Pith. sign in

REVIEW 3 major objections 4 minor 78 references

A holomorphic Kolmogorov-Arnold network framework for solving elliptic problems on arbitrary 2D domains

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that holomorphic neural networks, extended via Vekua operators and Laurent-series/logarithmic-conjugation terms, solve Laplace, biharmonic, elasticity, and Helmholtz problems on arbitrary 2D domains with boundary-only…

desk verdict A genuinely useful PIHKAN and a clean Helmholtz demo, but the multiply-connected extension is oversold: the log-conjugation transfer to Table 1 is unproven and false for the biharmonic row. read the letter →

arxiv 2507.22678 v1 pith:ERI4JNV4 submitted 2025-07-30 cs.CE

classification cs.CE
keywords physics-informedneuralnetworksKolmogorov-ArnoldholomorphicLaurentseriesmultiply-connecteddomainsVekuaoperatorsHelmholtzequationlinearelasticity
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

Physics-informed holomorphic networks solve an elliptic PDE by construction: the solution is written through holomorphic potentials, so the differential equation is satisfied identically and training only fits boundary data. This paper claims to remove three limits of that idea in one framework. A new Kolmogorov-Arnold variant, PIHKAN, uses monomial activations with variance-preserving initialization and reaches higher accuracy with fewer parameters than the earlier MLP-based PIHNN and vanilla PINNs. A Vekua operator extends the approach to the Helmholtz equation, and a Laurent-series/logarithmic-conjugation construction allows domains with holes without domain decomposition. If the claims hold, boundary-only neural surrogates become a practical option for a broad class of 2D elliptic problems, including acoustics and elasticity on perforated geometries.

What carries the argument

The load-bearing object is the holomorphic representation of PDE solutions: for Laplace, biharmonic, linear elasticity, and Helmholtz, the unknown field is built from holomorphic potentials by formulas such as those in the paper's Table 1, so the PDE is satisfied by construction and only boundary conditions need to be learned. The new architecture PIHKAN replaces exponential-activation MLPs with Kolmogorov-Arnold layers whose trainable activations are monomials $\sum_{p=0}^{P} W_{p}z^{p}$, chosen because the monomials are $L^2$-orthogonal on the unit disc and on the unit circle, and it is stabilized by a variance-preserving weight initialization. The Helmholtz extension uses the Vekua operator $V_\beta[g](x)=g(x)-\int_0^1 \frac{J_1(\beta|x|\sqrt{1-t})}{2\sqrt{1-t}\beta|x|}g(tx)\,dt$, which maps harmonic functions to Helmholtz solutions. The multiply connected extension uses the logarithmic-conjugation and Laurent form $\phi_n(z)=\phi_{n,0}(z)+\sum_s \phi_{n,s}(1/(z-z_s))+\sum_s c_{n,s}\log(z-z_s)$, with trainable real coefficients $c_{n,s}$, so a single network plus small singular subnetworks covers each hole without domain decomposition.

What would settle it

Run the Laurent-based PIHKAN on Helmholtz or biharmonic problems over an annular or multi-hole domain with non-constant boundary data and compare against a high-resolution FEM reference; if the relative $L^2(\Omega)$ error does not converge toward the FEM solution as epochs, polynomial degree, and boundary points are increased, the claimed completeness of the log-term extension for non-Laplace problems is false.

Watch

Extended reading notes

Core claim

The central claim is that every 2D elliptic problem admitting a holomorphic potential representation — Laplace, biharmonic, linear elasticity, and now Helmholtz via the Vekua operator $V_\beta$ — can be solved by a holomorphic neural network trained only on the boundary, and that multiply connected domains are no longer a barrier. The paper introduces PIHKAN, a KAN whose trainable activations are complex monomials $\sum_{p=0}^{P} W_{p}z^{p}$, with an initialization that keeps variance near unity across layers, and augments each potential with terms $\sum_{s} c_{n,s}\log(z-z_s)$ and $\sum_{s} \phi_{n,s}(1/(z-z_s))$ to cover holes. In the reported tests PIHKAN reaches relative $L^2(\Omega)$ errors of $1.76\times10^{-2}$ on the L-shaped Laplace benchmark and $4.89\times10^{-3}$ on a Helmholtz square with wave number $\beta\approx18$, versus $1.18\times10^{-1}$ and $2.66\times10^{-2}$ for vanilla PINNs; on the multiply connected elasticity benchmark the Laurent-based PIHKAN reduces the $\sigma_{yy}$ error to $8.43\times10^{-3}$, where domain-decomposition PIHNN reaches $4.70\times10^{-2}$.

Load-bearing premise

The method relies on the assumption that the logarithmic-conjugation correction proven for harmonic functions also completes the complex-potential representations for biharmonic, linear-elasticity, and Helmholtz equations on domains with holes, an assumption the paper verifies numerically only for one symmetric linear-elasticity case.

Editorial extensions

If this is right

  • On all tested benchmarks the holomorphic methods match or beat FEM reference accuracy using only boundary training points, so interior collocation becomes unnecessary for these elliptic PDE classes.
  • PIHKAN reaches the reported accuracies with roughly half to one-third the parameters of the MLP-based PIHNN, implying similar savings should appear in any 2D problem with holomorphic potentials.
  • The Laurent/log construction replaces domain decomposition for multiply connected domains, so holes, cracks, and inclusions can be handled by one base network plus small subnetworks per hole.
  • The Vekua-operator route opens boundary-only training for wave problems, and the Helmholtz square test at $\beta\approx18$ shows acoustic benchmarks are within reach.
  • Boundary-based residual adaptive sampling cuts the Laplace error from $1.76\times10^{-2}$ to $7.13\times10^{-3}$ at negligible extra cost, identifying adaptive boundary collocation as a cheap accuracy lever.

Reading between the lines

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

  • The paper validates the Laurent/log construction on only one symmetric elasticity example; if the log-completion conjecture transfers, the same one-network construction should handle asymmetric multi-hole Helmholtz and biharmonic problems, which is a direct next experiment the paper leaves open.
  • Because the PDE is hard-coded, inverse problems such as identifying hole location, shape, or boundary impedance from interior measurements would reduce to fitting a boundary/interior measurement loss; the paper does not explore this direction.
  • The monomial-basis KAN inherits polynomial approximation theory, so for smooth potentials error should decay rapidly as the polynomial degree $P$ increases; this implies $P$-refinement may be a cheaper accuracy path than adding layers, which the paper does not demonstrate.
  • The reported KAN-versus-MLP wall-clock gap is attributed to unoptimized evaluation of the polynomial activations; a closed-form or recurrence-based evaluation would likely close most of the gap, leaving PIHKAN both smaller and faster than PIHNN.
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 paper proposes three extensions to physics-informed holomorphic neural networks (PIHNNs): a Kolmogorov-Arnold variant called PIHKAN based on monomial (polynomial) activation functions; a Vekua-operator treatment of the 2D Helmholtz equation; and a Laurent-series/logarithmic-conjugation ansatz, Eq. (15)-(16), intended to remove the restriction of holomorphic networks to simply-connected domains. Numerical tests are reported for the Poisson/Laplace problem on an L-shaped domain, the Helmholtz equation on a square, and linear elasticity on a square with a circular hole, comparing PIHKAN, PIHNN, classical PINN, and FEM. The central claim is that holomorphic networks can now solve Laplace, biharmonic, elasticity, and Helmholtz problems on arbitrary multiply-connected 2D domains with boundary-only training and high accuracy.

Significance. If the mathematical claims were fully supported, the paper would make a valuable contribution to physics-informed machine learning: boundary-only training, hard PDE satisfaction, sharply smaller parameter counts, and an alternative to domain decomposition for multiply-connected geometries. The numerical study is reported with concrete hyperparameters, parameter counts, and FEM reference solutions, and the PIHKAN results on the three benchmarks show consistently lower relative L2 errors than the compared PINN/domain-decomposition baselines; the ablation study in Table 3 and the RAD experiment in Table 4 are useful additions. However, the multiply-connected extension -- the paper's headline novelty -- is not established for the biharmonic and Helmholtz rows of Table 1, and the single multiply-connected test does not exercise the logarithmic-conjugation terms at all. The central claim is therefore stronger than the evidence and mathematics in the manuscript support.

major comments (3)
  1. [Section 2.4, Eq. (16) and Table 1 (biharmonic row)] The assertion that applying Eq. (15) to each complex potential is sufficient to extend all representations in Table 1 to multiply-connected domains is false for the biharmonic equation. On the annulus 1<|z|<2, u(r)=r^2 log r is biharmonic. With z_s=0, the ansatz (16) generates terms of the form Re(zbar z^n) (radial only for n=1, giving r^2), Re(zbar z^{-k}) for k>=1 (all carrying angular factors e^{-i(k+1)theta}), and logarithmic terms whose real parts are log r or r theta sin theta; no combination, even with infinite Laurent series, produces the radial mode r^2 log r. The exact solution therefore lies outside the model class. A correct extension needs additional multivalued potentials (for example z log z in phi_1) or a separate biharmonic logarithmic-conjugation theorem with proof; the sentence 'Since all representations in Table 1 stem from the holomorphic characterization of harmonic functions' does not supply this.
  2. [Section 2.4, Eq. (3) (Helmholtz row)] For the Helmholtz equation, the claimed transfer to multiply-connected domains is unsupported. The Vekua operator in Eq. (3) integrates g(tx) along the segment from 0 to x; if the domain is not star-shaped with respect to the chosen origin, or if the origin lies inside a hole, the integral samples points outside Omega and the formula is not defined. The paper includes no multiply-connected Helmholtz test and no modification of the Vekua operator for such domains. Moreover, Eq. (2) requires the image of F* to equal the kernel of F, so the Helmholtz representation needs a concrete surjectivity statement; Section 4 itself notes that 'little is known' about whether the Vekua map is surjective. The authors should either state the precise hypotheses from [44] under which surjectivity holds or explicitly restrict the Helmholtz claim to simply-connected (e.g., star-shaped) domains.
  3. [Section 3.3, Table 6] The only multiply-connected numerical experiment, the centered-hole elasticity benchmark, does not validate the logarithmic-conjugation part of the proposed method. For this symmetric problem with a self-equilibrated traction system, the Kolosov-Muskhelishvili potentials are single-valued and no log(z-z_s) term is active; the Laurent ansatz is exercised only through the negative powers 1/(z-z_1). Table 6 therefore demonstrates that negative powers help here, but it gives no evidence that Eq. (15) transfers Axler's theorem to elasticity, let alone to the biharmonic and Helmholtz rows. A test in which log terms are genuinely required (for plate bending with a hole, or an elasticity problem with nonzero resultant force) would be needed to support the central claim.
minor comments (4)
  1. [Table 5 caption] The caption of Table 5 says 'comparison between the different methods employed in Section 3.1', but the table reports the Helmholtz experiment of Section 3.2.
  2. [Table 1 and Section 3.3] In the displayed Kolosov-Muskhelishvili formulas, the complex-conjugate bars appear to be missing: the standard displacement formula has kappa phi_1 - z \overline{phi_1'} - \overline{phi_2}, not kappa phi_1 - z phi_1' - phi_2. If the bars are present in the typeset version, the plain-text rendering is misleading; if they are absent, the formulas are incorrect.
  3. [Section 2.4, Eq. (16)] The phrase that z_s 'should be chosen as the most internal point of B_s' is ambiguous for non-circular holes; the operative mathematical requirement is that the Laurent expansion with center z_s converges on the annular region between the hole and the outer boundary, which depends on the distance to the boundary of the hole, not just on a point selected inside it.
  4. [Section 2.3] The paper gives no universal-approximation statement for PIHKANs. For the method to be a convergent solver as network size grows, the ansatz class must be dense in the relevant holomorphic function spaces; for polynomial activations this can presumably be obtained from polynomial/Runge approximation, but the assumption is not stated in the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: representations are external, training is boundary-only, and FEM is the independent reference.

full rationale

The derivation chain is self-contained with respect to its inputs. The holomorphic representations in Table 1 (Goursat, Kolosov–Muskhelishvili, Vekua) are taken from external classical references, and the Vekua operator for Helmholtz is attributed to Moiola et al.; none of these representations is derived from the paper's own fitted quantities. PIHNN/PIHKAN training enforces Equation (2) as a hard constraint and minimizes only the boundary residual in Equation (9), so the reported L2 errors are computed against independent FEM solutions rather than recovered from training data. The architecture contribution (monomial PIHKAN layer, Eq. (12)) is a genuine representation choice; its universal approximation property is cited from external work, and its weight initialization is a heuristic derivation, neither of which makes the benchmark results true by definition. Self-citations to Calafà et al. [30, 31] appear only as prior-baseline context (e.g., 'as already expected from Calafà et al. [30]' and the replicated Section 4.1.2 benchmark), and the multiply-connected benchmark is validated against FEM; no load-bearing claim reduces to these citations. The main mathematical weakness, the unproved transfer of Axler's logarithmic-conjugation theorem to the biharmonic, elasticity, and Helmholtz rows in Section 2.4, is a correctness/completeness concern rather than a circularity: the paper asserts the transfer by analogy ('Since all representations in Table 1 stem from the holomorphic characterization of harmonic functions'), but this assertion is not established by defining the conclusion into the premise. It should be scored as a correctness risk, not as a circular step.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The framework rests on classical external results (Axler, Moiola/Vekua, Kolosov-Muskhelishvili) and one unproved extension of the multiply-connected correction to non-harmonic representations. The weight-initialization derivation additionally assumes a complex Gaussian distribution of activations. No new physical entities or fitted physical parameters are introduced.

free parameters (1)
  • z_s (Laurent expansion centers) = origin (center of the circular hole) in Section 3.3
    Chosen by hand as the most internal point of each hole to keep |z - z_s| bounded away from zero; the choice affects numerical conditioning but not theoretical completeness.
assumptions (4)
  • standard math Axler's logarithmic conjugation theorem: every harmonic function on a multiply connected domain can be written as Re(phi) + sum c_s log|z - z_s|
    Invoked as Theorem 1, Section 2.4, with citation [64].
  • domain assumption The Vekua operator V_beta maps harmonic functions onto solutions of the Helmholtz equation with sufficient completeness for boundary training
    Relied on for the Helmholtz test, Section 3.2; citation [44] proves the mapping, but surjectivity is noted as not fully known in the conclusion.
  • ad hoc to paper The logarithmic-conjugation correction extends unchanged to biharmonic, elasticity, and Helmholtz representations in Table 1
    Asserted in Section 2.4 with no proof: 'Since all representations in Table 1 stem from the holomorphic characterization of harmonic functions, we conclude...'.
  • domain assumption Inputs after normalization are approximately standard complex Gaussian, used to derive the PIHKAN weight initialization
    Section 2.3 variance analysis; the paper itself notes the real-valued case differs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A holomorphic Kolmogorov-Arnold network framework for solving elliptic problems on arbitrary 2D domains." pith.science (2026). https://pith.science/paper/ERI4JNV4

@misc{pith2026250722678,
  author       = {Pith},
  title        = {Pith review of: A holomorphic Kolmogorov-Arnold network framework for solving elliptic problems on arbitrary 2D domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERI4JNV4}},
  note         = {Machine review of arXiv:2507.22678}
}
read the original abstract

Physics-informed holomorphic neural networks (PIHNNs) have recently emerged as efficient surrogate models for solving differential problems. By embedding the underlying problem structure into the network, PIHNNs require training only to satisfy boundary conditions, often resulting in significantly improved accuracy and computational efficiency compared to traditional physics-informed neural networks (PINNs). In this work, we improve and extend the application of PIHNNs to two-dimensional problems. First, we introduce a novel holomorphic network architecture based on the Kolmogorov-Arnold representation (PIHKAN), which achieves higher accuracy with reduced model complexity. Second, we develop mathematical extensions that broaden the applicability of PIHNNs to a wider class of elliptic partial differential equations, including the Helmholtz equation. Finally, we propose a new method based on Laurent series theory that enables the application of holomorphic networks to multiply-connected plane domains, thereby removing the previous limitation to simply-connected geometries.

Figures

Figures reproduced from arXiv: 2507.22678 by the authors.

Figure 1
Figure 1. Visualization of multi-layer feed-forward network. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Sketch of multiply-connected domain with [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Contour plots of the various solutions from Section 3.1. The color represents the value of [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Errors |u−u˜| of the various solutions from Section 3.1 with respect to the reference FEM solution. Different color ranges are adopted. Method Training time (s) #parameters Relative L 2 (Ω) error PINN 526 7851 1.18 · 10−1 PIHNN 13 7622 2.77 · 10−2 PIHKAN 12 4302 1.76 ·…
Figure 5
Figure 5. Figure 5: Comparison of the training points before (left) and after (right) the application of the boundary [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Training and test loss curves for the experiments in Section 3.1 and Section 3.2. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Contour plots of the various solutions from Section 3.2. The color represents the value of [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Errors |u−u˜| of the various solutions from Section 3.2 with respect to the reference FEM solution. Different color ranges are adopted. 3.3. Linear elasticity on multiply-connected domain We test a multiply-connected domain to assess the effectiveness of the method pro…
Figure 9
Figure 9. Figure 9: Geometry for the test in Section 3.3. A uniaxial tension with unitary intensity [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Contour plots of the various solutions from Section 3.3. The color represents the value of the [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Components of the error |σ − σ˜| from Section 3.3 with respect to the reference FEM solution. to previous works [30, 31] for further details on linear elasticity solvers through holomorphic neural networks. Learned solutions can be visually compared in [PITH_FULL_IMA…
Figure 12
Figure 12. Figure 12: Training and test loss curves for the experiment in Section 3.3. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 42 canonical work pages

  1. [44]

    Vekua theory for the helmholtz operator.Zeitschrift f¨ ur angewandte Mathematik und Physik, 62(5):779–807, 2011

    Andrea Moiola, Ralf Hiptmair, and Ilaria Perugia. Vekua theory for the helmholtz operator.Zeitschrift f¨ ur angewandte Mathematik und Physik, 62(5):779–807, 2011. doi: 10.1007/s00033-011-0142-3. 26

  2. [1]

    Deep learning.nature, 521(7553): 436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.nature, 521(7553): 436–444, 2015. doi: 10.1038/nature14539

  3. [2]

    Springer,

    Rudolf Kruse, Christian Borgelt, Christian Braune, Sanaz Mostaghim, Matthias Stein- brecher, Frank Klawonn, and Christian Moewes.Computational intelligence. Springer,

  4. [3]

    Approximation by superpositions of a sigmoidal function.Mathemat- ics of control, signals and systems, 2(4):303–314, 1989

    George Cybenko. Approximation by superpositions of a sigmoidal function.Mathemat- ics of control, signals and systems, 2(4):303–314, 1989. doi: 10.1007/BF02551274

  5. [4]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440,

    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,

  6. [5]

    Neural algorithm for solving differential equations.Journal of Computational Physics, 91(1):110–131, 11 1990

    Hyuk Lee and In Seok Kang. Neural algorithm for solving differential equations.Journal of Computational Physics, 91(1):110–131, 11 1990. ISSN 0021-9991. doi: 10.1016/ 0021-9991(90)90007-N

  7. [6]

    Psichogios and Lyle H

    Dimitris C. Psichogios and Lyle H. Ungar. A hybrid neural network-first principles approach to process modeling.AIChE Journal, 38(10):1499–1511, 1992. doi: 10.1002/ aic.690381003

  8. [7]

    M. W. M. G. Dissanayake and Nhan Phan-Thien. Neural-network-based approximations for solving partial differential equations.Communications in Numerical Methods in Engineering, 10(3):195–201, 1994. doi: doi.org/10.1002/cnm.1640100303

Show all 78 references
  1. [8]

    A. J. Meade and A. A. Fernandez. Solution of nonlinear ordinary differential equations by feedforward neural networks.Mathematical and Computer Modelling, 20(9):19–44, 11 1994. ISSN 0895-7177. doi: 10.1016/0895-7177(94)00160-X

  2. [9]

    A. J. Meade and A. A. Fernandez. The numerical solution of linear ordinary differential equations by feedforward neural networks.Mathematical and Computer Modelling, 19 (12):1–25, 6 1994. ISSN 0895-7177. doi: 10.1016/0895-7177(94)90095-7. 23

  3. [10]

    Artificial neural networks for solving ordinary and partial differential equations.IEEE Transactions on Neural Networks, 9(5):987–1000, 1998

    Isaac E Lagaris, Aristidis Likas, and Dimitrios I Fotiadis. Artificial neural networks for solving ordinary and partial differential equations.IEEE Transactions on Neural Networks, 9(5):987–1000, 1998. doi: 10.1109/72.712178

  4. [11]

    Karniadakis

    Maziar Raissi, Paris Perdikaris, and George 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,

  5. [12]

    Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92 (3):88, 2022

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92 (3):88, 2022. doi: 10.100...

  6. [13]

    KAN: Kolmogorov-Arnold networks

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇ ci´ c, Thomas Y Hou, and Max Tegmark. KAN: Kolmogorov-Arnold networks. arXiv preprint, 2024. doi: 10.48550/arXiv.2404.19756

  7. [14]

    On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition

    Andrei Nikolaevich Kolmogorov. On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. In Doklady Akademii Nauk, volume 114, pages 953–956. Russian Academy of Sciences, 1957

  8. [15]

    Kolmogorov- Arnold networks (KANs) for time series analysis.arXiv preprint, 2024

    Cristian J Vaca-Rubio, Luis Blanco, Roberto Pereira, and M` arius Caus. Kolmogorov- Arnold networks (KANs) for time series analysis.arXiv preprint, 2024. doi: 10.48550/ arXiv.2405.08790

  9. [16]

    Mohamed Abd Elaziz, Ibrahim Ahmed Fares, and Ahmad O. Aseeri. CKAN: Convolu- tional Kolmogorov–Arnold networks model for intrusion detection in IoT environment. IEEE Access, 12:134837–134851, 2024. doi: 10.1109/ACCESS.2024.3462297

  10. [17]

    KAN 2.0: Kolmogorov-Arnold networks meet science.arXiv preprint, 2024

    Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Matusik, and Max Tegmark. KAN 2.0: Kolmogorov-Arnold networks meet science.arXiv preprint, 2024. doi: 10.48550/ arXiv.2408.10205

  11. [18]

    Kolmogorov-Arnold networks for time series: Bridging predictive power and interpretability.arXiv preprint, 2024

    Kunpeng Xu, Lifei Chen, and Shengrui Wang. Kolmogorov-Arnold networks for time series: Bridging predictive power and interpretability.arXiv preprint, 2024. doi: 10. 48550/arXiv.2406.02496

  12. [19]

    Zhaoyang Zhang, Qingwang Wang, Yinxing Zhang, Tao Shen, and Weiyi Zhang. Physics-informed neural networks with hybrid Kolmogorov-Arnold network and aug- mented Lagrangian function for solving partial differential equations.Scientific Reports, 15(1):10523, 2025. doi: 10.1038/s4...

  13. [20]

    Kolmogorov-Arnold networks (KAN) for time series classification and robust analysis

    Chang Dong, Liangwei Zheng, and Weitong Chen. Kolmogorov-Arnold networks (KAN) for time series classification and robust analysis. InInternational Confer- ence on Advanced Data Mining and Applications, pages 342–355. Springer, 2024. doi: 10.1007/978-981-96-0840-9 24. 24

  14. [21]

    B-spline signal processing

    Michael Unser, Akram Aldroubi, and Murray Eden. B-spline signal processing. I. The- ory.IEEE transactions on signal processing, 41(2):821–833, 1993. doi: 10.1109/78. 193220

  15. [22]

    Wav-KAN: Wavelet Kolmogorov-Arnold networks

    Zavareh Bozorgasl and Hao Chen. Wav-KAN: Wavelet Kolmogorov-Arnold networks. arXiv preprint, 2024. doi: 10.48550/arXiv.2405.12832

  16. [23]

    Kolmogorov-Arnold networks are radial basis function networks.arXiv preprint, 2024

    Ziyao Li. Kolmogorov-Arnold networks are radial basis function networks.arXiv preprint, 2024. doi: 10.48550/arXiv.2405.06721

  17. [24]

    Chebyshev polynomial-based Kolmogorov- Arnold networks: An efficient architecture for nonlinear function approximation.arXiv preprint, 2024

    Sidharth SS, Keerthana A R, and Anas KP. Chebyshev polynomial-based Kolmogorov- Arnold networks: An efficient architecture for nonlinear function approximation.arXiv preprint, 2024. doi: 10.48550/arXiv.2405.07200

  18. [25]

    Deepokan: Deep operator network based on Kolmogorov arnold networks for mechanics problems.Computer Methods in Applied Mechanics and Engineering, 436:117699, 2025

    Diab W Abueidda, Panos Pantidis, and Mostafa E Mobasher. Deepokan: Deep operator network based on Kolmogorov arnold networks for mechanics problems.Computer Methods in Applied Mechanics and Engineering, 436:117699, 2025. doi: 10.1016/j.cma. 2024.117699

  19. [26]

    Learning nonlinear operators via DeepONet based on the universal approxi- mation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karni- adakis. Learning nonlinear operators via DeepONet based on the universal approxi- mation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021. doi: 10.1038/s42256-021-00302-5

  20. [27]

    From PINNs to PIKANs: Recent advances in physics-informed machine learning.Machine Learning for Compu- tational Science and Engineering, 1(1):1–43, 2025

    Juan Diego Toscano, Vivek Oommen, Alan John Varghese, Zongren Zou, Nazanin Ah- madi Daryakenari, Chenxi Wu, and George Em Karniadakis. From PINNs to PIKANs: Recent advances in physics-informed machine learning.Machine Learning for Compu- tational Science and Engineering, 1(1):...

  21. [28]

    Khemraj Shukla, Juan Diego Toscano, Zhicheng Wang, Zongren Zou, and George Em Karniadakis. A comprehensive and F AIR comparison between MLP and KAN rep- resentations for differential equations and operator networks.Computer Methods in Applied Mechanics and Engineering, 431:117...

  22. [29]

    Benjamin C Koenig, Suyong Kim, and Sili Deng. KAN-ODEs: Kolmogorov–Arnold network ordinary differential equations for learning dynamical systems and hidden physics.Computer Methods in Applied Mechanics and Engineering, 432:117397, 2024. doi: 10.1016/j.cma.2024.117397

  23. [30]

    Engsig-Karup, and Tito Andriollo

    Matteo Calaf` a, Emil Hovad, Allan P. Engsig-Karup, and Tito Andriollo. Physics- informed holomorphic neural networks (PIHNNs): Solving 2D linear elasticity prob- lems.Computer Methods in Applied Mechanics and Engineering, 432:117406, 2024. ISSN 0045-7825. doi: 10.1016/j.cma.2...

  24. [31]

    Solving plane crack prob- lems via enriched holomorphic neural networks.Engineering Fracture Mechanics, page 111133, 2025

    Matteo Calaf` a, Henrik Myhre Jensen, and Tito Andriollo. Solving plane crack prob- lems via enriched holomorphic neural networks.Engineering Fracture Mechanics, page 111133, 2025. doi: 10.1016/j.engfracmech.2025.111133. 25

  25. [32]

    Higher Mathematics Series

    Walter Rudin.Real and Complex Analysis. Higher Mathematics Series. McGraw-Hill, 3 edition, 1987. ISBN 0070542341

  26. [33]

    E. Goursat. Sur l’´ equation ∆∆u= 0.Bulletin de la Soci´ et´ e Math´ ematique de France, 26:236–237, 1898

  27. [34]

    Antony P. S. Selvadurai.Partial differential equations in mechanics 2: The Biharmonic equation, Poisson’s equation. Springer Science & Business Media, 2013. ISBN 978-3- 540-67284-5. doi: 10.1007/978-3-662-09205-7

  28. [35]

    Gould and Yuan Feng.Introduction to linear elasticity, volume 2

    Phillip L. Gould and Yuan Feng.Introduction to linear elasticity, volume 2. Springer,

  29. [36]

    Gury V. Kolosov. On an application of complex function theory to a plane problem of the mathematical theory of elasticity.Yuriev, Russia, 1909

  30. [37]

    Noordhoff Groningen, 1977

    Nikola ˘ ı Ivanovich Muskhelishvili.Some basic problems of the mathematical theory of elasticity. Noordhoff Groningen, 1977. ISBN 978-90-481-8245-9. doi: 10.1007/ 978-94-017-3034-1

  31. [38]

    An iterative method for the Helmholtz equation.Journal of Computational Physics, 49(3):443–457, 1983

    Alvin Bayliss, Charles I Goldstein, and Eli Turkel. An iterative method for the Helmholtz equation.Journal of Computational Physics, 49(3):443–457, 1983. doi: 10.1016/0021-9991(83)90139-0

  32. [39]

    Helmholtz equation–least-squares method for recon- structing the acoustic pressure field.The Journal of the Acoustical Society of America, 102(4):2020–2032, 1997

    Zhaoxi Wang and Sean F Wu. Helmholtz equation–least-squares method for recon- structing the acoustic pressure field.The Journal of the Acoustical Society of America, 102(4):2020–2032, 1997. doi: 10.1016/0898-1221(85)90162-2

  33. [40]

    A fast method for the solution of the Helmholtz equation.Journal of Computational Physics, 230(12):4403–4418, 2011

    Eldad Haber and Scott MacLachlan. A fast method for the solution of the Helmholtz equation.Journal of Computational Physics, 230(12):4403–4418, 2011. doi: 10.1016/j. jcp.2011.01.015

  34. [41]

    WORLD SCIENTIFIC, 1991

    Robert G Dean and Robert A Dalrymple.Water Wave Mechanics for Engineers and Scientists. WORLD SCIENTIFIC, 1991. doi: 10.1142/1232

  35. [42]

    Boundary integrated neural networks and code for acoustic radiation and scattering.International Journal of Mechanical System Dynamics, 4(2):131–141, 2024

    Wenzhen Qu, Yan Gu, Shengdong Zhao, Fajie Wang, and Ji Lin. Boundary integrated neural networks and code for acoustic radiation and scattering.International Journal of Mechanical System Dynamics, 4(2):131–141, 2024. doi: 10.1002/msd2.12109

  36. [43]

    Boundary integral neural networks for acoustic radiation prediction from noisy boundary data

    Johannes D Schmid, Simone Preuss, Lukas Maicher, and Steffen Marburg. Boundary integral neural networks for acoustic radiation prediction from noisy boundary data. Journal of Theoretical and Computational Acoustics, 2440002:36, 2025. doi: 10.1142/ S2591728524400024

  37. [45]

    I. N. Vekua.New Methods for Solving Elliptic Equations. North-Holland Series in Applied Mathematics & Mechanics. Elsevier Science Publishing Co Inc., 1967. ISBN 9780720423532

  38. [46]

    A survey of i

    Peter Henrici. A survey of i. n. vekua’s theory of elliptic partial differential equations with analytic coefficients.Zeitschrift f¨ ur angewandte Mathematik und Physik ZAMP, 8 (3):169–203, 1957. doi: 10.1007/BF01600500

  39. [47]

    Anastassiou

    George A. Anastassiou. Univariate hyperbolic tangent neural network approximation. Mathematical and Computer Modelling, 53(5-6):1111–1132, 2011. doi: 10.1016/j.mcm. 2010.11.072

  40. [48]

    Convergence analysis of two-layer neural networks with ReLU activation.Advances in neural information processing systems, 30, 2017

    Yuanzhi Li and Yang Yuan. Convergence analysis of two-layer neural networks with ReLU activation.Advances in neural information processing systems, 30, 2017

  41. [49]

    The influence of the sigmoid function parameters on the speed of backpropagation learning

    Jun Han and Claudio Moraga. The influence of the sigmoid function parameters on the speed of backpropagation learning. InInternational workshop on artificial neural networks, pages 195–201. Springer, 1995. doi: 10.1007/3-540-59497-3 175

  42. [50]

    An extension of the back-propagation algorithm to complex numbers

    Tohru Nitta. An extension of the back-propagation algorithm to complex numbers. Neural Networks, 10(8):1391–1415, 1997. doi: 10.1016/S0893-6080(97)00036-1

  43. [51]

    Studies in Computational Intelligence

    Akira Hirose.Complex-Valued Neural Networks. Studies in Computational Intelligence. Springer, 2 edition, 2012. ISBN 978-3-642-27632-3. doi: 10.1007/978-3-642-27632-3

  44. [52]

    Deep complex networks.arXiv preprint, 2017

    Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subramanian, Joao Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio, and Christo- pher J Pal. Deep complex networks.arXiv preprint, 2017. doi: 10.48550/arXiv.1705. 09792

  45. [53]

    Adam: A method for stochastic optimization.arXiv preprint, 2014

    Diederik P Kingma. Adam: A method for stochastic optimization.arXiv preprint, 2014. doi: 10.48550/arXiv.1412.6980

  46. [54]

    On the limited memory BFGS method for large scale op- timization.Mathematical programming, 45(1):503–528, 1989

    Dong C Liu and Jorge Nocedal. On the limited memory BFGS method for large scale op- timization.Mathematical programming, 45(1):503–528, 1989. doi: 10.1007/BF01589116

  47. [55]

    Zur formalen theorie der funktionen von mehr komplexen ver¨ anderlichen.Mathematische Annalen, 97(1):357–375, 1927

    Wilhelm Wirtinger. Zur formalen theorie der funktionen von mehr komplexen ver¨ anderlichen.Mathematische Annalen, 97(1):357–375, 1927. doi: 10.1007/ BF01447872

  48. [56]

    Springer Nature Switzerland, Cham, 2024

    Josiah Park and Stephan Wojtowytsch.Qualitative Neural Network Approximation over R and C: Elementary Proofs for Analytic and Polynomial Activation, chapter 3, pages 31–64. Springer Nature Switzerland, Cham, 2024. ISBN 978-3-031-66497-7. doi: 10.1007/978-3-031-66497-7 3

  49. [57]

    Gentile, Mario Dagrada, Chul Lee, Seong-Hyok S

    Atiyo Ghosh, Antonio A. Gentile, Mario Dagrada, Chul Lee, Seong-Hyok S. Kim, Hyukgeun Cha, Yunjun Choi, Dongho Kim, Jeong-Il Kye, and Vincent E. Elfving. Harmonic neural networks. InInternational Conference on Machine Learning, pages 11340–11359. PMLR, 2023. doi: 10.5555/36184...

  50. [58]

    A review on weight initialization strategies for neural networks.Artificial intelligence review, 55(1):291–322,

    Meenal V Narkhede, Prashant P Bartakke, and Mukul S Sutaone. A review on weight initialization strategies for neural networks.Artificial intelligence review, 55(1):291–322,

  51. [59]

    Understanding the difficulty of training deep feed- forward neural networks

    Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feed- forward neural networks. InProceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010

  52. [60]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015. doi: 10. 1109/ICCV.2015.123

  53. [61]

    Importance of input data normalization for the applica- tion of neural networks to complex industrial problems.IEEE Transactions on nuclear science, 44(3):1464–1468, 1997

    Jorge Sola and Joaquin Sevilla. Importance of input data normalization for the applica- tion of neural networks to complex industrial problems.IEEE Transactions on nuclear science, 44(3):1464–1468, 1997. doi: 10.1109/23.589532

  54. [62]

    Anderson.Fracture mechanics: fundamentals and applications

    Ted L. Anderson.Fracture mechanics: fundamentals and applications. CRC press, 2017. ISBN 9781315370293. doi: 10.1201/9781315370293

  55. [63]

    Jagtap and George Em Karniadakis

    Ameya D. Jagtap and George Em Karniadakis. Extended physics-informed neural net- works (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations.Communications in Computa- tional Physics, 28(5), 11 2020. ...

  56. [64]

    Harmonic functions from a complex analysis viewpoint.The American mathematical monthly, 93(4):246–258, 1986

    Sheldon Axler. Harmonic functions from a complex analysis viewpoint.The American mathematical monthly, 93(4):246–258, 1986. doi: 10.1080/00029890.1986.11971799

  57. [65]

    Howie.Laurent Series and the Residue Theorem, chapter 8, pages 137–

    John M. Howie.Laurent Series and the Residue Theorem, chapter 8, pages 137–

  58. [66]

    PyTorch: An imperative style, high-performance deep learning library.Advances in neural infor- mation processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library.Advances in neural infor- mation processing systems, 32...

  59. [67]

    New development in FreeFem++.Journal of numerical mathematics, 20(3-4):251–266, 2012

    Fr´ ed´ eric Hecht. New development in FreeFem++.Journal of numerical mathematics, 20(3-4):251–266, 2012. doi: 10.1515/jnum-2012-0013

  60. [68]

    Zhiping Mao and Xuhui Meng. Physics-informed neural networks with residual/gradient-based adaptive sampling methods for solving partial differential equa- tions with sharp solutions.Applied Mathematics and Mechanics, 44(7):1069–1084, 2023. doi: 10.1007/s10483-023-2994-7

  61. [69]

    Separable physics-informed neural networks

    Junwoo Cho, Seungtae Nam, Hyunmo Yang, Seok-Bae Yun, Youngjoon Hong, and Eunbyung Park. Separable physics-informed neural networks. In A. Oh, T. Naumann, 28 A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural In- formation Processing Systems, volume 36...

  62. [70]

    DeepXDE: A deep learning library for solving differential equations.SIAM Review, 63(1):208–228, 2021

    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. doi: 10.1137/19M1274067

  63. [71]

    A comparison between multilayer perceptrons and Kolmogorov-Arnold networks for multi-task classification in sitting posture recognition

    David Carneros-Prado, Luis Caba˜ nero-G´ omez, Esperanza Johnson, Iv´ an Gonz´ alez, Jes´ us Fontecha, and Ram´ on Herv´ as. A comparison between multilayer perceptrons and Kolmogorov-Arnold networks for multi-task classification in sitting posture recognition. IEEE Access, 20...

  64. [72]

    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. doi: 10.1016/j.cma.2022.115671. 29

  65. [152]

    ISBN 978-1-4471-0027-0

    Springer London, London, 2003. ISBN 978-1-4471-0027-0. doi: 10.1007/ 978-1-4471-0027-0 8

  66. [1994]

    doi: 10.1007/978-1-4614-4833-4

    ISBN 978-1-4614-4832-7. doi: 10.1007/978-1-4614-4833-4

  67. [2011]

    doi: 10.1007/978-3-030-42227-1

  68. [2019]

    doi: 10.1016/j.jcp.2018.10.045

  69. [2021]

    doi: 10.1038/s42254-021-00314-5

  70. [2022]

    doi: 10.1007/s10462-021-10033-z

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.