
keywords eigenvalueproblemsRayleighquotientGram-SchmidtorthogonalizationneuralnetworkdiscretizationsharmonicbasisparametricMonteCarlointegrationphysics-informedmachinelearning
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

Engineering eigenvalue problems, such as natural frequencies, buckling loads, and thermal decay rates, are usually solved by converting a continuous differential equation into a matrix problem, which neural-network discretizations cannot do directly because the network depends nonlinearly on its parameters. The paper shows that minimizing the Rayleigh quotient in a space orthogonalized by Gram-Schmidt is a simple alternative: each successive minimization returns one eigenvalue, namely the quotient's minimum value, together with its eigenfunction, in ascending order, without penalty terms and without prior knowledge of the spectrum. The claim is supported on one-dimensional and irregular-domain Laplace problems, a parameterized multi-material elasticity problem, a nonlinear p-Laplace problem, and Laplace eigenproblems up to dimension ten with Monte Carlo integration. A sympathetic reader would care because this gives physics-informed machine learning a workable route to spectral analysis, and because the learned eigenfunctions can double as a spectral basis for approximating PDE solutions.

What carries the argument

The load-bearing mechanism is the Rayleigh quotient min-characterization together with exact orthogonalization. The first eigenvalue is $\lambda_1=\min_u R(u)$, and the $i$-th eigenvalue is the minimum of the same quotient over the subspace of functions orthogonal to all previously found eigenfunctions. Gram-Schmidt orthogonalization turns that subspace constraint into a projection of the neural-network output, so no penalty term is needed to keep eigenfunctions distinct. Since the converged value of the quotient equals the eigenvalue, minimizing successive quotients in orthogonalized subspaces yields ordered eigenpairs while avoiding the eigenvalue as an additional unknown and lowering the order of differentiation required from the network.

What would settle it

Run the proposed scheme on the one-dimensional Fourier problem beyond $N=14$ and inspect whether the sixteenth converged quotient is $16^2\pi^2$; a spurious converged value would confirm that the method's validity hinges on the optimizer reaching the global minimum.

Watch

Extended reading notes

Core claim

The central claim is that a neural network discretization of the eigenfunction, paired with the Rayleigh quotient as the objective and Gram-Schmidt orthogonalization as the constraint enforcer, recovers ordered eigenpairs of continuous eigenvalue problems. For self-adjoint operators, the $i$-th eigenvalue is the minimum of $R(u)=\frac{\int_\Omega \nabla u\cdot\nabla u\,d\Omega}{\int_\Omega u^2\,d\Omega}$ over functions orthogonal to the previous $i-1$ eigenfunctions; the paper builds that orthogonality into the network output by projecting out the already-converged eigenfunctions rather than penalizing it. Because the objective's minimum value is the eigenvalue, the minimizer is the eigenfunction, and successive minima are forced to ascend. The paper demonstrates the combination on harmonic bases on the unit interval and a semicircle, on parameterized material and geometry problems, on the nonlinear $p$-Laplace eigenvalue problem, and on $d$-dimensional Laplace eigenproblems using Monte Carlo integration for both the quotient and the Gram-Schmidt inner products. It also shows that the learned harmonic functions work as a spectral basis for a parameterized steady-heat-conduction problem.

Load-bearing premise

The optimizer must find a global minimum of each successive nonconvex Rayleigh quotient in the Gram-Schmidt-orthogonalized subspace; the paper explicitly says there is no guarantee of this and relies on numerical examples to show that the proper ordering is obtained.

Editorial extensions

If this is right

  • Successive eigenvalues come out in ascending order automatically, so no shift parameter or spectrum prior is needed to reach higher modes, unlike power-method approaches.
  • A single network can parameterize eigenfunctions over stochastic material or geometry parameters, and after training the eigenvalue is available at negligible cost, enabling statistics such as mean and variance to be computed cheaply.
  • On the $p$-Laplace problem, the neural-network discretization obtains a lower converged Rayleigh quotient than a 100-term Fourier basis despite having fewer parameters, because no linear system needs to be assembled.
  • Monte Carlo evaluation of the Rayleigh quotient and the Gram-Schmidt projections keeps eigenanalysis tractable in high dimensions, with relative eigenvalue errors of about 2.4% at $d=9$ and about 5.1% for the second eigenvalue at $d=10$.
  • Learned harmonic eigenfunctions form a usable spectral basis: with 15 basis functions on the semicircular domain, the average relative Galerkin error for a parameterized Poisson-type problem is $7\times10^{-3}$.

Reading between the lines

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

  • The paper leaves open whether accumulated Gram-Schmidt error, rather than spectral bias alone, is the limiting factor at $N=14$; re-running with re-orthogonalization or a different orthonormalization scheme would isolate the cause.
  • A natural extension is to use the same Monte Carlo Gram-Schmidt procedure for the third and higher eigenfunctions in high dimensions, where exact integration is unavailable; the paper only demonstrates the second.
  • Because the method needs no mesh and no matrix assembly, it is a plausible building block for eigenanalysis inside larger optimization or operator-learning loops where the geometry changes repeatedly.
  • The parameterized-geometry basis suggests that eigenfunctions for unseen geometries could be obtained by interpolation, a direction the authors themselves flag for future work.
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

4 major / 6 minor

Summary. The paper proposes a method for solving continuous eigenvalue problems by discretizing the eigenfunctions with neural networks and minimizing the Rayleigh quotient, with successive eigenpairs obtained through Gram-Schmidt orthogonalization against previously converged approximations. The approach is demonstrated on a one-dimensional Fourier basis, a semicircular domain, a parameterized multi-material elasticity problem, a parameterized donut geometry, a nonlinear p-Laplace eigenproblem, and high-dimensional Laplace eigenvalue problems. Where possible, the results are compared against analytical solutions or finite element references, and the learned harmonic functions are used as a spectral basis for a model heat-conduction problem.

Significance. If its claims are fully supported, the paper offers a conceptually simple alternative to penalty-based and strong-form approaches for neural-network eigenvalue solvers, and it demonstrates useful capabilities for parametric and high-dimensional problems. The numerical validations against analytical and FEM references for the linear examples are a clear strength, as is the authors' transparent acknowledgment of the lack of a global-minimization guarantee. However, the central 'robustness' claim is not fully established: the p-Laplace comparison lacks a ground-truth reference, the 1D failure at N=14 is not diagnosed quantitatively, and the Monte Carlo Rayleigh quotient estimator introduces an unquantified bias in the high-dimensional section.

major comments (4)
  1. [Section 6, Eq. (19) and Figure 12] The claim that the neural network discretization 'outcompetes' the Fourier basis is supported only by the training histories of the Rayleigh quotient, with no reference eigenvalue or eigenfunction error metric. Because both methods minimize the same objective, a lower converged value is an improved upper bound but does not by itself establish higher accuracy; the paper should provide a reference solution (for example, a converged high-resolution discretization) and report errors for both the eigenvalue and eigenfunction, or temper the claim to 'achieves a lower Rayleigh quotient with fewer parameters.'
  2. [Section 4.1 and Section 8] The method is reported to break down beyond N=14 in the 1D Fourier test, with the cause attributed jointly to spectral bias and accumulated Gram-Schmidt errors, but no evidence separates these effects. Since the central claim is robustness and the Gram-Schmidt procedure is the central methodological alternative to penalty-based orthogonalization, the paper should include diagnostics such as the orthogonality residual of the stored basis, a comparison to re-orthogonalized or double-precision Gram-Schmidt, and/or a study of the optimization trajectories at the failure threshold, to delineate the reliable operating range of the method.
  3. [Section 7, Eq. (21)] The Monte Carlo Rayleigh quotient is a ratio of two random estimates, and the expectation of this ratio is not equal to the ratio of the underlying integrals. The paper does not quantify the resulting bias or its effect on the reported eigenvalue errors and on the noisy gradients used by ADAM. A bias analysis, variance reduction, or a small numerical study comparing the MC estimator against deterministic integration in low dimensions would be needed to support the high-dimensional accuracy claims.
  4. [Section 3, Eq. (8)] The Gram-Schmidt projection is performed against previously converged approximate eigenfunctions, so the search space at step i is not the exact orthogonal complement of the exact eigenspace. The paper acknowledges the lack of global-minimization guarantee but not the additional error from approximate projections; the observed eigenvalue ordering is therefore empirical rather than structural. A quantitative assessment (for example, the L2 orthogonality defect of the stored basis, or a comparison of the Rayleigh quotient against FEM eigenvalues for increasing N) would convert this observation into a validated property.
minor comments (6)
  1. [Section 4.1, Eq. (9)] The eigenfunction error metric E_u does not account for the arbitrary sign of each eigenfunction; if the learned function has a flipped sign, the integral is not small. Define the metric with a sign-corrected term or state that the sign is aligned during evaluation.
  2. [Section 4.2] The statement that better convergence is observed without the normalization penalty is a useful heuristic but not quantified; reporting the values of beta tested and the resulting differences would make the recommendation reproducible.
  3. [Section 7, preceding Figure 14] The text says 'the second eigenvalue is given by λ1 = (d-1)π² + 4π²'; this should read λ2, not λ1.
  4. [Section 7 and Figure 3] The phrase 'there would 109 integration points' is missing a word, and the label 'eigenf(nction error' in Figure 3 contains a typo; both should be corrected.
  5. [Section 3] The notation R(u(x)) for a generic, possibly vector-valued Rayleigh quotient is not formally defined; a one-line definition would avoid ambiguity in the subsequent examples.
  6. [Reference [50]] Reference [50] is listed only as a preprint with a year; if it has appeared in a journal or has a DOI, that information should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Rayleigh-quotient objective is a variational characterization of eigenvalues, and the results are externally benchmarked against analytical or FEM solutions.

full rationale

The paper's load-bearing step is the variational identity lambda_i = min R(u) over functions orthogonal to all previous eigenfunctions, used as an optimization objective for a neural-network-discretized candidate. This is a standard mathematical characterization of eigenvalues, not a parameter fitted to the output; the converged Rayleigh quotient is the eigenvalue by definition, but the reported accuracy is judged against independent references: closed-form sin(i*pi*x) eigenpairs in the 1D Fourier test, FEM eigenpairs for the semicircular and parametric problems, and exact high-dimensional Laplace eigenvalues with known analytic forms. These external benchmarks are not used to fit any parameter, so the computed eigenvalues are not predictions forced by construction. The Gram-Schmidt orthogonalization is applied to previously converged approximate eigenfunctions rather than exact ones, and the paper explicitly disclaims any guarantee of global minimization of the nonconvex quotients; both are honest limitations that affect robustness but do not make the derivation circular. The p-Laplace comparison uses the Rayleigh quotient value as a merit metric, but since the true first eigenvalue is the global minimum of that quotient, comparing final quotient values across discretizations is a legitimate optimization benchmark rather than a fitted-input prediction. No load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in via citation was found; the cited Rayleigh quotient theory, Deep Ritz method, spectral bias literature, and FEM references are standard external results. Overall, the paper is self-contained against external references and its central numerical claims are grounded in those references rather than in its own training objective alone.

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

The central claim rests on classical variational principles, the expressivity and trainability of the chosen neural networks, the reliability of the optimizer in finding global minima, and the accuracy of quadrature. No physical entities are invented; the hand-chosen algorithmic quantities are listed as free parameters.

free parameters (4)
  • Convergence threshold T = 300 (1D Fourier); 25 (semicircle); 200 (donut)
    Stopping criterion for each Rayleigh quotient optimization; chosen per example and part of the method's tuning.
  • Normalization penalty weight beta = active in 1D; set to 0 for 2D and later examples
    Hand-chosen; when nonzero it can cause stagnation, so it was dropped in favor of post-hoc normalization.
  • Monte Carlo batch size and learning rate schedule = B=1000+2000(d-1) for d=1..9; B=5000 for d=10; LR=2e-3 decaying by 0.95 per 1000 epochs
    Chosen per dimension; affects accuracy and convergence in high-dimensional examples.
  • Network architecture per problem = 1 hidden layer width 20 (1D); 2 hidden layers width 10 (semicircle); 2 hidden layers width 7 (p-Laplace); 2 hidden…
    Selected per problem; the paper notes wide and deep networks are unnecessary for low-frequency modes.
assumptions (5)
  • standard math For self-adjoint operators, the Rayleigh quotient is minimized by the first eigenfunction and successive eigenvalues are minima in subspaces orthogonal to previous eigenfunctions (Courant-Fischer).
    Used in Section 3, Eq. (7).
  • ad hoc to paper The ADAM optimizer converges to the global minimum of each nonconvex Rayleigh quotient loss in the orthogonalized search space.
    Explicitly disclaimed in Section 3 ('We have no guarantee that the optimizer will find such global minima'); the method relies on this empirically.
  • domain assumption The neural network discretization is expressive enough to represent the target eigenfunctions at the required accuracy, and its spectral bias does not prevent convergence for the frequencies targeted.
    Invoked throughout; Section 8 notes failure for n>14 in 1D attributed to spectral bias and Gram-Schmidt error accumulation.
  • domain assumption Monte Carlo integration with batch size B approximates the Rayleigh quotient and its gradient accurately enough for the high-dimensional optimizations.
    Section 7 uses random sampling per step; errors grow with dimension and reach 2.4% for d=9 and 5.1% for the second eigenvalue in d=10.
  • domain assumption FEM results and analytical solutions used as references are sufficiently accurate.
    Used in Sections 4 and 5 for comparison; mesh convergence is verified only for some cases.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solving engineering eigenvalue problems with neural networks using the Rayleigh quotient." pith.science (2026). https://pith.science/paper/MCVGPGUT

@misc{pith2026250604375,
  author       = {Pith},
  title        = {Pith review of: Solving engineering eigenvalue problems with neural networks using the Rayleigh quotient},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MCVGPGUT}},
  note         = {Machine review of arXiv:2506.04375}
}
read the original abstract

From characterizing the speed of a thermal system's response to computing natural modes of vibration, eigenvalue analysis is ubiquitous in engineering. In spite of this, eigenvalue problems have received relatively little treatment compared to standard forward and inverse problems in the physics-informed machine learning literature. In particular, neural network discretizations of solutions to eigenvalue problems have seen only a handful of studies. Owing to their nonlinearity, neural network discretizations prevent the conversion of the continuous eigenvalue differential equation into a standard discrete eigenvalue problem. In this setting, eigenvalue analysis requires more specialized techniques. Using a neural network discretization of the eigenfunction, we show that a variational form of the eigenvalue problem called the "Rayleigh quotient" in tandem with a Gram-Schmidt orthogonalization procedure is a particularly simple and robust approach to find the eigenvalues and their corresponding eigenfunctions. This method is shown to be useful for finding sets of harmonic functions on irregular domains, parametric and nonlinear eigenproblems, and high-dimensional eigenanalysis. We also discuss the utility of harmonic functions as a spectral basis for approximating solutions to partial differential equations. Through various examples from engineering mechanics, the combination of the Rayleigh quotient objective, Gram-Schmidt procedure, and the neural network discretization of the eigenfunction is shown to offer unique advantages for handling continuous eigenvalue problems.

Figures

Figures reproduced from arXiv: 2506.04375 by the authors.

Figure 1
Figure 1. Even when the eigenvalue differential equation is [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Learning the orthonormal 1D Fourier basis with hom [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The semicircular domain on which we find the harmoni [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: The first 9 harmonic basis functions on the semicirc [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Parameterized Young’s modulus for the multimater [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Verifying the implementation of the expected Rayl [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: The expected Rayleigh quotient is minimized to find [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Donut-shaped domain with given inner and outer rad [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: The first 9 harmonic functions on the donut with homo [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: The first 9 harmonic functions on the donut with hom [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: The eigenvalues for each of the 9 harmonic functio [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Comparing the performance of a two hidden layer ML [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Minimizing Rayleigh quotient with Monte Carlo in [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: Minimizing the Rayleigh quotient with Monte Carl [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Learning Gradient Flow: Using Equation Discovery to Accelerate Engineering Optimization

    math.OC 2026-02 conditional novelty 6.0 of 10

    An optimizer that fits a SINDy polynomial model to recent optimization-variable trajectories and then integrates that surrogate flow instead of evaluating the true objective/gradient can cut gradient-evaluation counts...

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [5]

    Deep Learning Solution of the Eigenvalue Problem for Differential Operators

    Ido Ben-Shaul, Leah Bar, Dalia Fishelov, and Nir Sochen. Deep Learning Solution of the Eigenvalue Problem for Differential Operators. Neural Computation, 35(6):1100–1134, May 2023

    work page 2023

  2. [49]

    Neural networks based on power method and inverse power method for solving linear eigenvalue prob lems

    Qihong Y ang, Y angtao Deng, Y u Y ang, Qiaolin He, and Shiquan Zhang. Neural networks based on power method and inverse power method for solving linear eigenvalue prob lems. Computers & Mathematics with Applications , 147:14–24, October 2023

    work page 2023

  3. [1]

    Abueidda, Seid Koric, Rashid Abu Al-Rub, Corey M

    Diab W . Abueidda, Seid Koric, Rashid Abu Al-Rub, Corey M. Parrott, Kai A. James, and Nahil A. Sobh. A deep learning energy method for hyperelasticity and viscoe lasticity. European Journal of Mechanics - A/Solids , 95:104639, September 2022. arXiv:2201.08690 [cs]

    arXiv 2022

  4. [2]

    Neural Operators for Accelerating Scientific Simula tions and Design, January 2024

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Mi guel Liu-Schiaffini, Jean Kossaifi, and Anima Anand- kumar. Neural Operators for Accelerating Scientific Simula tions and Design, January 2024. arXiv:2309.15325 [cs]

    arXiv 2024

  5. [3]

    Klaus-Jürgen Bathe and Edward L. Wilson. Solution metho ds for eigenvalue problems in structural mechanics. International Journal for Numerical Methods in Engineerin g, 6(2):213–226, January 1973

    work page 1973

  6. [4]

    O. A. Bauchau and J. I. Craig. Structural Analysis: With Applications to Aerospace Struc tures. Springer Science & Business Media, August 2009. Google-Books-ID: GYRX8ZYVN YQC

    work page 2009

  7. [6]

    Stress field prediction in fiber-reinforced composite materials using a deep learning approach

    Anindya Bhaduri, Ashwini Gupta, and Lori Graham-Brady. Stress field prediction in fiber-reinforced composite materials using a deep learning approach. Composites Part B: Engineering , 238:109879, June 2022

    work page 2022

  8. [7]

    William L. Brogan. Modern Control Theory. Prentice Hall, 1991. Google-Books-ID: OPFQAAAAMAAJ. 23

    work page 1991

Show all 52 references
  1. [8]

    De ep least-squares methods: an unsupervised learning- based numerical method for solving elliptic PDEs

    Zhiqiang Cai, Jingshuang Chen, Min Liu, and Xinyu Liu. De ep least-squares methods: an unsupervised learning- based numerical method for solving elliptic PDEs. Journal of Computational Physics , 420:109707, November 2020

    2020

  2. [9]

    The Deep Ritz method: A deep learning -based numerical algorithm for solving varia- tional problems, September 2017

    Weinan E and Bing Y u. The Deep Ritz method: A deep learning -based numerical algorithm for solving varia- tional problems, September 2017. arXiv:1710.00211 [cs]

    2017 arXiv

  3. [10]

    Karhunen–Loéve Ex pansion of Temporal and Spatio-Temporal Processes

    Lara Fontanella and Luigi Ippoliti. Karhunen–Loéve Ex pansion of Temporal and Spatio-Temporal Processes. In Handbook of Statistics , volume 30, pages 497–520. Elsevier, 2012

    2012

  4. [11]

    Y . C. Gerstenmaier and G. Wachutka. Heat Conduction as E igenvalue Problem. In Michael Breuer, Franz Durst, and Christoph Zenger, editors, High Performance Scientific And Engineering Computing, pages 401–408, Berlin, Heidelberg, 2002. Springer

    2002

  5. [12]

    Numerical integr ation using sparse grids

    Thomas Gerstner and Michael Griebel. Numerical integr ation using sparse grids. Numerical Algorithms , 18(3):209–232, January 1998

    1998

  6. [13]

    Jimack, and René de Bo rst

    Y ousef Ghaffari Motlagh, Peter K. Jimack, and René de Bo rst. Deep learning phase-field model for brittle fractures. International Journal for Numerical Methods in Engineerin g, 124(3):620–638, 2023

    2023

  7. [14]

    Laplacian Eigenfunction- Based Neural Operator for Learning Nonlinear Partial Differential Equations, February 2025

    Wenrui Hao and Jindong Wang. Laplacian Eigenfunction- Based Neural Operator for Learning Nonlinear Partial Differential Equations, February 2025. arXiv:2502.05571 [math-ph]

    2025

  8. [15]

    Deep autoen coders for physics-constrained data-driven nonlinear materials modeling

    Xiaolong He, Qizhi He, and Jiun-Shyan Chen. Deep autoen coders for physics-constrained data-driven nonlinear materials modeling. Computer Methods in Applied Mechanics and Engineering , 385:114034, November 2021

    2021

  9. [16]

    Latent Diffusion Models for Structural Component Design

    Ethan Herron, Jaydeep Rade, Anushrut Jignasu, Baskar G anapathysubramanian, Aditya Balu, Soumik Sarkar, and Adarsh Krishnamurthy. Latent Diffusion Models for Structural Component Design. Computer-Aided Design, 171:103707, June 2024

    2024

  10. [17]

    Holliday, John F

    Elliott G. Holliday, John F. Lindner, and William L. Dit to. Solving two-dimensional quantum eigenvalue prob- lems using physics-informed machine learning, February 20 23. arXiv:2302.01413 [physics]

    arXiv

  11. [18]

    Den sely connected neural networks for nonlinear regression

    Chao Jiang, Canchen Jiang, Dongwei Chen, and Fei Hu. Den sely connected neural networks for nonlinear regression. Entropy, 24(7):876, June 2022. arXiv:2108.00864 [cs]

    2022 arXiv

  12. [19]

    P hysics-Informed Neural Networks for Quantum Eigen- value Problems, February 2022

    Henry Jin, Marios Mattheakis, and Pavlos Protopapas. P hysics-Informed Neural Networks for Quantum Eigen- value Problems, February 2022. arXiv:2203.00451 [cs]

    2022 arXiv

  13. [20]

    NSFnets (Navier-Stokes Flow nets): Physics- informed neural networks for the incompressible Navier-St okes equations

    Xiaowei Jin, Shengze Cai, Hui Li, and George Em Karniada kis. NSFnets (Navier-Stokes Flow nets): Physics- informed neural networks for the incompressible Navier-St okes equations. Journal of Computational Physics , 426:109951, February 2021

    2021

  14. [21]

    Physics-informed ne ural network for modeling dynamic linear elasticity, January 2024

    Vijay Kag and V enkatesh Gopinath. Physics-informed ne ural network for modeling dynamic linear elasticity, January 2024. arXiv:2312.15175 [cs]

    2024

  15. [22]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, and G. E. Karniadakis. V ariation al Physics-Informed Neural Networks For Solving Partial Differential Equations, November 2019

    2019

  16. [23]

    hp-VPINNs: V ariational Physics-Informed Neural Networks With Domain Decomposition

    Ehsan Kharazmi, Zhongqiang Zhang, and George Em Karnia dakis. hp-VPINNs: V ariational Physics-Informed Neural Networks With Domain Decomposition. Computer Methods in Applied Mechanics and Engineering , 374:113547, February 2021

    2021

  17. [24]

    Zavlanos

    Reza Khodayi-Mehr and Michael M. Zavlanos. V arNet: V ar iational Neural Networks for the Solution of Partial Differential Equations, December 2019. arXiv:1912.07443 [cs]

    2019 arXiv

  18. [25]

    Conditional physics informed neural networ ks

    Alexander Kovacs, Lukas Exl, Alexander Kornell, Johan n Fischbacher, Markus Hovorka, Markus Gusenbauer, Leoni Breth, Harald Oezelt, Masao Y ano, Noritsugu Sakuma, Akihito Kinoshita, Tetsuya Shoji, Akira Kato, and Thomas Schrefl. Conditional physics informed neural networ ks. Co...

    2022 arXiv

  19. [26]

    Lagaris, A

    I.E. Lagaris, A. Likas, and D.I. Fotiadis. Artificial ne ural networks for solving ordinary and partial differentia l equations. IEEE Transactions on Neural Networks , 9(5):987–1000, September 1998. Conference Name: IEEE Transactions on Neural Networks

    1998

  20. [27]

    Ap- plication of neural networks to modelling nonlinear relati onships in ecology

    Sovan Lek, Marc Delacoste, Philippe Baran, Ioannis Dim opoulos, Jacques Lauga, and Stéphane Aulagnier. Ap- plication of neural networks to modelling nonlinear relati onships in ecology. Ecological Modelling, 90(1):39–52, September 1996

    1996

  21. [28]

    Fourier Neural Operator for Parametr ic Partial Differential Equations, May 2021

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, B urigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier Neural Operator for Parametr ic Partial Differential Equations, May 2021. arXiv:2010.08895 [cs]. 24

    2021 arXiv

  22. [29]

    Deep Ritz met hod with adaptive quadrature for linear elasticity

    Min Liu, Zhiqiang Cai, and Karthik Ramani. Deep Ritz met hod with adaptive quadrature for linear elasticity. Computer Methods in Applied Mechanics and Engineering , 415:116229, October 2023

    2023

  23. [30]

    Image Classi fication with Classic and Deep Learning Techniques, May 2021

    Òscar Lorente, Ian Riera, and Aditya Rana. Image Classi fication with Classic and Deep Learning Techniques, May 2021. arXiv:2105.04895 [cs]

    2021 arXiv

  24. [31]

    DeepONe t: Learning nonlinear operators for identifying differential equations based on the universal approximati on theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. DeepONe t: Learning nonlinear operators for identifying differential equations based on the universal approximati on theorem of operators. Nature Machine Intelligence , 3(3):218–229, March 2021. arXiv:1910.03193 [cs]

    2021 arXiv

  25. [32]

    Physics Informed Neural Net- works for heat conduction with phase change, October 2024

    Bahae-Eddine Madir, Francky Luddens, Corentin Lothod é, and Ionut Danaila. Physics Informed Neural Net- works for heat conduction with phase change, October 2024. a rXiv:2410.14216 [math]

    2024 arXiv

  26. [33]

    PRINC IP AL COMPONENTS ANALYSIS (PCA).Comput- ers and geosciences, 1992

    Andrzej Makiewicz, Waldemar Ratajczak, and I Xi. PRINC IP AL COMPONENTS ANALYSIS (PCA).Comput- ers and geosciences, 1992

    1992

  27. [34]

    Manav, R

    M. Manav, R. Molinaro, S. Mishra, and L. De Lorenzis. Pha se-field modeling of fracture with physics-informed deep learning. Computer Methods in Applied Mechanics and Engineering , 429:117104, September 2024

    2024

  28. [35]

    Implementation of CALFEM for Python

    Andreas Ottosson. Implementation of CALFEM for Python . Master’s thesis, Lund University, 2010

    2010

  29. [36]

    Hamprecht, Y oshua Bengio, and Aaron Courville

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred A. Hamprecht, Y oshua Bengio, and Aaron Courville. On the Spectral Bias of Neural Networks , May 2019. arXiv:1806.08734 [stat]

    2019 arXiv

  30. [37]

    Raissi, P

    M. Raissi, P . Perdikaris, and G.E. Karniadakis. Physic s-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonline ar partial differential equations. Journal of Computa- tional Physics, 378:686–707, February 2019

    2019

  31. [38]

    Deep Generative Models in Engineering Design: A Review

    Lyle Regenwetter, Amin Heyrani Nobari, and Faez Ahmed. Deep Generative Models in Engineering Design: A Review. Journal of Mechanical Design , 144(7):071704, July 2022

    2022

  32. [39]

    Numerical Methods for Large Eigenvalue Problems: Revised E dition

    Y ousef Saad. Numerical Methods for Large Eigenvalue Problems: Revised E dition. SIAM, January 2011. Google-Books-ID: gViDLbUDjZ8C

    2011

  33. [40]

    Sahin, M

    T. Sahin, M. von Danwitz, and A. Popp. Solving Forward an d Inverse Problems of Contact Mechanics using Physics-Informed Neural Networks. Advanced Modeling and Simulation in Engineering Sciences, 11(1):11, May

  34. [41]

    E. Seneta. Computing the stationary distribution for i nfinite Markov chains. Linear Algebra and its Applications , 34:259–267, December 1980

    1980

  35. [42]

    Deep Petrov-Galerkin Method for Solving Partial Differential Equations, January 2022

    Y ong Shang, Fei Wang, and Jingbo Sun. Deep Petrov-Galerkin Method for Solving Partial Differential Equations, January 2022. arXiv:2201.12995 [math]

    2022 arXiv

  36. [43]

    DGM: A deep learning algorithm for solving partial differential equations

    Justin Sirignano and Konstantinos Spiliopoulos. DGM: A deep learning algorithm for solving partial differential equations. Journal of Computational Physics , 375:1339–1364, December 2018. arXiv:1708.07469 [q-fin]

    2018 arXiv

  37. [44]

    Sukumar and Ankit Srivastava

    N. Sukumar and Ankit Srivastava. Exact imposition of bo undary conditions with distance functions in physics- informed deep neural networks. Computer Methods in Applied Mechanics and Engineering , 389:114333, Febru- ary 2022

    2022

  38. [45]

    George R. C. Tai and Richard Paul Shaw. Helmholtz-equat ion eigenvalues and eigenmodes for arbitrary domains. The Journal of the Acoustical Society of America , 56(3):796–804, September 1974

    1974

  39. [46]

    V allet and B

    B. V allet and B. Lévy. Spectral Geometry Processing wit h Manifold Harmonics. Computer Graphics F orum, 27(2):251–260, April 2008

    2008

  40. [47]

    Dee p Sturm–Liouville: Learnable orthogonal basis func- tions parameterized by neural networks

    David Vigouroux, Joseba Dalmau, and Louis Béthune. Dee p Sturm–Liouville: Learnable orthogonal basis func- tions parameterized by neural networks. February 2024

    2024

  41. [48]

    Mo, Bassam Izzuddin, and Chul-Woo Kim

    Jiaji Wang, Y .L. Mo, Bassam Izzuddin, and Chul-Woo Kim. Exact Dirichlet boundary Physics-informed Neural Network EPINN for solid mechanics. Computer Methods in Applied Mechanics and Engineering , 414:116184, September 2023

    2023

  42. [50]

    A Physics-Informed Neural Network Approach for Solving the Engineering Eigenvalue Problem

    Seongjoon Y oo, Min Seo Kang, Heonjun Y oon, and Taejin Kim. A Physics-Informed Neural Network Approach for Solving the Engineering Eigenvalue Problem. Preprint, 2024

    2024

  43. [51]

    A review of convo- lutional neural networks in computer vision

    Xia Zhao, Limin Wang, Y ufei Zhang, Xuming Han, Muhammet Deveci, and Milan Parmar. A review of convo- lutional neural networks in computer vision. Artificial Intelligence Review, 57(4):99, March 2024. 25

    2024

  44. [2024]

    arXiv:2308.12716 [math]

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.