Pith. sign in

REVIEW 2 major objections 6 minor 37 references

A discontinuous Galerkin plane wave neural network method for Helmholtz equation and Maxwell's equations

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper establishes that recursively augmenting discontinuous Galerkin subspaces with element-wise plane-wave neural-network maximizers converges for Helmholtz and Maxwell equations, with a computable error bound and no assumption that…

desk verdict The method and numerics are promising, but Lemma 4.1 drops the residual scale, so Theorem 4.1 is not proven as written. read the letter →

arxiv 2506.09309 v1 pith:GVDL42EX submitted 2025-06-11 math.NA cs.NA

classification math.NAcs.NA MSC 65N3065N5568T07
keywords HelmholtzequationMaxwell'sequationsplanewaveactivationdiscontinuousGalerkinneuralnetworksquasi-maximizationresidualminimizationerrorestimator
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

The paper introduces a discontinuous Galerkin plane wave neural network method that solves Helmholtz and Maxwell equations by iteratively growing a discrete solution subspace. At each step it searches for a plane-wave neural network function that nearly maximizes the residual functional, then adds that function to the subspace and recomputes the Galerkin projection. The central result is that this recursion converges in the DG energy norm with a known geometric contraction factor, provided each added function meets a quasi-maximization tolerance. A computable quantity is shown to be equivalent to the true error, so it can serve as a stopping criterion. The method avoids the bounded-parameter assumption common in Galerkin neural network analyses and keeps discrete condition numbers close to unity.

What carries the argument

The central object is the residual functional $\langle r(u_{i-1}), v\rangle = \operatorname{Re}\{L(v)-a(u_{i-1},v)\}$, whose maximum over the closed unit ball of the trial space is exactly the DG energy-norm error $|||u-u_{i-1}|||$. The approximating maximizer is sought in $V^\sigma_{n_i}\cap B$, where $V^\sigma_n$ is the set of element-wise single-hidden-layer neural networks with complex exponential plane-wave activation $e^{i\omega W\cdot x}$ and unit directions $|W|=1$. Because this set is not a linear space and not topologically closed, the paper uses a quasi-maximization formulation from residual-minimization theory; the resulting function $\varphi^{\mathrm{NN}}_i$ is then used to augment the discrete subspace $\Phi^{\mathrm{NN}}_i = \operatorname{span}\{\varphi^{\mathrm{NN}}_1,\dots,\varphi^{\mathrm{NN}}_i\}$, and the next approximation $u_i$ is the Galerkin projection onto this augmented subspace. The mechanism also yields the error estimator $\eta(u_{i-1},\varphi^{\mathrm{NN}}_i)$ and, when the quasi-maximization tolerances are small, a condition-number bound $\operatorname{cond}(K^{(j)}) < (1+\gamma)/(1-\gamma)$ for the discrete system.

What would settle it

Run Algorithm 1 on a two-dimensional Helmholtz problem with moderate wave number and, at each outer iteration, compute the optimization gap $\sup_{v\in V^\sigma_{n_i}\cap B}\langle r(u_{i-1}), v\rangle - \langle r(u_{i-1}), \varphi^{\mathrm{NN}}_i\rangle$ by evaluating the residual over a dense grid of candidate propagation directions; if this gap exceeds $(2\tau_i/(2-\tau_i))^2$ for the claimed $\tau_i$ while the error reduction predicted by Theorem 4.1 fails, the quasi-maximizer assumption is the point of collapse.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 4.1: if at iteration $j$ the network function $\varphi^{\mathrm{NN}}_j$ is a $\delta_j$-quasi-maximizer of the residual functional in $V^\sigma_{n_j}\cap B$, with $\delta_j = (2\tau_j/(2-\tau_j))^2$, then the DG energy-norm error after $i$ iterations satisfies $|||u-u_i||| \le |||u-u_0|||\prod_{j=1}^i \min\{1, 2^{3/2}\tau_j/(2-\tau_j)\}$. This gives a guaranteed contraction whenever the quasi-maximization condition is met, with no boundedness requirement on the neural network parameters. Lemma 4.1 establishes that the computable quantity $\eta(u_{i-1},\varphi^{\mathrm{NN}}_i)$, defined as the residual functional normalized by the DG norm, is equivalent to the true error $|||u-u_{i-1}|||$ up to explicit constants. The proof uses the fact that the residual maximizer is the normalized error itself, then approximates that maximizer by an element of the plane-wave neural network set via the universal approximation property and a relaxed quasi-maximization condition.

Load-bearing premise

The entire convergence and error-estimator theory assumes that the practical training procedure in Algorithm 2 actually produces a network satisfying the quasi-maximization inequality (4.19) for the chosen tolerance; the paper provides no proof, verification check, or convergence analysis for the optimization step itself.

Editorial extensions

If this is right

  • If the quasi-maximization condition (4.19) is actually realized by the training routine, Theorem 4.1 guarantees a strictly geometric reduction of the DG energy-norm error at every outer iteration.
  • The quantity $\eta(u_{i-1},\varphi^{\mathrm{NN}}_i)$ gives a computable, provably equivalent error estimate that can be used as a stopping criterion, including a companion estimate for the $L^2$ error.
  • The discrete linear systems built from the augmented plane-wave neural basis functions are provably well conditioned, with condition numbers close to one when the tolerances are small.
  • With a judicious $hp$-refinement strategy, increasing the network width and decreasing the mesh size together, the method can handle large wave numbers using roughly $h = O(\pi/\omega)$ and no more than about ten outer iterations.
  • The same convergence analysis carries over to three-dimensional time-harmonic Maxwell equations using vector plane-wave neural networks built from polarization directions.

Reading between the lines

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

  • The practical reliability of the method hinges on whether the randomly shuffled Adam optimizer actually delivers a function satisfying Definition 4.1 for the chosen tolerance; a runtime verification of inequality (4.19) would convert the theoretical guarantee into a certified adaptive solver.
  • Because the learned plane-wave directions encode information about the error's oscillatory structure, the final directions from one residual step could plausibly be reused as warm starts for neighboring frequencies or for repeated solves with different right-hand sides.
  • The same residual-augmentation mechanism should extend naturally to nonhomogeneous, anisotropic, and time-dependent wave problems, since only the local plane-wave space and the sesquilinear form need to be adapted.
  • The equivalence between $\eta$ and the true error suggests that $\eta\|\varphi^{\mathrm{NN}}_i\|_{L^2}$ could serve not only as a stopping criterion but also as a spatial error indicator for adaptive mesh refinement.
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 / 6 minor

Summary. The manuscript proposes a discontinuous Galerkin plane wave neural network (DGPWNN) method for the Helmholtz equation and time-harmonic Maxwell equations. The method recursively augments a discontinuous Galerkin ansatz space with element-wise single-hidden-layer neural network basis functions whose activation is a plane wave, trained to quasi-maximize the residual functional in the intersection of the network set and the closed unit ball. The paper proves a contraction estimate for the energy-norm error (Theorem 4.1) under a quasi-maximization assumption (Definition 4.1), gives an error-equivalence estimator (Lemma 4.1), a condition-number bound (Lemma 4.2), and presents numerical experiments for 2D/3D Helmholtz and 3D Maxwell problems, including hp-refinement and a comparison with the plane wave least squares method.

Significance. If the theoretical results were sound, the framework would be a valuable extension of Galerkin neural networks to wave problems: it removes the bounded-parameter assumption of [1], provides a computable error estimator and a guaranteed contraction factor in the energy norm, and the plane wave activation is well adapted to oscillatory solutions. The numerical experiments are extensive, covering several regimes, reporting condition numbers close to one, and comparing favorably with the PWLS method. However, the central convergence proof contains a scaling error that invalidates Lemma 4.1 as stated, and the implemented algorithm is not shown to satisfy the theoretical quasi-maximization assumption; these issues undermine the main claims as they currently stand.

major comments (2)
  1. [Section 4.4, Lemma 4.1, Eq. (4.19)-(4.29)] The proof of Lemma 4.1 drops the factor M = |||u - u_{i-1}|||. From (4.19), since <r(u_{i-1}), v> = M Re{a(phi_i, v)}, the correct consequence is Re{a(phi_i, phi^NN_i)} >= sup_{v in V_sigma_n_i ∩ B} Re{a(phi_i, v)} - delta_i / M, not - delta_i as written in the proof. Consequently the displayed bound should contain an additional term 2*delta_i/M, giving |||phi_i - phi^NN_i|||^2 <= |||phi_i - phi_hat_i|||^2 + 2*delta_i/M. When M is small relative to delta_i, inequality (4.19) is nearly vacuous: for tau=0.1 and M=0.01, a vector v with Re{a(phi_i,v)}=0 satisfies (4.19), but |||phi_i - v||| ≈ sqrt(2), far exceeding 2^{3/2} tau/(2-tau) ≈ 0.298. Thus the bound (4.28), the error equivalence (4.29), and the strict contraction in Theorem 4.1 are not established as written. This is an internal inconsistency in a load-bearing proof, not a minor typo.
  2. [Section 4.1 (Definition 4.1), Algorithm 2, and Theorem 4.1] There is a gap between the theoretical hypothesis and the implemented algorithm. Theorem 4.1 assumes that each phi^NN_j is a delta_j-quasi-maximizer satisfying (4.19). However, Algorithm 2 trains only with a randomly shuffled Adam optimizer and terminates based on a gradient-norm condition (described in Section 6); it never verifies (4.19), and the paper gives no proof that the training output satisfies it. The numerical experiments therefore do not test the assumptions of the theorem. To make the convergence claim apply to the method as presented, the algorithm must be modified to certify (4.19) (or a tractable condition that implies the needed bound), or an optimization analysis must be supplied that establishes the quasi-maximization inequality.
minor comments (6)
  1. [Section 4.1, Eq. (4.21)] The displayed inequality has mismatched angle brackets; it should read eta(u_{i-1}, phi^NN_i) >= sup_{v in V_sigma_{n_i} ∩ B} eta(u_{i-1}, v) - delta_i.
  2. [Sections 2 and 5.1] The term 'semilinear form' should be 'sesquilinear form', and 'Risez representation theorem' should be 'Riesz representation theorem'.
  3. [Algorithm 2 and Section 6] The phrase 'an randomly shuffled Adam optimizer' should be 'a randomly shuffled Adam optimizer'; additionally, the training termination condition in Section 6 references u0, which appears to be a typo for the current iterate u_{i-1}.
  4. [Section 4.4, Theorem 4.1] The text refers to 'the resulting approximate solution u_r', but the theorem concerns u_i; also the product symbol should be a standard product sign rather than the current symbol.
  5. [Section 6.1.4, Figure 6] For the case omega=64*pi, the caption gives n1=21 for the L2-error panel and n1=13 for the energy-norm panel; please clarify whether this difference is intentional.
  6. [Section 6.1.1, Figure 1 caption] The label '(Up-Bottom)' appears to be a typo; it likely should read '(Right-Bottom)' or simply '(Bottom)'.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity: the convergence result is conditional on an explicit quasi-maximization assumption; the main proof gap in Lemma 4.1 is a dropped residual scale, not a circular identification.

full rationale

The claimed results are not circular in the input/output sense. Theorem 4.1 is an implication: if φNN_j satisfies the explicit quasi-maximization inequality (4.19) of Definition 4.1 with δ_j = (2τ_j/(2−τ_j))², then the energy error contracts by the stated factor. The proof uses only Galerkin orthogonality and Lemma 4.1. Lemma 4.1 in turn is conditional on (4.19) and on the plane-wave approximation estimate (3.15) imported from [16, Lemma 4.4] and [23, Corollary 5.5]; these are external, independently published approximation results, not the target theorem. Self-citations ([1], [14]–[16], [34]) supply framework facts such as the norm property of a(·,·) and universal approximation, which are published with independent derivation and are not used to presuppose convergence. No fitted parameter is renamed as a prediction; η(ui−1,φNN_i) is proposed as an equivalent error estimator only under the same conditional inequality. Two caveats prevent a clean endorsement. First, the proof of Lemma 4.1 is not valid as written: from (4.16), ⟨r(u_{i−1}),v⟩ = |||u−u_{i−1}||| Re{a(φ_i,v)}, so (4.19) gives Re{a(φ_i,φNN_i)} ≥ sup_{Vσ∩B} Re{a(φ_i,v)} − δ_i/|||u−u_{i−1}|||, not the −δ_i used in the proof; the derived bound (4.28) therefore needs a relative-tolerance version of (4.19) and is not established. This is an internal proof gap, not circularity. Second, Algorithm 2 gives no guarantee or verification that the training output satisfies (4.19), so the practical claim is conditional. These concerns affect correctness, not whether the derivation reduces to its inputs.

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

The ledger shows that the method rests on plane wave approximation theory from the authors' earlier work and on the unproved optimizer assumption. The only hand-chosen numbers are DG weights and tolerances; no physical constants or fitted data enter the derivation.

free parameters (4)
  • Helmholtz DG weights alpha, beta = alpha = omega^2, beta = 1
    Chosen hand-tuned in the variational functional (2.4) following [14]; not fitted to target solutions, but the norm property of a(.,.) and hence the theory depends on such weights.
  • Maxwell DG weights rho1, rho2 = rho1 = rho2 = 1
    Chosen in Section 6 for the Maxwell functional (5.40), following [15]; needed for well-posedness.
  • Tolerance parameters tau_i, delta_i = delta_i = (2 tau_i / (2 - tau_i))^2, tau_i < 0.522
    Theoretical parameters controlling quasi-maximization accuracy; not fitted to data, but the convergence rate depends on them.
  • Training and stopping tolerances = tol = 1e-6, epsilon = 1e-6 or 1e-8
    Empirical stopping criteria in numerical experiments (Section 6); not part of the convergence theorem.
assumptions (4)
  • domain assumption a(.,.) defines a norm on V(Th) for the Helmholtz and Maxwell variational problems
    Used to define the energy norm and invoke the Riesz representation theorem in Sections 2 and 5.1; cited from [14,15].
  • standard math Plane wave neural networks V^sigma_n are dense in H^s(Th) via Lemma 3.1 from [16,23]
    Foundation of Lemma 4.1; approximation of homogeneous Helmholtz solutions by plane waves requires s at least 2 and sufficient width.
  • domain assumption The normalized error phi_i admits a sufficiently regular representation in H^s(Th)
    The universal approximation property used in Lemma 4.1 needs phi_i in H^s, which holds if the exact solution has enough regularity on each element.
  • ad hoc to paper Adam training returns a phi^NN_i satisfying the quasi-maximization inequality (4.19)
    Algorithm 2 does not verify (4.19); no convergence analysis of the optimizer is given, so this is an unproved practical premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A discontinuous Galerkin plane wave neural network method for Helmholtz equation and Maxwell's equations." pith.science (2026). https://pith.science/paper/GVDL42EX

@misc{pith2026250609309,
  author       = {Pith},
  title        = {Pith review of: A discontinuous Galerkin plane wave neural network method for Helmholtz equation and Maxwell's equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GVDL42EX}},
  note         = {Machine review of arXiv:2506.09309}
}
abstract

In this paper we propose a discontinuous Galerkin plane wave neural network (DGPWNN) method for approximately solving Helmholtz equation and Maxwell's equations. In this method, we define an elliptic-type variational problem as in the plane wave least square method with $h-$refinement and introduce the adaptive construction of recursively augmented discontinuous Galerkin subspaces whose basis functions are realizations of element-wise neural network functions with $hp-$refinement, where the activation function is chosen as a complex-valued exponential function like the plane wave function. A sequence of basis functions approaching the unit residuals are recursively generated by iteratively solving quasi-maximization problems associated with the underlying residual functionals and the intersection of the closed unit ball and discontinuous plane wave neural network spaces. The convergence results of the DGPWNN method are established without the assumption on the boundedness of the neural network parameters. Numerical experiments confirm the effectiveness of the proposed method.

Figures

Figures reproduced from arXiv: 2506.09309 by the authors.

Figure 1
Figure 1. Displacement of a string (Isotropic Helmholtz equation in two dimension). (Left-Up) Esti [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. Displacement of a membrane (Helmholtz equation in two dimensions). Exact error [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Displacement of a string (Isotropic Helmholtz equation in two dimension). (Left-Up) Es [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Displacement of a string (Isotropic Helmholtz equation in two dimension). (Left) Estimated [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: Displacement of a membrane (Helmholtz equation in two dimensions). Exact error [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Comparison between the DGPWNN and the PWLS method. (Left) True errors in the [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Displacement of a string (Isotropic Helmholtz equation in three dimension). (Left-Up) [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Displacement of a string (Maxwell’s equations in three dimension). (Left-Up) Estimated [PITH_FULL_IMAGE:figures/full_fig_p028_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [1]

    Ainsworth and J

    M. Ainsworth and J. Dong, Galerkin neural networks: A framework for approximating varia- tional equations with error control, SIAM J. Sci. Comput., 43(2021), pp. A2474-A2501

  2. [2]

    Ammari, B

    H. Ammari, B. Li and J. Zou, Mathematical analysis of electromagnetic scattering by dielectric nanoparticles with high refractive indices, Trans. Amer. Math. Soc., 376(2023), pp. 39-90

  3. [3]

    G. Bao, H. Zhang and J. Zou, Unique determination of periodic polyhedral structures by scat- tered electromagnetic fields, Trans. Amer. Math. Soc., 363 (2011), pp. 4527-4551

  4. [4]

    Berg and K

    J. Berg and K. Nystr ¨om, A unified deep artificial neural network approach to partial differential equations in complex geometries, Neurocomputing, 317 (2018), pp. 28-41

  5. [5]

    Cybenko, Approximation by superpositions of a sigmoidal function, Math

    G. Cybenko, Approximation by superpositions of a sigmoidal function, Math. Control Signals Systems, 2 (1989), pp. 303-314

  6. [6]

    W. E and B. Yu, The deep Ritz method: A deep learning-based numerical algorithm for solving variational problems, Commun. Math. Stat., 6 (2018), pp. 1-12

  7. [7]

    Folland, Real Analysis: Modern Techniques and Their Applications, Pure Appl

    G. Folland, Real Analysis: Modern Techniques and Their Applications, Pure Appl. Math. (N. Y .) 40, John Wiley & Sons, New York, 1999

  8. [8]

    Gittelson, R

    C. Gittelson, R. Hiptmair and I. Perugia, Plane wave discontinuous Galerkin methods: Analysis of the h-version, ESAIM Math. Model. Numer. Anal., 43 (2009), pp. 297-331

Show all 37 references
  1. [9]

    J. He, L. Li, J. Xu, and C. Zheng, ReLU deep neural networks and linear finite elements, J. Comput. Math., 38 (2020), pp. 502-527. 29

  2. [10]

    Hiptmair, A

    R. Hiptmair, A. Moiola, and I. Perugia, Plane wave discontinuous Galerkin methods for the 2D Helmholtz equation: analysis of the p-version. SIAM J. Numer. Anal., 49(2011), pp. 264-284

  3. [11]

    Hiptmair, A

    R. Hiptmair, A. Moiola, and I. Perugia, Error analysis of Trefftz-discontinuous Galerkin methods for the time-harmonic Maxwell equations, Math. Comp., 82(2013), pp. 247-268

  4. [12]

    Hiptmair, A

    R. Hiptmair, A. Moiola and I. Perugia, A survey of Tre fftz methods for the Helmholtz equation. In: Building Bridges: Con- nections and Challenges in Modern Approaches to Numerical Partial Differential Equations, Springer International Publishing, (2015), pp. 237-279

  5. [13]

    Hornik, Approximation capabilities of multilayer feedforward networks, Neural Netw., 4(1991), pp

    K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural Netw., 4(1991), pp. 251-257

  6. [14]

    Hu and L

    Q. Hu and L. Yuan, A weighted variational formulation based on plane wave basis for discretiza- tion of Helmholtz equations, Int. J. Numer. Anal. Model., 11(2014), pp. 587-607

  7. [15]

    Hu and L

    Q. Hu and L. Yuan, A Plane Wave Least-Squares Method for Time-Harmonic Maxwell’s Equa- tions in Absorbing Media, SIAM J. Sci. Comput., 36(2014), pp. A1911-A1936

  8. [16]

    Hu and L

    Q. Hu and L. Yuan, A Plane wave method combined with local spectral elements for nonho- mogeneous Helmholtz equation and time-harmonic Maxwell equations, Adv. Comput. Math., 44(2018), pp. 245-275

  9. [17]

    Huttunen, P

    T. Huttunen, P. Gamallo and R. Astley, Comparison of two wave element methods for the Helmholtz problem, Commun. Numer. Meth. Engng., 25(2009), pp. 35-52

  10. [18]

    Huttunen, J

    T. Huttunen, J. Kaipio and P. Monk. The perfectly matched layer for the ultra weak variational formulation of the 3D Helmholtz equation, Int. J. Numer. Meth. Engng., 61(2004), pp. 1072- 1092

  11. [19]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, and G. Karniadakis, Variational Physics-Informed Neural Networks for Solving Partial Differential Equations, preprint, https://arxiv.org/abs/1912.00873, 2019

  12. [20]

    Leshno, V

    M. Leshno, V . Y . Lin, A. Pinkus, and S. Schocken, Multilayer feedforward networks with a nonpolynomial activation function can approximate any function, Neural Netw., 6 (1993), pp. 861-867

  13. [21]

    A. Moiola, Tre fftz-discontinuous Galerkin Methods for Time-Harmonic Wave Prob- lems, dissertation, ETH Zurich, Switzerland, 2012; also available online from http://ecollection.library.ethz.ch/view/eth:4515?q=moiola

  14. [22]

    Moiola and I

    A. Moiola and I. Perugia, A space-time Tre fftz discontinuous Galerkin method for the acoustic wave equation in first-order formulation, Numer. Math., 138(2018), pp. 389-435

  15. [23]

    Moiola, R

    A. Moiola, R. Hiptmair, and I. Perugia, Plane wave approximation of homogeneous Helmholtz solutions, Z. Angew. Math. Phys., 62(2011), pp. 809-837

  16. [24]

    Monk and D

    P. Monk and D. Wang, A least-squares method for the helmholtz equation, Comput. Methods Appl. Mech. Engrg., 175(1999), pp. 121-136

  17. [25]

    Parolin, D

    E. Parolin, D. Huybrechs and A. Moiola, Stable approximation of Helmholtz solutions by evanescent plane waves, ESAIM: M2AN, 57(2023), pp. 3499-3536. 30

  18. [26]

    J. Peng, J. Wang and S. Shu, Adaptive BDDC algorithms for the system arising from plane wave discretization of Helmholtz equations, Int. J. Numer. Methods Eng., 116(2018), pp. 683-707

  19. [27]

    J. Peng, S. Shu, J. Wang and L. Zhong, Adaptive-Multilevel BDDC algorithm for three- dimensional plane wave Helmholtz systems, J. Comput. Appl. Math., 381(2021), Article ID: 113011

  20. [28]

    Petersen, M

    P. Petersen, M. Raslan, and F. V oigtlaender, Topological properties of the set of functions gener- ated by neural networks of fixed size, Found. Comput. Math., 21 (2021), pp. 375-444

  21. [29]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. Karniadakis, Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys., 378 (2019), pp. 686-707

  22. [30]

    Y . Shin, Z. Zhang, G. Karniadakis, Error estimates of residual minimization using neural net- works for linear PDEs, 2020, arXiv:2010.08019

  23. [31]

    Yeung, K

    T. Yeung, K. Cheung, E. Chung, S. Fu, J. Qian, Learning rays via deep neural network in a ray-based IPDG method for high-frequency Helmholtz equations in inhomogeneous media, J. Comput. Phys., 465 (2022), 111380

  24. [32]

    Yuan and Q

    L. Yuan and Q. Hu, A PWDG method for the Maxwell system in anisotropic media with piece- wise constant coefficient matrix, ESAIM:M2AN, 58(2024), pp. 1-22

  25. [33]

    L. Yuan, X. Wang and X. Yue, A space-time Tre fftz dG method for the second order time- dependent Maxwell system in anisotropic media, J. Comput. Math., doi:10.4208 /jcm.2410- m2024-0141, 2024

  26. [34]

    Yuan and Q

    L. Yuan and Q. Hu, A discontinuous plane wave neural network method for Helmholtz equation and time-harmonic Maxwell’s equations, Adv. Comput. Math., 51(2025): 18

  27. [35]

    Y . Zang, G. Bao, X. Ye, and H. Zhou, Weak adversarial networks for high-dimensional partial differential equations, J. Comput. Phys., 411 (2020), Article ID: 109409

  28. [36]

    Zhang, C

    Y . Zhang, C. Chen, N. Shi, R. Sun and Z. Luo, Adam Can Converge Without Any Modification On Update Rules, Part of Advances in Neural Information Processing Systems 35, (NeurIPS 2022)

  29. [37]

    M. Zhao, N. Zhu and L. Wang, The electromagnetic scattering from multiple arbitrarily shaped cavities with inhomogeneous anisotropic media, J. Comput. Phys., 489(2023), 112274. 31

Pith tools

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