Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Complex Physics-Informed Neural Network

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A physics-informed network with one hidden Cauchy layer solves the tested PDEs more accurately than deeper baselines.

desk verdict Solid empirical application of an existing activation to PINNs, with an overreaching high-dimensional claim that the theory and benchmarks do not actually support. read the letter →

arxiv 2502.04917 v2 pith:KM5PCBPR submitted 2025-02-07 cs.LG cs.AI

classification cs.LGcs.AI MSC 68T0765N3535J0535K05
keywords physics-informedneuralnetworksCauchyactivationfunctionintegralformulasinglehiddenlayerhigh-dimensionalPDEsHelmholtzequationuniversalapproximationPoisson
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 proposes compleX-PINN, a physics-informed neural network whose single hidden layer uses a learnable activation $\Phi(x;\mu_1,\mu_2,d)=\frac{\mu_1 x}{x^2+d^2}+\frac{\mu_2}{x^2+d^2}$, motivated by the Cauchy integral formula. It aims to establish that this one Cauchy layer is enough to approximate the smooth solutions of several PDEs that usually require deep networks, and that training the activation parameters is cheaper than adding depth. On its 2D and 3D Helmholtz, 3D heat, and 5D and 10D Poisson benchmarks, the reported relative $L^2$ errors are typically about an order of magnitude below those of RBA-PINN, BsPINN, and PIKAN, with fewer parameters than the deep baselines. If the claim holds, PINN users could replace deep stacks with a single wide layer of rational activations and get both faster training and better accuracy on the tested classes of problems.

What carries the argument

The central object is the Cauchy activation function $\Phi(x;\mu_1,\mu_2,d)=\mu_1 x/(x^2+d^2)+\mu_2/(x^2+d^2)$, one learnable rational bump per neuron with the three trainable parameters $\mu_1$, $\mu_2$, and $d$. It is derived from the 1D Cauchy integral formula: the Riemann-sum quadrature of $(1/2\pi i)\oint f(\zeta)/(\zeta-z)\,d\zeta$ yields exactly the real part of a sum of terms $\lambda_k/(\zeta_k-x)$, each of which matches the Cauchy activation form. The machinery carries the argument because the quadrature error decays exponentially for analytic $f$, so a single layer of these activations can, in principle, approximate an analytic target to high precision with modest width; the trainable parameters let the network learn the quadrature nodes and weights rather than fixing them. For $N$ dimensions, the paper abandons the tensor-product form for the scalable ansatz $\sum_k \Phi(W_{k1}x_1+\dots+W_{kN}x_N+b_k;\mu_{k1},\mu_{k2},d_k)$, and invokes Theorem 3 to show this class is still universal in $C(\mathbb{R}^N)$. That simplified form is what all experiments actually use, so it is the load-bearing mechanism for the empirical results.

What would settle it

Run the simplified high-dimensional network (Eq. 15) on a 10D Poisson problem whose exact solution couples the coordinates, for instance $u(x)=\sin(\pi x_1 x_2)+\sin(\pi x_3 x_4)+\dots+\sin(\pi x_9 x_{10})$, and record the relative $L^2$ error as the layer width $m$ grows; if the error decays only algebraically, or fails to beat the baselines on such a coupled analytic solution, the claimed transfer of exponential accuracy to the linear-combination layer is undercut.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that discretizing Cauchy's integral formula turns function approximation into a sum of rational terms of the form $\lambda_k/(\zeta_k-z)$, and that a one-layer network whose neurons realize that rational form can learn the sample points and weights directly, since the terms coincide with the trainable Cauchy activation of Eq. (8). The 1D quadrature derivation predicts exponential decay $O(r^{-m})$ of the approximation error in the number of sample points for analytic functions, an order of accuracy not available from ReLU or sigmoid layers. In high dimensions, the paper replaces the tensor-product quadrature with a simplified network that applies the same activation to linear combinations $Wx+b$ of the inputs, proving universal approximation (Theorem 3) rather than a rate. The numerical claim is that this single Cauchy layer matches or beats deep PINN architectures, often by an order of magnitude in relative $L^2$ and $L^\infty$ error, on smooth Helmholtz, heat, and high-dimensional Poisson equations.

Load-bearing premise

The central assumption is that a single Cauchy layer applied to linear combinations of the inputs keeps the exponential approximation power that the paper derives for one-dimensional Cauchy quadrature, even though only universal approximation, with no convergence rate, is proved for that high-dimensional form; the smooth, separable test solutions are also assumed to represent the high-dimensional PDEs that challenge PINNs.

Editorial extensions

If this is right

  • For the tested smooth PDEs, one Cauchy hidden layer appears sufficient, so network depth can be traded for three extra parameters per neuron.
  • Relative $L^2$ errors on the reported Helmholtz, heat, and Poisson problems are typically an order of magnitude smaller than those of RBA-PINN, BsPINN, and PIKAN under the paper's comparison protocols.
  • The network stays small: in the 2D Helmholtz case with $a_1=a_2=8$, compleX-PINN uses 3,501 parameters against BsPINN's 124,929 while reporting a smaller error.
  • The method is compatible with the usual PINN toolkit, including hard boundary constraints, Adam plus L-BFGS training, and learning-rate schedules, so the reported accuracy gain does not require a new training algorithm.

Reading between the lines

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

  • Editorial inference: if the 1D exponential convergence transfers to the linear-combination layer, the same network could serve as a parametric surrogate by feeding PDE coefficients in as extra input coordinates; the paper only gestures at operator learning.
  • Editorial inference: measuring the error of Eq. (15) as the Cauchy-layer width grows on the paper's own Helmholtz cases would separate the architecture's intrinsic approximation rate from optimizer effects, a test the paper does not run.
  • Editorial inference: the reported divergence at $d=0.01$ suggests that constraining $d$ away from zero or reparameterizing it would be needed before the method can be pushed toward stiff or multiscale problems.
  • Editorial inference: for non-smooth solutions, the analyticity assumption behind the Cauchy quadrature would need a piecewise version or domain decomposition; extending the layer to shocks is a natural next test given the paper's stated limitation.
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

3 major / 4 minor

Summary. The manuscript proposes compleX-PINN, a physics-informed neural network with a single hidden layer whose neurons use the learnable Cauchy activation Phi(x; mu1, mu2, d) = mu1 x / (x^2 + d^2) + mu2 / (x^2 + d^2). Section 3 attempts to motivate this activation from a quadrature approximation of the 1D Cauchy integral formula and, after noting the exponential cost of the multidimensional formula, replaces it in Eq. (15) by a layer of Cauchy activations applied to linear projections Wx+b. The paper reports relative L2 and L-infty errors for 2D/3D Helmholtz, 3D heat, and 5D/10D Poisson problems, comparing against RBA-PINN, BsPINN, and PIKAN, and claims order-of-magnitude improvements with far fewer parameters. A universal-approximation theorem (Theorem 3) is the only theoretical guarantee for the implemented architecture, and the conclusion explicitly acknowledges that analyticity assumptions may limit applicability to non-smooth problems.

Significance. If the reported numerical advantages withstand closer scrutiny, compleX-PINN is a worthwhile empirical contribution: it is parameter-efficient, simple to implement, compatible with standard hard constraints and optimizers, and the experimental section takes care to match baseline training protocols and to report wall-clock times. The high-dimensional evidence, however, is currently too narrow to support the abstract's general claim, and the main tables lack trial-to-trial variability. The algebraic slip in Eq. (11) is correctable, and the rate-transfer gap between Eq. (9) and Eq. (15) can be addressed either by additional theory or by tempering the claims. I see no circularity in the accuracy measurements themselves.

major comments (3)
  1. [§3.1, Eq. (11)] Equation (11) claims that the real part in Eq. (10) can be written as Phi(x - Re(zeta_k); -Re(lambda_k), Re(lambda_k) Re(zeta_k) + Im(lambda_k) Im(zeta_k), Im(zeta_k)^2). Substituting the definition (8) gives numerator -Re(lambda_k)(x - Re(zeta_k)) + Re(lambda_k) Re(zeta_k) + Im(lambda_k) Im(zeta_k) = -Re(lambda_k)x + 2Re(lambda_k) Re(zeta_k) + Im(lambda_k) Im(zeta_k), which differs from the correct numerator Re(lambda_k) Re(zeta_k) + Im(lambda_k) Im(zeta_k) - Re(lambda_k)x by an extra Re(lambda_k) Re(zeta_k) term. The correct second parameter appears to be Im(lambda_k) Im(zeta_k). This error breaks the claimed equivalence between the Cauchy quadrature and the activation form; while the trained network can absorb the discrepancy, the derivation should be corrected.
  2. [§3.2–§3.3 and §4.4] The exponential-rate argument in Section 3.1 applies to the 1D quadrature (Eq. (9)) and does not transfer to the implemented high-dimensional architecture (Eq. (15)). For Eq. (15), the paper offers only Theorem 3, the generic one-hidden-layer universal approximation theorem, with no convergence rate and no control with respect to dimension. The abstract and conclusion nevertheless claim that compleX-PINN solves high-dimensional problems that challenge PINNs. The only d=5 and d=10 experiments use the additive solution u(x)=sum_i sin(pi x_i), which is structurally aligned with a sum of ridge activations and does not probe coupled high-dimensional behavior; the product-sine benchmarks are limited to dimensions 2 and 3. This is a load-bearing gap: either add genuinely coupled or non-smooth high-dimensional benchmarks, or substantially moderate the headline claim.
  3. [§4, Tables 3–5] All main comparisons report only the best of three random seeds, with no standard deviation or per-trial values, even though Section 4.1.2 demonstrates appreciable seed/initialization sensitivity: in Table 1, varying the initialization of mu2 changes the Adam-phase error from 2.69e-5 to 3.47e-5, a spread of roughly 30%. Best-of-three point estimates therefore cannot by themselves support the asserted order-of-magnitude advantages. Please report the full set of trials or mean +/- std for every benchmark, and state how the 'best' configuration is selected.
minor comments (4)
  1. [§3.1] The claimed exponential decay should read O(r^m) with r<1, not O(r^{-m}); as written, the error bound grows as m increases.
  2. [§4.4, Table 5] The caption of Table 5 says 'fixed number of iterations (20k)' while the table header and the text specify 50,000 iterations; this inconsistency should be fixed.
  3. [§3.3, Theorem 3] The proof of Theorem 3 is only a sketch and relies on [47] and on the unstated ability of Phi to approximate monomials in one dimension; a self-contained derivation, or at least a precise statement of what is assumed from [47], would strengthen the paper.
  4. [§4] The paper does not include a code availability or reproducibility statement; providing the training scripts would help verify the wall-clock and error numbers.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found: the Cauchy activation is re-derived in-text, Theorem 3's proof is sketched, and the accuracy claims are fresh benchmark measurements; self-citations [47,48] are minor and not load-bearing.

full rationale

The claimed derivation chain is not circular. The Cauchy activation (Eq. 8) is derived in Section 3.1 from the 1D Cauchy integral formula: Eqs. (9)-(11) show that a sum of Cauchy activations reproduces the quadrature approximation of a real analytic function, so the one-dimensional representation is a derivation rather than an input assumed as output. The high-dimensional architecture (Eq. 15) is justified by Theorem 3; although the theorem is attributed to [47], a proof sketch is given in Section 3.3 using Stone-Weierstrass and the Waring decomposition [51], and it is a generic universal-approximation statement, not a uniqueness claim that forces the Cauchy choice. The central accuracy claims are new relative L2 and L-infinity measurements on Helmholtz, heat, and Poisson benchmarks with known solutions, evaluated against RBA-PINN, BsPINN, and PIKAN using reported results or public code; these errors are not derived from the model's trainable parameters, so this is not a fitted-input-called-prediction case. Two self-citations ([47] Li-Xia-Zhang and [48] Li-Xia-Zheng) share authors with the present paper, but they provide background on the activation function and a comparison architecture, while the load-bearing empirical evidence here is independent. The most serious weakness is in Section 3.3: Eq. (15) is supported only by Theorem 3's universal approximation, not by the O(r^{-m}) exponential rate of the 1D Cauchy quadrature cited from [50], and the paper does not prove a convergence rate for the simplified high-dimensional form; the Conclusion explicitly concedes that the analyticity assumptions may limit applicability to non-smooth or shock problems. That is an unproven transfer and a limitation, not a circular reduction. Accordingly, no definitional or fitted-input circularity is exhibited, and the paper is self-contained against external benchmarks; the score reflects only the minor, non-load-bearing author-overlapping citations.

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

The theoretical motivation relies on standard complex analysis and approximation facts, plus the domain assumption that PDE solutions are analytic. The practical high-dimensional architecture relies on trained network parameters and hand-chosen widths, loss weights, and schedules. The paper does not prove exponential convergence for the architecture actually used in Eq. (15), and it introduces no new physical entities.

free parameters (4)
  • Per-neuron Cauchy parameters mu1, mu2, d = trained by gradient descent; initialized to 0.1
    These are the trainable constants in the activation, Eq. (8). Table 1 shows the initialization of d materially affects results, including NaN for d=0.01.
  • Hidden layer weights W and biases b = trained by Adam and L-BFGS
    The high-dimensional representation in Eq. (15) is a generic trainable linear projection layer; no formula fixes these values.
  • Layer width N_Cauchy and default initialization = 110, 200, 300, or 500 depending on experiment; init 0.1
    Neuron counts are chosen by hand per problem and are not derived from the theory; parameter counts and accuracy depend on this choice.
  • Loss weights lambda_F and lambda_B = e.g., 0.01/1, 1/100 across experiments
    Loss weights are hyperparameters set manually in each benchmark and affect the reported accuracy.
assumptions (4)
  • standard math Cauchy's integral formula and its discretized quadrature error decay exponentially for analytic functions
    Used in Section 3.1 to motivate the activation and the claimed O(r^-m) convergence for analytic functions on compact domains.
  • domain assumption The test PDE solutions are smooth or analytic on the domains
    All benchmarks use sine products or sums; the conclusion acknowledges the method may not extend to shocks or sharp interfaces.
  • domain assumption The Cauchy activation function has the universal approximation property in one dimension
    Assumed in Section 3.3 before Theorem 3; the property is cited to reference [47] rather than reproved in this paper.
  • standard math Waring decomposition expresses multivariate monomials as sums of powers of linear forms
    Used in the proof sketch of Theorem 3 to argue that a single-layer Cauchy network can approximate polynomials and hence continuous functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Complex Physics-Informed Neural Network." pith.science (2026). https://pith.science/paper/KM5PCBPR

@misc{pith2026250204917,
  author       = {Pith},
  title        = {Pith review of: Complex Physics-Informed Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KM5PCBPR}},
  note         = {Machine review of arXiv:2502.04917}
}
read the original abstract

We propose compleX-PINN, a novel physics-informed neural network (PINN) architecture incorporating a learnable activation function inspired by the Cauchy integral theorem. By optimizing the activation parameters, compleX-PINN achieves high accuracy with just a single hidden layer. Empirically, we demonstrate that compleX-PINN solves high-dimensional problems that pose significant challenges for PINNs. Our results show that compleX-PINN consistently achieves substantially greater precision, often improving accuracy by an order of magnitude, on these complex tasks.

Figures

Figures reproduced from arXiv: 2502.04917 by the authors.

Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 1
Figure 1. Relative L 2 error history of the 2D Helmholtz equation using the Adam optimizer (red: compleX-PINN; blue: RBA-PINN), followed by 1,000 iterations of the L-BFGS optimizer. A dashed vertical line indicates the transition from Adam to L-BFGS. The Adam optimizer is configured with a learning rate of 0.005 and an exponential decay scheduler (decay rate: 0.7; decay step: 1,000). Although the large initial learning rate c… view at source ↗
Figure 2
Figure 2. Prediction and absolute point-wise errors of compleX-PINN for the 2D Helmholtz equation. The results are [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Relative L 2 error and L∞ norm history of compleX-PINN for the 2D Helmholtz equation (a1 = a2 = 8). 4.1.2 Robustness of the trainable parameters in Cauchy layers Recall that the Cauchy activation function introduced in (8) contains trainable parameters µ1, µ2, d, which…
Figure 4
Figure 4. Figure 4: Prediction and absolute point-wise errors of compleX-PINN for the 2D Helmholtz equation when [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Relative L 2 error history of compleX-PINN (red), RBA-PINN (blue), BsPINN (cyan), and PIKAN (magenta) on the 3D heat equation. The left plot shows error evolution over 20,000 training iterations. The right plot presents the error history when models are trained for the…
Figure 6
Figure 6. Figure 6: Relative L 2 error history of compleX-PINN (red), RBA-PINN (blue), BsPINN (cyan), and PIKAN (magenta) for the Poisson equation in 5D (top row) and 10D (bottom row). The left column presents error trajectories over 50,000 training iterations, while the right column comp…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Alternating Levenberg-Marquardt Training of Physics-Informed Neural Networks with Fourier-Enhanced Features

    cs.LG 2026-08 conditional novelty 5.0 of 10

    FALM-PINN alternates basis learning with Levenberg-Marquardt coefficient fitting to solve high-frequency and nonlinear PDEs with orders-of-magnitude better accuracy than baselines.

  2. Kolmogorov-Arnold Representation for Symplectic Learning: Advancing Hamiltonian Neural Networks

    cs.LG 2025-08 reject novelty 2.0 of 10

    KAR-HNN, an HNN built from univariate KAN blocks, shows mixed accuracy gains but fails to consistently reduce energy drift versus MLP-HNN.

Reference graph

Works this paper leans on

52 extracted references · 38 canonical work pages · cited by 2 Pith papers

  1. [47]

    Cauchy activation function and XNet

    Xin Li, Zhihong Xia, and Hongkun Zhang. Cauchy activation function and XNet. Neural Networks, 188:107375, 2025. 16 A PREPRINT - AUGUST 18, 2025

  2. [1]

    Raissi, P

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

  3. [2]

    Physics- informed machine learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics- informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021

  4. [3]

    Physics-informed neural networks for inverse problems in nano-optics and metamaterials

    Yuyao Chen, Lu Lu, George Em Karniadakis, and Luca Dal Negro. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Optics express, 28(8):11618–11633, 2020. 14 A PREPRINT - AUGUST 18, 2025

  5. [4]

    Non-invasive inference of thrombus material properties with physics-informed neural networks

    Minglang Yin, Xiaoning Zheng, Jay D Humphrey, and George Em Karniadakis. Non-invasive inference of thrombus material properties with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 375:113603, 2021

  6. [5]

    Physics-informed neural networks for heat transfer problems

    Shengze Cai, Zhicheng Wang, Sifan Wang, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks for heat transfer problems. Journal of Heat Transfer, 143(6):060801, 2021

  7. [6]

    Heat transfer prediction with unknown thermal boundary conditions using physics-informed neural networks

    Shengze Cai, Zhicheng Wang, Chryssostomos Chryssostomidis, and George Em Karniadakis. Heat transfer prediction with unknown thermal boundary conditions using physics-informed neural networks. In Fluids Engineering Division Summer Meeting, volume 83730, page V003T05A054. American Society of Mechanical Engineers, 2020

  8. [7]

    Physics-informed graph convolutional neural network for modeling fluid flow and heat convection

    Jiang-Zhou Peng, Yue Hua, Yu-Bai Li, Zhi-Hua Chen, Wei-Tao Wu, and Nadine Aubry. Physics-informed graph convolutional neural network for modeling fluid flow and heat convection. Physics of Fluids, 35(8), 2023

Show all 52 references
  1. [8]

    A physics-informed neural network technique based on a modified loss function for computational 2D and 3D solid mechanics.Computational Mechanics, 71(3):543–562, 2023

    Jinshuai Bai, Timon Rabczuk, Ashish Gupta, Laith Alzubaidi, and Yuantong Gu. A physics-informed neural network technique based on a modified loss function for computational 2D and 3D solid mechanics.Computational Mechanics, 71(3):543–562, 2023

  2. [9]

    A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics

    Ehsan Haghighat, Maziar Raissi, Adrian Moure, Hector Gomez, and Ruben Juanes. A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics. Computer Methods in Applied Mechanics and Engineering, 379:113741, 2021

  3. [10]

    Analyses of internal structures and defects in materials using physics-informed neural networks

    Enrui Zhang, Ming Dao, George Em Karniadakis, and Subra Suresh. Analyses of internal structures and defects in materials using physics-informed neural networks. Science advances, 8(7):eabk0644, 2022

  4. [11]

    Physics-informed neural networks (PINNs) for fluid mechanics: A review

    Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (PINNs) for fluid mechanics: A review. Acta Mechanica Sinica, 37(12):1727–1738, 2021

  5. [12]

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

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

  6. [13]

    Physics-informed neural networks for inverse problems in supersonic flows

    Ameya D Jagtap, Zhiping Mao, Nikolaus Adams, and George Em Karniadakis. Physics-informed neural networks for inverse problems in supersonic flows. Journal of Computational Physics, 466:111402, 2022

  7. [14]

    Stochastic physics-informed neural ordinary differential equations

    Jared O’Leary, Joel A Paulson, and Ali Mesbah. Stochastic physics-informed neural ordinary differential equations. Journal of Computational Physics, 468:111466, 2022

  8. [15]

    Physics-informed generative adversarial networks for stochastic differential equations

    Liu Yang, Dongkun Zhang, and George Em Karniadakis. Physics-informed generative adversarial networks for stochastic differential equations. SIAM Journal on Scientific Computing, 42(1):A292–A317, 2020

  9. [16]

    Adversarial uncertainty quantification in physics-informed neural networks

    Yibo Yang and Paris Perdikaris. Adversarial uncertainty quantification in physics-informed neural networks. Journal of Computational Physics, 394:136–152, 2019

  10. [17]

    B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data

    Liu Yang, Xuhui Meng, and George Em Karniadakis. B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data. Journal of Computational Physics, 425:109913, 2021

  11. [18]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

  12. [19]

    When and why PINNs fail to train: A neural tangent kernel perspective

    Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022

  13. [20]

    Numerical Analysis (10th)

    RL Burden, JD Faires, and AM Burden. Numerical Analysis (10th). Cengage Learning, 2015

  14. [21]

    Self-adaptive physics-informed neural networks

    Levi D McClenny and Ulisses M Braga-Neto. Self-adaptive physics-informed neural networks. Journal of Computational Physics, 474:111722, 2023

  15. [22]

    Loss-attentional physics-informed neural networks

    Yanjie Song, He Wang, He Yang, Maria Luisa Taccari, and Xiaohui Chen. Loss-attentional physics-informed neural networks. Journal of Computational Physics, 501:112781, 2024

  16. [23]

    DeepXDE: A deep learning library for solving differential equations

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE: A deep learning library for solving differential equations. SIAM review, 63(1):208–228, 2021

  17. [24]

    A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks

    Chenxi Wu, Min Zhu, Qinyang Tan, Yadhu Kartha, and Lu Lu. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 403:115671, 2023

  18. [25]

    Active learning based sampling for high-dimensional nonlinear partial differential equations

    Wenhan Gao and Chunmei Wang. Active learning based sampling for high-dimensional nonlinear partial differential equations. Journal of Computational Physics, 475:111848, 2023

  19. [26]

    Failure-informed adaptive sampling for PINNs.SIAM Journal on Scientific Computing, 45(4):A1971–A1994, 2023

    Zhiwei Gao, Liang Yan, and Tao Zhou. Failure-informed adaptive sampling for PINNs.SIAM Journal on Scientific Computing, 45(4):A1971–A1994, 2023. 15 A PREPRINT - AUGUST 18, 2025

  20. [27]

    Respecting causality for training physics-informed neural networks

    Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Respecting causality for training physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 421:116813, 2024

  21. [28]

    Mitigating propagation failures in physics- informed neural networks using retain-resample-release (R3) sampling

    Arka Daw, Jie Bu, Sifan Wang, Paris Perdikaris, and Anuj Karpatne. Mitigating propagation failures in physics- informed neural networks using retain-resample-release (R3) sampling. In International Conference on Machine Learning, pages 7264–7302. PMLR, 2023

  22. [29]

    Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations

    Ameya D Jagtap and George Em Karniadakis. Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations. Communications in Computational Physics, 28(5), 2020

  23. [30]

    Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems

    Ameya D Jagtap, Ehsan Kharazmi, and George Em Karniadakis. Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems. Computer Methods in Applied Mechanics and Engineering, 365:113028, 2020

  24. [31]

    Parallel physics-informed neural networks via domain decomposition

    Khemraj Shukla, Ameya D Jagtap, and George Em Karniadakis. Parallel physics-informed neural networks via domain decomposition. Journal of Computational Physics, 447:110683, 2021

  25. [32]

    Initialization-enhanced physics-informed neural network with domain decomposition (IDPINN)

    Chenhao Si and Ming Yan. Initialization-enhanced physics-informed neural network with domain decomposition (IDPINN). Journal of Computational Physics, 530:113914, 2025

  26. [33]

    Quantifying training difficulty and accelerating convergence in neural network-based PDE solvers

    Chuqi Chen, Qixuan Zhou, Yahong Yang, Yang Xiang, and Tao Luo. Quantifying training difficulty and accelerating convergence in neural network-based PDE solvers. arXiv preprint arXiv:2410.06308, 2024

  27. [34]

    Two-stage initial-value iterative physics- informed neural networks for simulating solitary waves of nonlinear wave equations

    Jin Song, Ming Zhong, George Em Karniadakis, and Zhenya Yan. Two-stage initial-value iterative physics- informed neural networks for simulating solitary waves of nonlinear wave equations. Journal of Computational Physics, 505:112917, 2024

  28. [35]

    Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems

    Jeremy Yu, Lu Lu, Xuhui Meng, and George Em Karniadakis. Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems. Computer Methods in Applied Mechanics and Engineering, 393:114823, 2022

  29. [36]

    KAN: Kolmogorov-Arnold Networks

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇci´c, Thomas Y Hou, and Max Tegmark. KAN: Kolmogorov-Arnold Networks. arXiv preprint arXiv:2404.19756, 2024

  30. [37]

    A comprehen- sive and fair comparison between MLP and KAN representations for differential equations and operator networks

    Khemraj Shukla, Juan Diego Toscano, Zhicheng Wang, Zongren Zou, and George Em Karniadakis. A comprehen- sive and fair comparison between MLP and KAN representations for differential equations and operator networks. Computer Methods in Applied Mechanics and Engineering, 431:117...

  31. [38]

    Yizheng Wang, Jia Sun, Jinshuai Bai, Cosmin Anitescu, Mohammad Sadegh Eshaghi, Xiaoying Zhuang, Timon Rabczuk, and Yinghua Liu. Kolmogorov–Arnold-informed neural network: A physics-informed deep learning framework for solving forward and inverse problems based on Kolmogorov–Ar...

  32. [39]

    Pinnsformer: A transformer-based framework for physics- informed neural networks

    Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics- informed neural networks. arXiv preprint arXiv:2307.11833, 2023

  33. [40]

    Physics-informed radial basis network (PIRBN): A local approximating neural network for solving nonlinear partial differential equations

    Jinshuai Bai, Gui-Rong Liu, Ashish Gupta, Laith Alzubaidi, Xi-Qiao Feng, and YuanTong Gu. Physics-informed radial basis network (PIRBN): A local approximating neural network for solving nonlinear partial differential equations. Computer Methods in Applied Mechanics and Enginee...

  34. [41]

    Binary structured physics-informed neural networks for solving equations with rapidly changing solutions

    Yanzhi Liu, Ruifan Wu, and Ying Jiang. Binary structured physics-informed neural networks for solving equations with rapidly changing solutions. arXiv preprint arXiv:2401.12806, 2024

  35. [42]

    f-PICNN: A physics-informed convolutional neural network for partial differential equations with space-time domain

    Biao Yuan, He Wang, Ana Heitor, and Xiaohui Chen. f-PICNN: A physics-informed convolutional neural network for partial differential equations with space-time domain. Journal of Computational Physics, 515:113284, 2024

  36. [43]

    Piratenets: Physics-informed deep learning with residual adaptive networks

    Sifan Wang, Bowen Li, Yuhan Chen, and Paris Perdikaris. Piratenets: Physics-informed deep learning with residual adaptive networks. Journal of Machine Learning Research, 25(402):1–51, 2024

  37. [44]

    Dasa-pinns: Differentiable adversarial self-adaptive pointwise weighting scheme for physics-informed neural networks

    Guangtao Zhang, Huiyu Yang, Fang Zhu, Yang Chen, et al. Dasa-pinns: Differentiable adversarial self-adaptive pointwise weighting scheme for physics-informed neural networks. SSRN, 2023

  38. [45]

    Augmented Physics-Informed Neural Networks (APINNs): A gating network-based soft domain decomposition methodology

    Zheyuan Hu, Ameya D Jagtap, George Em Karniadakis, and Kenji Kawaguchi. Augmented Physics-Informed Neural Networks (APINNs): A gating network-based soft domain decomposition methodology. Engineering Applications of Artificial Intelligence, 126:107183, 2023

  39. [46]

    Residual- based attention in physics-informed neural networks

    Sokratis J Anagnostopoulos, Juan Diego Toscano, Nikolaos Stergiopulos, and George Em Karniadakis. Residual- based attention in physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 421:116805, 2024

  40. [48]

    Model comparisons: XNet outperforms KAN

    Xin Li, Zhihong Jeff Xia, and Xiaotao Zheng. Model comparisons: XNet outperforms KAN. arXiv preprint arXiv:2410.02033, 2024

  41. [49]

    Wavelets based physics informed neural networks to solve non-linear differential equations

    Ziya Uddin, Sai Ganga, Rishi Asthana, and Wubshet Ibrahim. Wavelets based physics informed neural networks to solve non-linear differential equations. Scientific Reports, 13(1):2882, 2023

  42. [50]

    The exponentially convergent trapezoidal rule

    Lloyd N Trefethen and JAC Weideman. The exponentially convergent trapezoidal rule. SIAM review, 56(3):385– 458, 2014

  43. [51]

    Varieties of sums of powers

    Kristian Ranestad and Frank-Olaf Schreyer. Varieties of sums of powers. Journal fur die Reine und Angewandte Mathematik, pages 147–182, 2000

  44. [52]

    Enhanced physics-informed neural networks with augmented lagrangian relaxation method (AL-PINNs)

    Hwijae Son, Sung Woong Cho, and Hyung Ju Hwang. Enhanced physics-informed neural networks with augmented lagrangian relaxation method (AL-PINNs). Neurocomputing, 548:126424, 2023. 17

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.