Pith. sign in

REVIEW 4 major objections 10 minor 45 references

Weak TransNet: A Petrov-Galerkin based neural network method for solving elliptic PDEs

T0 review · 4 major / 10 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 Weak TransNet claims that minimizing the weak residual in a fixed neural-feature trial space against radial-basis test functions turns linear elliptic PDEs into a single linear least-squares problem, with Fourier-feature and…

desk verdict A promising training-free weak-form neural solver with a genuinely new trial/test combination, but a missing κ in the Darcy boundary term needs fixing and the baselines are quadrature-confounded. read the letter →

arxiv 2506.14812 v1 pith:4ITPVFLP submitted 2025-06-05 math.NA cs.NA

Zhihang Xu , Min Wang , Zhu Wang This is my paper
classification math.NAcs.NA MSC 65N3065N1268T07
keywords weakformulationPetrov-GalerkinmethodTransNetradialbasisfunctionsleast-squaresneuralnetworkellipticpartialdifferentialequationsmultiscaleDarcyflowpartitionofunity
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 proposes a way to solve elliptic partial differential equations without training a neural network. The solution is written as a linear combination of fixed neural features, the TransNet trial space, and the coefficients are found by minimizing the squared residual of the weak Petrov-Galerkin equations against a test space of locally supported radial basis functions. Because the PDE is linear in the coefficients, the whole computation becomes a single linear least-squares solve. The same framework is extended with Fourier feature mapping for multiscale coefficients and a partition-of-unity construction for sharp gradients and corner singularities. Numerical experiments compare the method against strong-form and Ritz-energy losses and report smaller relative L2 errors, especially for problems whose solutions lack strong regularity.

What carries the argument

The carrying object is the neural feature space U_TN = span{phi_0 = 1, phi_j = $\sigma$(gamma_j (a_j^T x + r_j))}, where a_j are random unit vectors and r_j uniform scalars, so the associated hyperplanes are distributed uniformly over the domain and the shape parameter gamma_j controls steepness of the basis. The test space is composed of radial basis functions psi_i with local support, so each stiffness entry a(phi_j, psi_i) = integral over the local support of kappa grad phi_j dot grad psi_i minus the boundary term can be evaluated by composite Simpson quadrature on a small region. The method then solves the least-squares system [A; $\beta$-tilde B] $\alpha$ approximately [f; $\beta$-tilde g]. PoU-WTN multiplies local neural bases by partition-of-unity functions and enforces solution and flux continuity across subdomain interfaces through penalty matrices.

What would settle it

Run the Darcy-flow example of Sec. 5.2 with a known weak solution, fix the number of trial and test functions, and refine the composite Simpson grid while increasing the boundary penalty beta. The central claim predicts the relative L2 error should decrease toward the quadrature and penalty limits; if instead the error plateaus above the finite element reference or diverges as beta grows, the discrete bilinear form is not a faithful discretization. A direct check of the printed boundary formula in Sec. 5.2 against integration by parts of -div(kappa grad u) also settles whether the implemented operator is the stated one.

Watch

Extended reading notes

Core claim

The paper seeks to establish that the weak residual is a more effective objective than the strong residual when the trial space is a frozen neural feature space. For an elliptic problem with operator L and boundary operator B, the approximate solution u_WTN is a linear combination of TransNet neural basis functions, and the coefficients are chosen to minimize the sum of squared differences a(u_WTN, psi_i) - l(psi_i) over radial-basis test functions psi_i, plus a boundary penalty. This converts the PDE into an overdetermined linear system whose least-squares solution gives the approximation, thereby avoiding the nonconvex optimization and ill-conditioning typical of PINN-style training. The paper demonstrates the approach on Darcy flow with discontinuous sources, multiscale permeabilities, channelized permeabilities, a Poisson problem with sharp gradients, and an L-shape domain with a solution singularity, and claims that the Fourier-enhanced F-WTN and partition-of-unity PoU-WTN variants improve accuracy further.

Load-bearing premise

The load-bearing premise is that the finite discrete least-squares Petrov-Galerkin system, built from a finite set of radial-basis test functions, composite Simpson quadrature, and a soft boundary penalty, faithfully represents the continuous weak problem, including a boundary term consistent with the integration by parts of the differential operator.

Editorial extensions

If this is right

  • Linear elliptic problems become a single least-squares solve: with M neural bases and N radial-basis tests, the coefficients are obtained from one overdetermined linear system, bypassing nonconvex optimization.
  • On the Darcy problem with a discontinuous source, where no strong solution exists, WTN reaches relative L2 errors around 1.8e-3 to 3.3e-3 with M=200 and N=200-300, beating both strong-form and deep-Ritz TransNet baselines.
  • Adding Fourier features (F-WTN) drops the multiscale Darcy-flow relative error from about 6.1e-2 to 5.8e-3, an order-of-magnitude improvement over standard WTN and the other comparators in the paper.
  • Partition-of-unity localization with interface constraints reduces the sharp-gradient Poisson error to 1.35e-4 and, with mixed shape parameters, reduces the L-shape singular problem error to 3.06e-4.
  • For nonlinear operators, the same least-squares kernel can be wrapped in Picard iteration, linearizing around the current approximation at each step and reusing the linear solve.

Reading between the lines

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

  • Editorial extension: the paper does not supply an inf-sup condition or a priori error estimate, so the empirical accuracy may not persist for high-contrast coefficients or irregular domains beyond the tested cases.
  • Editorial extension: the bilinear form printed in Sec. 5.2 writes the boundary term as -integral over dOmega of psi_i dphi_j/dn ds without the permeability kappa, which is inconsistent with integration by parts for -div(kappa grad u); if that formula is what was implemented, the reported experiments may solve a slightly modified operator.
  • Editorial extension: the mixing of shape parameters in PoU-WTN suggests a cheap way to construct multiscale trial spaces without tuning a single gamma, and this idea could transfer to other random-feature PDE methods.
  • Editorial extension: a natural testable extension is to apply F-WTN and PoU-WTN to time-dependent or nonlinear elliptic problems, or to choose RBF centers adaptively based on the weak residual, which the paper does not explore.
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 / 10 minor

Summary. This paper proposes a Petrov-Galerkin neural network method, Weak TransNet (WTN), for elliptic PDEs. The trial space is the neural feature space of a single-hidden-layer TransNet with frozen random weights, and the test space consists of radial basis functions; the coefficients are obtained by minimizing the weak residual in the least-squares sense with a soft boundary penalty. The paper also proposes two extensions: Fourier-WTN (F-WTN), which applies Fourier feature mapping to the TransNet inputs for multiscale problems, and PoU-WTN, which combines the method with a piecewise-constant partition of unity and interface penalties for problems with sharp gradients or singularities. The method is tested on 2D Darcy flow problems with discontinuous sources, multiscale permeability, channelized high-contrast permeability, a Poisson problem with a sharp gradient, and an L-shaped domain with a corner singularity, and is compared with strong-form and Ritz-energy TransNet baselines. The numerical experiments show that WTN and its extensions can be competitive and often more accurate than these baselines.

Significance. If the numerical results are supported by a consistent discretization, the paper offers an attractive workflow: for linear elliptic PDEs, the solution is obtained from a single linear least-squares solve, avoiding nonconvex optimization and the associated training difficulties. The reported gains are substantial, for example F-WTN reduces the multiscale Darcy error from 6.11% to 0.58%, and PoU-WTN with mixed shape parameters reduces the L-shape error from 4.30e-2 to 3.06e-4. The use of composite Simpson quadrature instead of Monte Carlo is a practical strength, and the paper is explicit about the empirical nature of the shape-parameter selection. However, the theoretical gaps and the incorrect boundary integral in Sec. 5.2 currently prevent the results from supporting the paper's central claim.

major comments (4)
  1. [§5.2, displayed formula for A_ij] The bilinear form A_ij = a(φ_j, ψ_i) = ∫_Ω κ∇φ_j·∇ψ_i dx − ∫_{∂Ω} ψ_i ∂φ_j/∂n ds omits the permeability κ from the boundary integral. For L[u] = −div(κ∇u), integration by parts gives ∫_Ω ψ L[u] dx = ∫_Ω κ∇u·∇ψ dx − ∫_{∂Ω} ψ κ ∂u/∂n ds. Since the RBF test functions do not vanish on ∂Ω and the boundary condition is only imposed as a soft penalty, the boundary term contributes to the discrete least-squares system. With κ ≠ 1 on ∂Ω, the corrected and displayed forms differ for any solution with a nonzero normal derivative, so the method as written solves a different boundary-value problem in general. The Sec. 5.2 experiment is insensitive to this error because the exact solution is independent of y and has zero normal derivative on all four boundary segments; the reported tables therefore do not validate the displayed formula. The bilinear form should be corrected and the experiments rerun for data with nonzero boundary flux.
  2. [§3, Eqs. (6)–(8)] The paper provides no stability or error analysis for the least-squares Petrov-Galerkin system. The method minimizes ∥Aα−f∥_2^2 with a boundary penalty, but no discrete inf-sup condition linking the RBF test space to U_TN is stated, no relation between N and M is derived, and no bound on the combined quadrature and boundary-penalty errors is given. The experiments only suggest that N slightly larger than M works well. The claim that the minimizer of (6) approximates the weak solution is therefore not supported; at minimum, a consistency estimate and a stability condition should be supplied, or the paper should be framed as purely empirical.
  3. [§4.2, Eqs. (11)–(14) and Alg. 2] The PoU-WTN method enforces the interface continuity and flux conditions only through soft penalties, with λ fixed to 1 in all experiments. No analysis is given for the consistency of this penalty formulation with the weak problem, and no sensitivity study with respect to λ is reported. Since the strong results in Secs. 5.4–5.6 depend on these interface constraints, the reader cannot assess whether the accuracy is robust or an artifact of the chosen penalty weight. A discretization error estimate or a systematic λ-study is needed.
  4. [§3.1 and §5.2–5.6] The stiffness and load entries are computed by composite Simpson quadrature on truncated RBF supports (N_l = 10), but no quadrature error analysis is provided. The support truncation itself introduces a modeling error because the Gaussian tails outside Ω_ψ are discarded, and this error is not quantified. A convergence study with respect to N_l and the Simpson mesh size would establish that the reported errors are not dominated by quadrature.
minor comments (10)
  1. [Abstract] The abstract cites 'Zhang et al., 2023' but reference [1] is dated 2024; please align the citation.
  2. [§3.1] There is a typo: 'homoegeneous' should be 'homogeneous'.
  3. [Algorithm 1] Lines 5 and 7 contain duplicated 'the the'; the wording should be cleaned up.
  4. [Figure 4 caption] The caption labels panels (a), (c), (d), (e), (f), (g) with no panel (b); the panel numbering should be corrected.
  5. [§5.2] The sentence 'The weight β for the boundary loss is set to as 1' should be 'set to 1'.
  6. [§5.3] In the discussion after Fig. 7, 'error or 7.12%' should be 'error of 7.12%'.
  7. [§5.6] The sentence 'the shape parameter for all neurons are to set to 1' should be 'the shape parameters for all neurons are set to 1'.
  8. [§4.2] The phrase 'partity of unity' should be 'partition of unity'.
  9. [§4.2] The symbol M denotes both the total number of localized basis functions and the number of trial basis functions in earlier sections; this overloaded notation should be clarified.
  10. [Appendix B] The formula for α* writes (L_DRM^T L_DRM)^{-1}; the text mentions adding a perturbation ϵI for numerical stability, but the displayed formula does not include it. Please update the equation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the WTN coefficients are determined by a linear least-squares weak residual with fixed trial/test spaces, not by fitting the reported solutions.

full rationale

The central derivation is a direct Petrov-Galerkin least-squares formulation: the trial space is the frozen TransNet feature space, the test space is a set of RBFs, and the coefficients are obtained by minimizing the weak residual against the PDE data. The reported solutions are evaluated on separate test grids against exact or FEM reference solutions, so the accuracy claims are not forced by construction. The shape parameter gamma is selected on auxiliary Gaussian projection problems rather than on the target PDE, so its choice is a hyperparameter study, not a disguised fit. The paper relies on TransNet distributional results from [1] and [31]; these are prior published theorems and do not appear to involve the current authors. There is one minor self-citation, [20], used for the hard-constraint multiplier h=x(1-x)y(1-y) in the quadrature comparison, but that device is not load-bearing for the main error tables. A genuine consistency concern is present in Sec. 5.2, where the boundary term in Aij omits the permeability factor kappa (Aij = integral of kappa grad phi_j dot grad psi_i minus boundary integral of psi_i dphi_j/dn, while integration by parts for -div(kappa grad u) requires the boundary integrand psi_i kappa dphi_j/dn), and no inf-sup or consistency analysis is supplied. This is a correctness or validation gap, not a circular reduction: the method's predictions are not equivalent to its inputs by construction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central method rests on standard weak-solution theory, the cited TransNet distributional theorem, and several unproved discretization assumptions: the fidelity of the least-squares Petrov-Galerkin system, the consistency of the PoU interface penalties, and the accuracy of Simpson quadrature on truncated RBF supports. The hyperparameters gamma, sigma, N_l, beta, lambda, and Fourier frequency settings are chosen by hand per experiment.

free parameters (7)
  • shape parameter gamma = gamma=1 for smooth problems, gamma=5 for sharp gradients, mixture {1,5,10} near singularity
    Chosen by hand in Sec 5.1 based on projection error of Gaussian targets; controls steepness of neural basis functions and is critical to expressivity.
  • RBF width sigma = sigma=0.05, with 0.03 in Sec 3.1
    Set by hand; controls the local support of test functions and the quadrature savings.
  • truncation factor N_l = 10
    Truncates Gaussian test support to reduce integration cost; used in the computational-saving estimate in Sec 3.
  • boundary penalty weight beta = 1 for WTN and PoU-WTN; optimized for SF and DRM baselines
    Weight for the soft boundary constraint in Eq (7), set per experiment.
  • interface penalty weight lambda = 1
    Weight for jump penalties across subdomain interfaces in PoU-WTN, Eq (14).
  • Fourier feature dimension P and frequency sigma_B = P=64, sigma_B in {1,3}
    Chosen without prior knowledge to capture both low and high frequencies in F-WTN, Sec 5.3.
  • trial and test basis counts M and N = M=100 to 2000, N=50 to 2000 depending on experiment
    Discretization sizes chosen per problem; the empirical rule that N should equal or exceed M is stated in Sec 5.2.
assumptions (5)
  • domain assumption Existence and uniqueness of weak solutions for the elliptic boundary value problems considered.
    Invoked implicitly in Sec 2 and Sec 5, where reference weak solutions are used; standard Lax-Milgram conditions are assumed to hold.
  • standard math Random hyperplane construction yields uniformly distributed neural basis functions (Theorem 1 of [1], Theorem 2 of [31]).
    Used in Sec 2.1 and Sec 4.1 to justify that the TransNet feature space covers the domain; cited, not reproved in this paper.
  • ad hoc to paper The finite-dimensional least-squares Petrov-Galerkin system is solvable and stable.
    No inf-sup condition, conditioning estimate, or rank analysis is given for Eq (6)-(8); the method assumes the least-squares solution is meaningful for the chosen M and N.
  • ad hoc to paper The piecewise-constant partition-of-unity basis with soft interface penalties is consistent with the weak solution.
    The PoU functions in Eq (9) are discontinuous at subdomain boundaries; the paper relies on the penalty term (14) to enforce continuity and flux conditions without a consistency proof.
  • ad hoc to paper Composite Simpson quadrature on truncated RBF supports is accurate enough for the stiffness matrix.
    Sec 5 states Simpson's rule is used for the stiffness matrix; no quadrature error analysis is provided, and the comparison in Sec 3.1 shows integration accuracy is a bottleneck.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weak TransNet: A Petrov-Galerkin based neural network method for solving elliptic PDEs." pith.science (2026). https://pith.science/paper/4ITPVFLP

@misc{pith2026250614812,
  author       = {Pith},
  title        = {Pith review of: Weak TransNet: A Petrov-Galerkin based neural network method for solving elliptic PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ITPVFLP}},
  note         = {Machine review of arXiv:2506.14812}
}
read the original abstract

While deep learning has achieved remarkable success in solving partial differential equations (PDEs), it still faces significant challenges, particularly when the PDE solutions have low regularity or singularities. To address these issues, we propose the Weak TransNet (WTN) method, based on a Petrov-Galerkin formulation, for solving elliptic PDEs in this work, though its framework may extend to other classes of equations. Specifically, the neural feature space defined by TransNet (Zhang et al., 2023) is used as the trial space, while the test space is composed of radial basis functions. Since the solution is expressed as a linear combination of trial functions, the coefficients can be determined by minimizing the weak PDE residual via least squares. Thus, this approach could help mitigate the challenges of non-convexity and ill-conditioning that often arise in neural network training. Furthermore, the WTN method is extended to handle problems whose solutions exhibit multiscale features or possess sharp variations. Several numerical experiments are presented to demonstrate the robustness and efficiency of the proposed methods.

Figures

Figures reproduced from arXiv: 2506.14812 by the authors.

Figure 1
Figure 1. Relative errors of uWTN using Monte Carlo compared to the composite Simpson’s rule for numerical quadratures. Note that γ = 1 is used in the neural basis functions, with σ = 0.03 and Nl = 10 for all test functions. Although finding the weak solution can handle problems with low regular￾ity, many practical problems involve solutions that exhibit complex behavior. To address these challenges, we next propose several e… view at source ↗
Figure 2
Figure 2. Projection error versus the shape parameter γ for different values of σf : (a) M = 200 and (b) M = 400. Colored curves correspond to different σf values, with matching markers indicating the optimal γ associated to the least projection error for each case. x ∈ Ω = [0, 1]2 , κ(x) = 1 + |x| 2 = 1 + x 2 + y 2 as shown in [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Darcy flow in Sec. 5.2: (a) permeability κ(x); and (b) reference solution. When SF is used, κ has to be differentiated, and the corresponding loss is LSF = Z Ω ∥ − κ(x)∆uTN(x) − ∇κ(x) · ∇uTN(x) − f(x)∥ 2 2 dx + βSF Z ∂Ω ∥B[uTN(x)] − g(x)∥ 2 2 ds , where βSF is a weight that has been optimized. If DRM method is instead used, the corresponding loss is LDRM = Z Ω  1 2 κ(x)|∇uTN(x)| 2 − f(x)uTN(x)  dx + βDRM Z ∂Ω ∥B[u… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Darcy flow in Sec. 5.2: (a) uWTN obtained by Alg.1; (b) uSF obtained by minimizing the strong form loss; (c) uDRM obtained by minimizing the DRM loss; (d), (e), (f) show the pointwise errors u⋆ − u e for ⋆ = WTN, SF, and DRM, respectively. (a) κ(x) (b) u e [PITH_FULL_…
Figure 5
Figure 5. Figure 5: Darcy flow in Sec. 5.3: (a) permeability κ(x); and (b) reference solution. As discussed in [44], the PINN method fails to learn this solution in a finite training budget. From the perspective of training, this is because it 22 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Darcy flow in Sec. 5.3: (a) uWTN obtained by Alg. 1; (b) uSF obtained by strong form loss; (c) uDRM obtained by DRM loss; (e),(f), (g) show pointwise errors u⋆ − u e for ⋆ = WTN, SF, and DRM, respectively. takes much more efforts for a neural network to learn the high-…
Figure 7
Figure 7. Figure 7: Darcy flow in Sec. 5.3: (a) uF−WTN; (b) uF−DRM; (c), (d) Pointwise error u⋆ − u e for ⋆ = F − WTN and F − DRM, respectively. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: Darcy flow in Sec. 5.4: (a) permeability κ(x); and (b) reference solution u e . Solving such a problem using PINN is challenging due to the high contrast permeability. In fact, when using a fully-connected neural network (FCNN) with three hidden layers, each with 50 ne…
Figure 9
Figure 9. Figure 9: Darcy flow in Sec. 5.4: (a) training loss of PINN; and (b) numerical solution obtained by PINN. To resolve the sharp gradients induced by the high-contrast permeability field, we use the PoU-WTN method developed in Sec. 4.2. The domain is divided into three nonoverlapp…
Figure 10
Figure 10. Figure 10: Darcy flow in Sec. 5.4: (a) uPoU−WTN approximation; and (b) pointwise error uPoU−WTN − u e . (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]
Figure 11
Figure 11. Figure 11: Poisson equation in Sec.5.5: (a) Exact solution u e ; (b) The domain decomposition used for PoU. compute the strong form solution and DRM solution using the same number of trial basis functions, the corresponding solutions and errors are shown in [PITH_FULL_IMAGE:fig…
Figure 12
Figure 12. Figure 12: Poisson equation in Sec.5.5: (a) uWTN obtained by Alg. 1; (c) uSF obtained by strong form loss; (e) uDRM obtained by DRM loss; (b), (d), (f) show the pointwise errors u⋆ − u e for ⋆ = WTN, SF, and DRM, respectively. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Poisson equation in Sec. 5.5: (a) uPoU−WTN obtained by Alg. 2; and (b) pointwise error [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: L-shape problem in Sec. 5.6: (a) The exact solution u e ; (b)uWTN obtained by Alg. 1; (c) uSF obtained by strong form loss; (d) and (e) Pointwise error u⋆ − u e for ⋆ = WTN and SF, respectively. Observing that the error is predominantly localized near the origin, in￾s…
Figure 15
Figure 15. Figure 15: L-shape problem in Sec. 5.6: Each row shows a PoU strategy used in PoU-WTN and the associated numerical error. Top: The global domain is decomposed into three subdomains as shown in (a), with the shape parameter set to 1 for all neural basis functions in each subdomai…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 28 canonical work pages

  1. [19]

    Shang, F

    Y. Shang, F. Wang, J. Sun, Randomized neural network with Petrov– Galerkin methods for solving linear and nonlinear partial differential equations, Communications in Nonlinear Science and Numerical Simu- lation 127 (2023) 107518

    work page 2023

  2. [1]

    Zhang, F

    Z. Zhang, F. Bao, L. Ju, G. Zhang, Transferable neural networks for partial differential equations, Journal of Scientific Computing 99 (2024) 2

    work page 2024

  3. [2]

    Whitaker, Flow in porous media I: A theoretical derivation of Darcy’s law, Transport in porous media 1 (1986) 3–25

    S. Whitaker, Flow in porous media I: A theoretical derivation of Darcy’s law, Transport in porous media 1 (1986) 3–25

    work page 1986

  4. [3]

    Efendiev, J

    Y. Efendiev, J. Galvis, T. Y. Hou, Generalized multiscale finite element methods (GMsFEM), Journal of computational physics 251 (2013) 116– 135

    work page 2013

  5. [4]

    C. Yang, X. Yang, X. Xiao, Data-driven projection method in fluid simulation, Computer Animation and Virtual Worlds 27 (3-4) (2016) 415–424. 38

    work page 2016

  6. [5]

    Y. Zhu, N. Zabaras, Bayesian deep convolutional encoder–decoder net- works for surrogate modeling and uncertainty quantification, Journal of Computational Physics 366 (2018) 415–447

    work page 2018

  7. [6]

    Y. Zhu, N. Zabaras, P.-S. Koutsourelakis, P. Perdikaris, Physics- constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data, Journal of Computa- tional Physics 394 (2019) 56–81

    work page 2019

  8. [7]

    Raissi, P

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

    2019

Show all 45 references
  1. [8]

    Y. Chen, B. Hosseini, H. Owhadi, A. M. Stuart, Solving and learn- ing nonlinear PDEs with Gaussian processes, Journal of Computational Physics 447 (2021) 110668

    2021

  2. [9]

    S. Mo, Y. Zhu, N. Zabaras, X. Shi, J. Wu, Deep convolutional encoder- decoder networks for uncertainty quantification of dynamic multiphase flow in heterogeneous media, Water Resources Research 55 (1) (2019) 703–728

    2019

  3. [10]

    S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: A review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738

    2021

  4. [11]

    DeVore, B

    R. DeVore, B. Hanin, G. Petrova, Neural network approximation, Acta Numerica 30 (2021) 327–444

    2021

  5. [12]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. Hamprecht, Y. Bengio, A. Courville, On the spectral bias of neural networks, in: International conference on machine learning, PMLR, 2019, pp. 5301– 5310

    2019

  6. [13]

    Z.-Q. J. Xu, Y. Zhang, T. Luo, Y. Xiao, Z. Ma, Frequency principle: Fourier analysis sheds light on deep neural networks, arXiv preprint arXiv:1901.06523 (2019). 39

    2019 arXiv

  7. [14]

    Sirignano, K

    J. Sirignano, K. Spiliopoulos, DGM: A deep learning algorithm for solv- ing partial differential equations, Journal of computational physics 375 (2018) 1339–1364

    2018

  8. [15]

    Yu, et al., The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems, Communications in Mathe- matics and Statistics 6 (1) (2018) 1–12

    B. Yu, et al., The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems, Communications in Mathe- matics and Statistics 6 (1) (2018) 1–12

    2018

  9. [16]

    Y. Zang, G. Bao, X. Ye, H. Zhou, Weak adversarial networks for high- dimensional partial differential equations, Journal of Computational Physics 411 (2020) 109409

    2020

  10. [17]

    Kharazmi, Z

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

    2021

  11. [18]

    Sheng, C

    H. Sheng, C. Yang, PFNN: A penalty-free neural network method for solving a class of second-order boundary-value problems on complex geometries, Journal of Computational Physics 428 (2021) 110085

    2021

  12. [20]

    P. Liu, Z. Xu, Z. Sheng, Subspace method based on neural networks for solving the partial differential equation in weak form, arXiv preprint arXiv:2405.08513 (2024)

    2024 arXiv

  13. [21]

    Y. Liao, P. Ming, Deep Nitsche method: Deep Ritz method with essen- tial boundary conditions, arXiv preprint arXiv:1912.01309 (2019)

    2019 arXiv

  14. [22]

    N. Wang, D. Zhang, H. Chang, H. Li, Deep learning of subsurface flow via theory-guided neural network, Journal of Hydrology 584 (2020) 124700

    2020

  15. [23]

    R. Xu, D. Zhang, M. Rong, N. Wang, Weak form theory-guided neural network (TgNN-wf) for deep learning of subsurface single-and two-phase flow, Journal of Computational Physics 436 (2021) 110318. 40

    2021

  16. [24]

    Z. Liu, W. Cai, Z.-Q. J. Xu, Multi-scale deep neural network (MscaleDNN) for solving Poisson-Boltzmann equation in complex do- mains, arXiv preprint arXiv:2007.11207 (2020)

    2020 arXiv

  17. [25]

    B. Wang, W. Zhang, W. Cai, Multi-scale deep neural network (MscaleDNN) methods for oscillatory stokes flows in complex domains, arXiv preprint arXiv:2009.12729 (2020)

    2020 arXiv

  18. [26]

    M.Raissi, P.Perdikaris, G.E.Karniadakis, Physicsinformeddeeplearn- ing (part I): Data-driven solutions of nonlinear partial differential equa- tions, arXiv preprint arXiv:1711.10561 (2017)

    2017 arXiv

  19. [27]

    Markidis, The old and the new: Can physics-informed deep-learning replace traditional linear solvers?, Frontiers in big Data 4 (2021) 669097

    S. Markidis, The old and the new: Can physics-informed deep-learning replace traditional linear solvers?, Frontiers in big Data 4 (2021) 669097

    2021

  20. [28]

    S. Dong, Z. Li, Local extreme learning machines and domain decom- position for solving linear and nonlinear partial differential equations, Computer Methods in Applied Mechanics and Engineering 387 (2021) 114129

    2021

  21. [29]

    J. Chen, X. Chi, Z. Yang, et al., Bridging traditional and ma- chine learning-based algorithms for solving PDEs: the random feature method, J Mach Learn 1 (2022) 268–98

    2022

  22. [30]

    Y. Wang, S. Dong, An extreme learning machine-based method for com- putational PDEs in higher dimensions, Computer Methods in Applied Mechanics and Engineering 418 (2024) 116578

    2024

  23. [31]

    T. Lu, L. Ju, L. Zhu, A multiple transferable neural network method with domain decomposition for elliptic interface problems, Journal of Computational Physics 530 (2025) 113902

    2025

  24. [32]

    S. Ding, H. Zhao, Y. Zhang, X. Xu, R. Nie, Extreme learning ma- chine: algorithm, theory and applications, Artificial Intelligence Review 44 (2015) 103–115

    2015

  25. [33]

    P. B. Bochev, M. D. Gunzburger, Least-squares finite element methods, Vol. 166, Springer Science & Business Media, 2009

    2009

  26. [34]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, G. E. Karniadakis, Variational physics-informed neural networks for solving partial differential equations, arXiv preprint arXiv:1912.00873 (2019). 41

    2019 arXiv

  27. [35]

    Shang, F

    Y. Shang, F. Wang, Randomized neural networks with Petrov–Galerkin methods for solving linear elasticity and Navier–Stokes equations, Jour- nal of Engineering Mechanics 150 (4) (2024) 04024010

    2024

  28. [36]

    J. P. Roop, A randomized neural network based Petrov–Galerkin method for approximating the solution of fractional order boundary value problems, Results in Applied Mathematics 23 (2024) 100493

    2024

  29. [37]

    Krishnapriyan, A

    A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney, Char- acterizing possible failure modes in physics-informed neural networks, Advances in neural information processing systems 34 (2021) 26548– 26560

    2021

  30. [38]

    M. D. Buhmann, Radial basis functions, Acta numerica 9 (2000) 1–38

    2000

  31. [39]

    Tancik, P

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, R. Ng, Fourier features let networks learn high frequency functions in low dimensional domains, Advances in neural information processing systems 33 (2020) 7537–7547

    2020

  32. [40]

    S. Wang, H. Wang, P. Perdikaris, On the eigenvector bias of fourier fea- ture networks: From regression to solving multi-scale pdes with physics- informedneuralnetworks, ComputerMethodsinAppliedMechanicsand Engineering 384 (2021) 113938

    2021

  33. [41]

    S. Li, Y. Xia, Y. Liu, Q. Liao, A deep domain decomposition method based on fourier features, Journal of Computational and Applied Math- ematics 423 (2023) 114963

    2023

  34. [42]

    J. M. Melenk, I. Babuška, The partition of unity finite element method: basic theory and applications, Computer methods in applied mechanics and engineering 139 (1-4) (1996) 289–314

    1996

  35. [43]

    Alnæs, J

    M. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Ring, M. E. Rognes, G. N. Wells, The fenics project version 1.5, Archive of Numerical Software 3 (100) (2015)

    2015

  36. [44]

    W. T. Leung, G. Lin, Z. Zhang, NH-PINN: Neural homogenization- based physics-informed neural network for multiscale problems, Journal of Computational Physics 470 (2022) 111539. 42

    2022

  37. [45]

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, S. G. Johnson, Physics-informed neural networks with hard constraints for inverse de- sign, SIAMJournalon ScientificComputing43(6)(2021)B1105–B1132. 43

    2021

Pith tools

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