pith. sign in

arxiv: 2509.12483 · v4 · submitted 2025-09-15 · 💻 cs.LG

Benchmarking Physics-Informed Neural Networks and Boundary Elements Methods for Wave Scattering

Oscar Rinc\'on-Cardeno , Gregorio P\'erez Bernal , Silvana Montoya Noguera , Nicol\'as Guar\'in-Zapata This is my paper

Pith reviewed 2026-05-18 15:42 UTC · model grok-4.3

classification 💻 cs.LG
keywords Physics-informed neural networksBoundary element methodHelmholtz equationWave scatteringComputational timing comparisonNumerical methods benchmark
0
0 comments X

The pith

For the same 2D wave-scattering problem, boundary-element assembly and solve finishes in 0.01 seconds while a tuned PINN takes 100 seconds to train, after which interior-point queries run 100 times faster than BEM.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets up an identical two-dimensional Helmholtz scattering problem and solves it once with boundary element discretization and once with a physics-informed neural network whose architecture and training schedule were selected by hyperparameter search. The search settled on three hidden layers of 25 neurons each, sine activations, and a learning rate of 0.01. At accuracy levels the authors judge comparable, the BEM matrix assembly plus direct solve required roughly 10 to the minus 2 seconds while PINN training required roughly 10 to the 2 seconds. Once training is complete, however, the network evaluates the scattered field at interior locations in 10 to the minus 2 seconds, two orders of magnitude quicker than repeated BEM interior-point evaluation. The study therefore supplies concrete wall-clock numbers that can guide method selection for future wave-propagation calculations.

Core claim

The authors establish that, for the chosen scattering configuration, a hyperparameter-optimized PINN reaches accuracy levels comparable to a standard BEM discretization, yet the BEM system assembles and solves in 10^{-2} seconds while the PINN requires 10^{2} seconds to train; once trained, the PINN evaluates the field at interior points in 10^{-2} seconds, two orders of magnitude faster than BEM interior-point evaluation.

What carries the argument

Side-by-side timing of a hyperparameter-tuned PINN (3 layers, 25 neurons, sine activation) against boundary-discretized BEM on the identical Helmholtz scattering problem, separating assembly, training, and per-query evaluation costs.

If this is right

  • BEM remains the faster choice when only one solution is required and matrix assembly dominates the cost.
  • A trained PINN becomes preferable whenever the same scattering problem must be queried at many interior points or for many nearby configurations.
  • The hyperparameter search overhead must be amortized over multiple uses before the PINN's evaluation speed advantage pays off.
  • The reported timing ratios supply quantitative expectations that can be tested on other wave numbers or boundary shapes.
  • The direct comparison procedure itself can be repeated on three-dimensional or time-domain problems to map broader performance regimes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the accuracy equivalence survives stricter error measures, PINNs could serve as fast surrogate models inside real-time optimization or inverse-scattering loops after one-time training.
  • Repeating the benchmark at higher wave numbers would reveal whether the evaluation-speed advantage grows or shrinks as oscillations increase.
  • A hybrid workflow that uses BEM to generate training data or to initialize network weights could reduce the 100-second training penalty.
  • The sine activation may be specially suited to oscillatory solutions; swapping it for other activations on the same problem would test robustness of the timing result.

Load-bearing premise

The tuned PINN produces solution accuracy comparable to the BEM discretization without the paper specifying the exact error norms, convergence tolerances, or validation against known solutions used to declare the accuracies comparable.

What would settle it

Recompute both solutions for the identical scattering geometry, measure the L2 or maximum-norm error of each against an independent high-accuracy reference, and verify whether the reported four-order training gap and two-order evaluation advantage persist at equal error thresholds.

read the original abstract

This study compares the Boundary Element Method (BEM) and Physics-Informed Neural Networks (PINNs) for solving the two-dimensional Helmholtz equation in wave scattering problems. The objective is to evaluate the performance of both methods under the same conditions. We solve the Helmholtz equation using BEM and PINNs for the same scattering problem. PINNs are trained by minimizing the residual of the governing equations and boundary conditions with their configuration determined through hyperparameter optimization, while BEM is applied using boundary discretization. Both methods are evaluated in terms of solution accuracy and computation time. We conducted numerical experiments by varying the number of boundary integration points for the BEM and the number of hidden layers and neurons per layer for the PINNs. We performed a hyperparameter tuning to identify an adequate PINN configuration for this problem as a network with 3 hidden layers and 25 neurons per layer, using a learning rate of $10^{-2}$ and a sine activation function. At comparable levels of accuracy, the assembly and solution of the BEM system required a computational time on the order of $10^{-2}$~s, whereas the training time of the PINN was on the order of $10^{2}$~s, corresponding to a difference of approximately four orders of magnitude. However, once trained, the PINN achieved evaluation times on the order of $10^{-2}$~s, which is about two orders of magnitude faster than the evaluation of the BEM solution at interior points. This work establishes a procedure for comparing BEM and PINNs. It also presents a direct comparison between the two methods for the scattering problem. The analysis provides quantitative data on their performance, supporting their use in future research on wave propagation problems and outlining challenges and directions for further investigation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper benchmarks the Boundary Element Method (BEM) and Physics-Informed Neural Networks (PINNs) for the 2D Helmholtz equation in wave scattering. It tunes a PINN (3 hidden layers, 25 neurons per layer, sine activation, learning rate 10^{-2}), varies BEM boundary integration points and PINN architecture, and reports that at comparable accuracy BEM assembly/solution takes O(10^{-2}) s while PINN training takes O(10^2) s, but PINN inference is O(10^{-2}) s and two orders of magnitude faster than BEM interior-point evaluation.

Significance. If the accuracy equivalence holds under quantitative verification, the work supplies useful practical data on training-versus-inference trade-offs for wave-scattering problems and demonstrates direct wall-clock timing on independent implementations, which is a methodological strength.

major comments (1)
  1. Abstract: the claim that the two methods achieve 'comparable levels of accuracy' is unsupported by any reported error norm (L2 residual, boundary L2, pointwise maximum, etc.), numerical error magnitudes, or convergence tolerances. Without these quantities the headline timing comparison cannot be verified as comparing solutions of equivalent quality.
minor comments (1)
  1. The numerical-experiments description would benefit from an explicit table or figure listing achieved errors versus number of BEM points and versus PINN layer/neuron counts so that the 'comparable accuracy' statement can be directly inspected.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback on our manuscript. The single major comment is addressed point-by-point below, and we will revise the manuscript accordingly to strengthen the presentation of accuracy metrics.

read point-by-point responses

  1. Referee: Abstract: the claim that the two methods achieve 'comparable levels of accuracy' is unsupported by any reported error norm (L2 residual, boundary L2, pointwise maximum, etc.), numerical error magnitudes, or convergence tolerances. Without these quantities the headline timing comparison cannot be verified as comparing solutions of equivalent quality.

    Authors: We agree that the abstract claim requires explicit quantitative support to allow verification of the timing comparison. While the manuscript body describes the PINN training via residual minimization of the Helmholtz equation and boundary conditions, and BEM via boundary discretization, with accuracy assessed through solution plots and residual behavior, we did not report specific numerical error norms (such as L2 boundary error or maximum pointwise interior error) in the abstract itself. In the revised manuscript we will add these metrics to the abstract, reporting the achieved values for the tuned configurations (3 hidden layers, 25 neurons, sine activation for PINN; varying integration points for BEM). These will confirm that both methods reach comparable accuracy levels (typically on the order of 10^{-3} or better) before the reported wall-clock timings are compared. A supporting table of error norms versus discretization parameters will also be added to the results section. revision: yes

Circularity Check

0 steps flagged

No circularity: direct empirical timings from independent solver runs

full rationale

The paper reports measured wall-clock times for BEM matrix assembly/solution versus PINN training and inference on the same 2D Helmholtz scattering problem. These timings come from separate numerical experiments in which BEM integration points and PINN architecture (layers/neurons) are varied until accuracies are declared comparable; the reported orders of magnitude (10^{-2} s for BEM, 10^2 s training / 10^{-2} s inference for PINN) are therefore direct observations rather than quantities derived from any internal equation or fitted parameter. No derivation chain exists that reduces a claimed result to its own inputs by construction, and the central comparison rests on external, reproducible runtime measurements rather than self-citation or ansatz smuggling.

Axiom & Free-Parameter Ledger

4 free parameters · 1 axioms · 0 invented entities

The comparison rests on the standard Helmholtz governing equation and the assumption that a hyperparameter search yields a representative PINN; no new physical entities are postulated and the only free parameters are the network hyperparameters chosen by tuning.

free parameters (4)
  • number of hidden layers = 3
    Determined by hyperparameter optimization to reach adequate accuracy for the scattering problem.
  • neurons per hidden layer = 25
    Selected via tuning as part of the final PINN configuration.
  • learning rate = 10^{-2}
    Chosen during hyperparameter search for training stability and speed.
  • activation function = sine
    Identified by tuning as sine for this wave problem.
axioms (1)
  • domain assumption The two-dimensional Helmholtz equation governs time-harmonic acoustic or electromagnetic scattering.
    Invoked as the governing PDE whose residual is minimized by the PINN and discretized by BEM.

pith-pipeline@v0.9.0 · 5877 in / 1598 out tokens · 83588 ms · 2026-05-18T15:42:57.056805+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    We conducted numerical experiments by varying the number of boundary integration points for the BEM and the number of hidden layers and neurons per layer for the PINNs... At comparable levels of accuracy, the assembly and solution of the BEM system required a computational time on the order of 10^{-2} s, whereas the training time of the PINN was on the order of 10^{2} s

  • IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    The analytical solution... usct(r, θ) = sum Fn H(1)n(kr) e^{inθ} with Fn determined by Neumann condition on circular obstacle

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    Berm´ udez, L

    A. Berm´ udez, L. Hervella-Nieto, A. Prieto, and R. Rodrı´guez. An optimal perfectly matched layer with unbounded absorbing function for time-harmonic acoustic scattering prob- lems.Journal of Computational Physics, 223(2):469–488, May 2007. ISSN 0021-9991. doi: 10.1016/j.jcp.2006.09.018. URLhttps://www.sciencedirect.com/science/article/pii/ S0021999106004487

    work page doi:10.1016/j.jcp.2006.09.018 2007

  2. [2]

    Chandler-Wilde, Ivan G

    Simon N. Chandler-Wilde, Ivan G. Graham, Stephen Langdon, and Euan A. Spence. Numerical-asymptotic boundary integral methods in high-frequency acoustic scattering.Acta Numerica, 21:89–305, May 2012. ISSN 1474-0508, 0962-4929. doi: 10.1017/S0962492912000

    work page doi:10.1017/s0962492912000 2012

  3. [3]

    URLhttps://www.cambridge.org/core/journals/acta-numerica/article/abs/nu mericalasymptotic-boundary-integral-methods-in-highfrequency-acoustic-scatter ing/5FDA12EF93DD25E89AF7B57A01F8CAEE

    work page

  4. [4]

    Deep learning methods for partial differen- tial equations and related parameter identification problems.Inverse Problems, 39(10): 103001, August 2023

    Derick Nganyu Tanyu, Jianfeng Ning, Tom Freudenberg, Nick Heilenk¨ otter, Andreas Rademacher, Uwe Iben, and Peter Maass. Deep learning methods for partial differen- tial equations and related parameter identification problems.Inverse Problems, 39(10): 103001, August 2023. ISSN 0266-5611. doi: 10.1088/1361-6420/ace9d4. URLhttps: //dx.doi.org/10.1088/1361-6...

    work page doi:10.1088/1361-6420/ace9d4 2023

  5. [5]

    Lagaris, A

    I.E. Lagaris, A. Likas, and D.I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations.IEEE Transactions on Neural Networks, 9(5):987–1000, Septem- ber 1998. ISSN 1941-0093. doi: 10.1109/72.712178. URLhttps://ieeexplore.ieee.org/ document/712178. Conference Name: IEEE Transactions on Neural Networks

    work page doi:10.1109/72.712178 1998

  6. [6]

    M. W. M. G. Dissanayake and N. Phan-Thien. Neural-network-based approximations for solving partial differential equations.Communications in Numerical Methods in Engineer- ing, 10(3):195–201, 1994. ISSN 1099-0887. doi: 10.1002/cnm.1640100303. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/cnm.1640100303. eprint: https://onlinelibrary.wiley.com/doi/pdf...

    work page doi:10.1002/cnm.1640100303 1994

  7. [7]

    Journal of Computational Physics 378, 686–707

    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 dif- ferential equations.Journal of Computational Physics, 378:686–707, February 2019. doi: 10.1016/J.JCP.2018.10.045. Publisher: Academic Press

    work page doi:10.1016/j.jcp.2018.10.045 2019

  8. [8]

    Paul Escapil-Inchausp´ e and Gonzalo A. Ruz. Hyper-parameter tuning of physics-informed neural networks: Application to Helmholtz problems.Neurocomputing, 561:126826, December 18 /Acoustic Scattering by a Circular Obstacle in an Infinite Domain

    work page

  9. [9]

    doi: 10.1016/j.neucom.2023.126826

    ISSN 0925-2312. doi: 10.1016/j.neucom.2023.126826. URLhttps://www.sciencedir ect.com/science/article/pii/S0925231223009499

    work page doi:10.1016/j.neucom.2023.126826 2023

  10. [10]

    Walsh, Greg Pickrell, and Fabio Semperlotti

    Siddharth Nair, Timothy F. Walsh, Greg Pickrell, and Fabio Semperlotti. Multiple scattering simulation via physics-informed neural networks.Engineering with Computers, July 2024. ISSN 1435-5663. doi: 10.1007/s00366-024-02038-3. URLhttps://doi.org/10.1007/s00366 -024-02038-3

    work page doi:10.1007/s00366-024-02038-3 2024

  11. [11]

    Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations, July 2024

    Nick McGreivy and Ammar Hakim. Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations, July 2024. URLhttp: //arxiv.org/abs/2407.07218. arXiv:2407.07218 [physics]

    work page arXiv 2024

  12. [12]

    On the generaliza- tion of PINNs outside the training domain and the hyperparameters influencing it.Neural Computing and Applications, 36(36):22677–22696, December 2024

    Andrea Bonfanti, Roberto Santana, Marco Ellero, and Babak Gholami. On the generaliza- tion of PINNs outside the training domain and the hyperparameters influencing it.Neural Computing and Applications, 36(36):22677–22696, December 2024. ISSN 1433-3058. doi: 10.1007/s00521-024-10178-2. URLhttps://doi.org/10.1007/s00521-024-10178-2

    work page doi:10.1007/s00521-024-10178-2 2024

  13. [13]

    A perfectly matched layer for the absorption of electromagnetic-waves , year =

    Jean-Pierre Berenger. A perfectly matched layer for the absorption of electromagnetic waves. Journal of Computational Physics, 114(2):185–200, October 1994. ISSN 0021-9991. doi: 10.1 006/jcph.1994.1159. URLhttps://www.sciencedirect.com/science/article/pii/S002 1999184711594

    work page arXiv 1994

  14. [14]

    An unsplit convolutional Perfectly Matched Layer improved at grazing incidence for the seismic wave equation.Geophysics, 72, September 2007

    Dimitri Komatitsch and Roland Martin. An unsplit convolutional Perfectly Matched Layer improved at grazing incidence for the seismic wave equation.Geophysics, 72, September 2007. doi: 10.1190/1.2757586

    work page doi:10.1190/1.2757586 2007

  15. [15]

    Princeton University Press,

    Philip McCord Morse and K Uno Ingard.Theoretical Acoustics. Princeton University Press,

    work page

  16. [16]

    ISBN 978-0-691-02401-1

    work page

  17. [17]

    Eighty years of Sommerfeld’s radiation condition.Historia Mathematica, 19 (4):385–401, November 1992

    Steven H Schot. Eighty years of Sommerfeld’s radiation condition.Historia Mathematica, 19 (4):385–401, November 1992. ISSN 0315-0860. doi: 10.1016/0315-0860(92)90004-U. URL https://www.sciencedirect.com/science/article/pii/031508609290004U

    work page doi:10.1016/0315-0860(92)90004-u 1992

  18. [18]

    Katsikadelis.The Boundary Element Method for Engineers and Scientists, Second Edition: Theory and Applications

    John T. Katsikadelis.The Boundary Element Method for Engineers and Scientists, Second Edition: Theory and Applications. Academic Press, London, 2nd edition edition, July 2016. ISBN 978-0-12-804493-3

    work page 2016

  19. [19]

    January 2007

    Stephen Kirkup.The Boundary Element Method in Acoustics, volume 8. January 2007. ISBN 978-0-9534031-0-3. Journal Abbreviation: Journal of Computational Acoustics Publication Title: Journal of Computational Acoustics

    work page 2007

  20. [20]

    Grossmann, U

    Tamara G Grossmann, Urszula Julia Komorowska, Jonas Latz, and Carola-Bibiane Sch¨ onlieb. Can physics-informed neural networks beat the finite element method?IMA Journal of Applied Mathematics, 89(1):143–174, January 2024. ISSN 0272-4960. doi: 10.1093/imamat/hxae011. URLhttps://doi.org/10.1093/imamat/hxae011

    work page doi:10.1093/imamat/hxae011 2024

  21. [21]

    A Rand Corporation Research Study

    Yih-Hsing Pao and Chao-Chow Mow.Diffraction of Elastic Waves and Dynamic Stress Con- centrations. A Rand Corporation Research Study. Russak, New York, 1973. ISBN 978-0- 85274-244-0 978-0-8448-0155-1

    work page 1973

  22. [22]

    URLhttps://dlmf.nist.gov/

    DLMF.NIST Digital Library of Mathematical Functions.https://dlmf.nist.gov/, Release 1.2.4 of 2025-03-15. URLhttps://dlmf.nist.gov/. F. W. J. Olver, A. B. Olde Daalhuis, Conclusions/ 19 D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds

    work page 2025

  23. [23]

    Monika Nagy-Huber and Volker Roth. Physics-informed boundary integral networks (PIBI- Nets): A data-driven approach for solving partial differential equations.Journal of Computa- tional Science, 81:102355, September 2024. ISSN 1877-7503. doi: 10.1016/j.jocs.2024.102355. URLhttps://www.sciencedirect.com/science/article/pii/S1877750324001480

    work page doi:10.1016/j.jocs.2024.102355 2024

  24. [24]

    Boundary integrated neural networks and code for acoustic radiation and scattering.International Journal of Mechan- ical 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 Mechan- ical System Dynamics, 4(2):131–141, 2024. ISSN 2767-1402. doi: 10.1002/msd2.12109. URLhttps://onlinelibrary.wiley.com/doi/abs/10.1002/msd2.12109. eprint: https://onlinelibrary.wiley...

    work page doi:10.1002/msd2.12109 2024