pith. sign in

arxiv: 2605.20514 · v1 · pith:IM4YTYYYnew · submitted 2026-05-19 · 💻 cs.LG · cs.NA· math.NA· stat.ML

Fast Reconstruction of Exact Maxwell Dynamics from Sparse Data

Dan DeGenaro , Xin Li , Obed Amo , Michael Pokojovy , Sarah Adel Bargal , Markus Lange-Hegermann , Bogdan Rai\c{t}\u{a} This is my paper

Pith reviewed 2026-05-21 07:13 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NAstat.ML
keywords Maxwell equationsneural networkselectromagnetic fieldssparse reconstructionphysics-informed learningexact solutionsuniversal approximation
0
0 comments X

The pith

A neural network architecture for homogeneous electromagnetic fields embeds exact Maxwell solutions in each neuron to satisfy the equations by construction.

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

This paper introduces a shallow neural network called FLASH-MAX designed specifically for reconstructing homogeneous electromagnetic fields. Each neuron in the hidden layer corresponds to an exact solution of Maxwell's equations, making the entire network obey these equations symbolically without any penalty terms in the loss. The authors demonstrate that this setup allows end-to-end training from sparse point observations in just seconds while keeping the PDE residual at zero. They also establish a universal approximation theorem for this model class on any domain. Such an approach could change how physical constraints are incorporated into machine learning models for scientific applications by moving them into the architecture rather than the training objective.

Core claim

FLASH-MAX is a shallow exact-by-construction neural network for predicting homogeneous electromagnetic fields from sparse pointwise observations. Each hidden neuron represents a separate exact solution to Maxwell's equations, ensuring the network satisfies the governing equations symbolically by construction. A universal approximation result is proven, showing that this exact model class remains universal on arbitrary domains. The network reaches sub-1% relative validation error from about 1K sparse observations in seconds while maintaining a zero PDE residual, and keeps single-digit errors even for only 100 observations sampled from 3D space.

What carries the argument

The exact-by-construction hidden neurons, each representing an independent analytic solution to the homogeneous Maxwell equations, which enforce the physics directly in the network structure.

If this is right

  • Training completes in seconds rather than requiring extensive optimization.
  • The PDE residual remains exactly zero throughout and after training.
  • Low error is achieved with as few as 100 sparse observations in three-dimensional space.
  • The model class is universal, capable of approximating any homogeneous electromagnetic field on arbitrary domains.
  • Sub-1 percent relative error is attainable with approximately one thousand pointwise observations.

Where Pith is reading between the lines

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

  • Architectures of this type may be adapted for other systems of linear partial differential equations that admit known families of exact solutions.
  • Real-time applications in electromagnetic sensing or imaging could benefit from the rapid training and exact compliance with the governing laws.
  • The separation of the physical structure into the hypothesis class rather than the loss function offers a template for improving efficiency in other physics-constrained learning problems.
  • Further work might explore how the number of neurons scales with the complexity of the field or the sparsity of the data.

Load-bearing premise

The fields being modeled must be homogeneous, without sources or material variations that would prevent the exact solutions from spanning the necessary function space.

What would settle it

Training the model on 100 random points sampled from a known homogeneous solution such as a plane wave and checking whether the output matches the true field to single-digit percent relative error while the computed curl, divergence, and other Maxwell residuals remain at machine zero.

Figures

Figures reproduced from arXiv: 2605.20514 by Bogdan Rai\c{t}\u{a}, Dan DeGenaro, Markus Lange-Hegermann, Michael Pokojovy, Obed Amo, Sarah Adel Bargal, Xin Li.

Figure 1
Figure 1. Figure 1: Sparse-data Maxwell Reconstruction with Exact Hidden-unit Structure. From sparse pointwise observations of a single field instance, FLASH-MAX fits a shallow trainable architecture whose hidden neurons are exact Maxwell solutions, so the predicted field solves Maxwell’s system throughout optimization. The resulting model is exact by construction, expressive on arbitrary domains, and yields accurate reconstr… view at source ↗
Figure 2
Figure 2. Figure 2: Experiment 2: Performance on a Time Budget (BC setup). The curves depict the achievable error rates (y-axis) given a specific time budget (x-axis) for the setting with boundary conditions. Uncertainties are shown (standard error of the mean over 5 random seeds). FLASH-MAX (ours, blue solid line) clearly converges significantly faster than S-EPGP across all four problems. FEM and PINN are excluded since the… view at source ↗
Figure 3
Figure 3. Figure 3: Experiment 2: Performance on a Time Budget (IC setup). The curves depict the achievable error rates (y-axis) given a specific time budget (x-axis) for the setting with initial conditions. Uncertainties are shown (standard error of the mean over 5 random seeds). FLASH-MAX (blue solid line) converges more reliably and usually significantly faster than existing methods across all four problems: Plane waves, R… view at source ↗
Figure 4
Figure 4. Figure 4: Graphical representation of the forward pass [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Time evolution of the magnitude of the electric field for the [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Time evolution of the magnitude of the electric field for the [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Planar streamlines of the Hopf Fibration solution. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Random Solution. For the last solution we add 100 plane waves similar to those used in the Plane Waves solution. The direction and shifts of these plane waves are chosen randomly as we explain below. This ensures that the solutions we obtain are indistinguishable from an arbitrary solution of (MAX) [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Time evolution of the magnitude of the electric field for the [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
read the original abstract

We introduce FLASH-MAX, a shallow, exact-by-construction neural network architecture for predicting homogeneous electromagnetic fields from sparse pointwise observations. Each hidden neuron represents a separate exact solution to Maxwell's equations, so that the network satisfies the governing equations symbolically by construction and can be trained end-to-end from sparse data within seconds. We prove a universal approximation result showing that this exact model class remains universal on arbitrary domains. FLASH-MAX reaches sub-1% relative validation error from about 1K sparse pointwise observations in seconds, all while maintaining a zero PDE residual, and keeps single-digit errors even for only 100 observations sampled from 3D space. These results suggest that moving governing structure from the loss into the hypothesis class can dramatically improve the trade-off between precision and optimization speed in scientific machine learning.

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

2 major / 2 minor

Summary. The manuscript introduces FLASH-MAX, a shallow neural network architecture for reconstructing homogeneous electromagnetic fields from sparse pointwise observations. Each hidden neuron is constructed to represent an exact solution to the homogeneous Maxwell equations, so the network satisfies the governing PDEs symbolically by construction rather than through the loss. The authors report that the model can be trained end-to-end from roughly 1K sparse observations in seconds while maintaining zero PDE residual, achieves sub-1% relative validation error, and remains accurate even with only 100 observations. They also claim a universal approximation theorem establishing that this exact-by-construction model class is dense on arbitrary domains.

Significance. If the universal approximation result is rigorously established and the exactness property holds under training, the work would meaningfully advance physics-informed machine learning by moving governing equations from the loss into the hypothesis class. This could yield faster optimization and guaranteed physical consistency for sparse-data reconstruction tasks in electromagnetics. The reported performance metrics on limited observations are practically relevant, though the restriction to homogeneous source-free fields narrows the immediate applicability.

major comments (2)
  1. [Universal approximation theorem] Universal approximation theorem (abstract and corresponding proof section): The claim that the exact model class remains universal on arbitrary domains is load-bearing for the central contribution. For homogeneous Maxwell equations any linear combination of exact solutions is exact, yet density in the target space on a general bounded domain requires the parameterized family of solutions used in the hidden neurons to be complete with respect to the boundary conditions and topology. The manuscript must explicitly identify the solution family (e.g., plane waves, spherical harmonics, or eigenfunction expansions) and supply the completeness argument; without it the universality statement does not follow from the architectural design alone.
  2. [Results and experiments] Experimental validation (results section): The reported sub-1% relative error and zero residual are presented without baseline comparisons to standard PINNs or other exact-solution embeddings, without error bars across random seeds or observation samplings, and without ablation on the number of hidden neurons. These controls are necessary to substantiate that the speed and precision gains are attributable to the exact-by-construction property rather than to the specific test cases chosen.
minor comments (2)
  1. [Introduction and methods] Notation for the electromagnetic fields and the precise parameterization inside each hidden neuron should be introduced earlier and used consistently to aid readability.
  2. [Abstract] The abstract states 'sub-1% relative validation error' but does not specify the norm (L2, L-infinity, etc.) or the validation sampling strategy; this detail belongs in the main text as well.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and have revised the manuscript to strengthen the presentation of the universal approximation result and the experimental validation.

read point-by-point responses

  1. Referee: [Universal approximation theorem] Universal approximation theorem (abstract and corresponding proof section): The claim that the exact model class remains universal on arbitrary domains is load-bearing for the central contribution. For homogeneous Maxwell equations any linear combination of exact solutions is exact, yet density in the target space on a general bounded domain requires the parameterized family of solutions used in the hidden neurons to be complete with respect to the boundary conditions and topology. The manuscript must explicitly identify the solution family (e.g., plane waves, spherical harmonics, or eigenfunction expansions) and supply the completeness argument; without it the universality statement does not follow from the architectural design alone.

    Authors: We agree that the proof requires greater explicitness to be fully rigorous. The original manuscript states that the model class is universal but does not name the specific family or detail the completeness argument in the main text. In the revision we explicitly identify the hidden neurons as parameterized plane-wave solutions (with learnable wave vectors, frequencies, and polarizations) to the source-free homogeneous Maxwell equations. We have expanded Section 4 to include a self-contained completeness argument: these plane waves are dense in the space of smooth, divergence-free vector fields on bounded Lipschitz domains (invoking the known completeness of plane-wave expansions for the Helmholtz equation together with the divergence-free constraint). The revised proof now directly shows that finite linear combinations can approximate any target field in the appropriate Sobolev norm, thereby establishing universality on arbitrary domains. revision: yes

  2. Referee: [Results and experiments] Experimental validation (results section): The reported sub-1% relative error and zero residual are presented without baseline comparisons to standard PINNs or other exact-solution embeddings, without error bars across random seeds or observation samplings, and without ablation on the number of hidden neurons. These controls are necessary to substantiate that the speed and precision gains are attributable to the exact-by-construction property rather than to the specific test cases chosen.

    Authors: We concur that additional controls are needed to isolate the benefit of the exact-by-construction architecture. The revised manuscript now includes direct comparisons against a standard PINN baseline and a generic neural network without the Maxwell embedding, all trained on identical sparse observation sets. We report mean and standard deviation of relative error over 10 independent random seeds for both observation sampling and weight initialization. We have also added an ablation study that varies the number of hidden neurons from 20 to 400 while keeping all other factors fixed; the results confirm that zero residual is maintained for any neuron count and that validation error plateaus beyond approximately 80 neurons. These new experiments appear in an expanded Section 5 together with the corresponding tables and figures. revision: yes

Circularity Check

0 steps flagged

No significant circularity; exactness is architectural by design

full rationale

The paper's central claim is that the network satisfies Maxwell's equations exactly by construction because each hidden neuron is defined to be an exact solution, with the PDE residual therefore identically zero without any parameter fitting or loss term enforcing it. This is a direct architectural choice rather than a derived or fitted result, and the universal approximation statement is presented as a separate theorem for the resulting model class. No equations or steps in the provided abstract reduce a prediction or theorem back to its own inputs by definition, nor do they rely on self-citations whose content is unverified or load-bearing for the exactness property. The derivation chain is therefore self-contained against external benchmarks of PDE satisfaction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that the target fields are homogeneous Maxwell solutions and on the mathematical fact that linear combinations of exact solutions remain exact. No free parameters or invented entities are introduced beyond the standard neural-network weights.

axioms (1)
  • domain assumption The electromagnetic fields are homogeneous and satisfy the source-free Maxwell equations.
    The abstract explicitly limits the method to homogeneous electromagnetic fields whose exact solutions form the neuron basis.

pith-pipeline@v0.9.0 · 5685 in / 1331 out tokens · 44989 ms · 2026-05-21T07:13:07.599144+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.

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

80 extracted references · 80 canonical work pages

  1. [1]

    Raissi, P

    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, 2019. doi: 10.1016/j.jcp.2018.10.045

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

  2. [2]

    Gaussian Process Priors for Systems of Linear Partial Differential Equations with Constant Coefficients

    Marc Härkönen, Markus Lange-Hegermann, and Bogdan Rai¸ t˘a. Gaussian Process Priors for Systems of Linear Partial Differential Equations with Constant Coefficients. InProceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 12587–12615. PMLR, 2023. URLhttps://proceedings.mlr.press/v...

    work page 2023

  3. [3]

    Noetherian Operators in Macaulay2.Journal of Software for Algebra and Geometry, 12(1):33–41, 2022

    Justin Chen, Yairon Cid-Ruiz, Marc Härkönen, Robert Krone, and Anton Leykin. Noetherian Operators in Macaulay2.Journal of Software for Algebra and Geometry, 12(1):33–41, 2022. doi: 10.2140/jsag.2022. 12.33

    work page doi:10.2140/jsag.2022 2022

  4. [4]

    The Construction of Noetherian Operators.Journal of Algebra, 222(2):595–620, 1999

    Ulrich Oberst. The Construction of Noetherian Operators.Journal of Algebra, 222(2):595–620, 1999. doi: 10.1006/jabr.1999.8035

    work page doi:10.1006/jabr.1999.8035 1999

  5. [5]

    Alberto Damiano, Irene Sabadini, and Daniele C. Struppa. Computational Methods for the Construction of a Class of Noetherian Operators.Experimental Mathematics, 16(1):41–53, 2007. doi: 10.1080/10586458. 2007.10128986

    work page doi:10.1080/10586458 2007

  6. [6]

    Effective Algorithm for Computing Noetherian Operators of Zero-Dimensional Ideals.Applicable Algebra in Engineering, Communication and Computing, 33(6): 867–899, 2022

    Katsusuke Nabeshima and Shinichi Tajima. Effective Algorithm for Computing Noetherian Operators of Zero-Dimensional Ideals.Applicable Algebra in Engineering, Communication and Computing, 33(6): 867–899, 2022. doi: 10.1007/s00200-022-00570-7

    work page doi:10.1007/s00200-022-00570-7 2022

  7. [7]

    Grayson and Michael E

    Daniel R. Grayson and Michael E. Stillman. Macaulay2, a Software System for Research in Algebraic Geometry, 2023. URLhttps://macaulay2.com

    work page 2023

  8. [8]

    Linear PDE with Constant Coefficients

    Rida Ait El Manssour, Marc Härkönen, and Bernd Sturmfels. Linear PDE with Constant Coefficients. Glasgow Mathematical Journal, 65(S1):S2–S27, 2023. doi: 10.1017/S0017089521000355

    work page doi:10.1017/s0017089521000355 2023

  9. [9]

    Primary Decomposition with Differential Operators.International Mathematics Research Notices, 2023(14):12119–12147, 2023

    Yairon Cid-Ruiz and Bernd Sturmfels. Primary Decomposition with Differential Operators.International Mathematics Research Notices, 2023(14):12119–12147, 2023. doi: 10.1093/imrn/rnac178

    work page doi:10.1093/imrn/rnac178 2023

  10. [10]

    Primary Decomposition of Modules: A Computational Differential Approach.Journal of Pure and Applied Algebra, 226(10):107080, 2022

    Justin Chen and Yairon Cid-Ruiz. Primary Decomposition of Modules: A Computational Differential Approach.Journal of Pure and Applied Algebra, 226(10):107080, 2022

    work page 2022

  11. [11]

    Primary Ideals and Their Differential Equations

    Yairon Cid-Ruiz, Roser Homs, and Bernd Sturmfels. Primary Ideals and Their Differential Equations. Foundations of Computational Mathematics, 21(5):1363–1399, 2021. doi: 10.1007/s10208-020-09485-6

    work page doi:10.1007/s10208-020-09485-6 2021

  12. [12]

    Noetherian operators and primary decomposition.Journal of Symbolic Computation, 110:1–23, 2022

    Justin Chen, Marc Härkönen, Robert Krone, and Anton Leykin. Noetherian operators and primary decomposition.Journal of Symbolic Computation, 110:1–23, 2022. doi: 10.1016/j.jsc.2021.09.002

    work page doi:10.1016/j.jsc.2021.09.002 2022

  13. [13]

    Noetherian operators, primary submodules and symbolic powers.Collectanea Mathe- matica, 72:175–202, 2021

    Yairon Cid-Ruiz. Noetherian operators, primary submodules and symbolic powers.Collectanea Mathe- matica, 72:175–202, 2021. doi: 10.1007/s13348-020-00285-3

    work page doi:10.1007/s13348-020-00285-3 2021

  14. [14]

    Wiley, 1970

    Leon Ehrenpreis.Fourier analysis in several complex variables. Wiley, 1970

    work page 1970

  15. [15]

    V . P. Palamodov.Linear differential operators with constant coefficients. Springer, 1970

    work page 1970

  16. [16]

    North-Holland, 3 edition, 1990

    Lars Hörmander.An Introduction to Complex Analysis in Several Variables, volume 7 ofNorth-Holland Mathematical Library. North-Holland, 3 edition, 1990

    work page 1990

  17. [17]

    Björk.Rings of Differential Operators, volume 21 ofNorth-Holland Mathematical Library

    J.-E. Björk.Rings of Differential Operators, volume 21 ofNorth-Holland Mathematical Library. North- Holland, 1979. ISBN 9780444852922

    work page 1979

  18. [18]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. InAdvances in Neural Information Processing Systems, volume 33, 2020

    work page 2020

  19. [19]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. InAdvances in Neural Information Processing Systems, volume 33, 2020

    work page 2020

  20. [20]

    Random features for large-scale kernel machines

    Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. InAdvances in Neural Information Processing Systems, volume 20, 2007. 11

    work page 2007

  21. [21]

    Yoh-Han Pao, Gwang Hoon Park, and Dejan J. Sobajic. Learning and generalization characteristics of the random vector functional-link net.Neurocomputing, 6(2):163–180, 1994. doi: 10.1016/0925-2312(94) 90053-1

    work page doi:10.1016/0925-2312(94 1994

  22. [22]

    Extreme learning machine: Theory and applica- tions.Neurocomputing, 70(1–3):489–501, 2006

    Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: Theory and applica- tions.Neurocomputing, 70(1–3):489–501, 2006. doi: 10.1016/j.neucom.2005.12.126

    work page doi:10.1016/j.neucom.2005.12.126 2006

  23. [23]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. InProceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 5301–5310. PMLR, 2019

    work page 2019

  24. [24]

    Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

    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. doi: 10.1137/20M1318043

    work page doi:10.1137/20m1318043 2021

  25. [25]

    Krishnapriyan, Amir Gholami, Shandian Zhe, Robert M

    Aditi S. Krishnapriyan, Amir Gholami, Shandian Zhe, Robert M. Kirby, and Michael W. Mahoney. Characterizing Possible Failure Modes in Physics-Informed Neural Networks. InAdvances in Neural Information Processing Systems, volume 34, pages 26548–26560, 2021. URL https://proceedings. neurips.cc/paper/2021/hash/df438e5206f31600e6ae4af72f2725f1-Abstract.html

    work page 2021

  26. [26]

    A Maxwell’s Equations Based Deep Learning Method for Time Domain Electromagnetic Simulations.IEEE Journal on Multiscale and Multiphysics Computational Techniques, 6:35–40, 2021

    Pan Zhang, Yanyan Hu, Yuchen Jin, Shaogui Deng, Xuqing Wu, and Jiefu Chen. A Maxwell’s Equations Based Deep Learning Method for Time Domain Electromagnetic Simulations.IEEE Journal on Multiscale and Multiphysics Computational Techniques, 6:35–40, 2021

    work page 2021

  27. [27]

    Castonguay

    Pratanu Roy and Stephen T. Castonguay. Exact Enforcement of Temporal Continuity in Sequential Physics- Informed Neural Networks.Computer Methods in Applied Mechanics and Engineering, 430:117197, 2024. doi: 10.1016/j.cma.2024.117197

    work page doi:10.1016/j.cma.2024.117197 2024

  28. [28]

    Shiyuan Piao, Hong Gu, Aina Wang, and Pan Qin. A Domain-Adaptive Physics-Informed Neural Network for Inverse Problems of Maxwell’s Equations in Heterogeneous Media.IEEE Antennas and Wireless Propagation Letters, 23(10):2905–2909, 2024. doi: 10.1109/LAWP.2024.3413851

    work page doi:10.1109/lawp.2024.3413851 2024

  29. [29]

    TCAS-PINN: Physics-Informed Neural Networks with a Novel Temporal Causality-Based Adaptive Sampling Method.Chinese Physics B, 33(5):050701,

    Jia Guo, Haifeng Wang, Shilin Gu, and Chenping Hou. TCAS-PINN: Physics-Informed Neural Networks with a Novel Temporal Causality-Based Adaptive Sampling Method.Chinese Physics B, 33(5):050701,

    work page

  30. [30]

    doi: 10.1088/1674-1056/ad21f3

    work page doi:10.1088/1674-1056/ad21f3

  31. [31]

    Electromagnetic Scattering from a Three-Dimensional Object Using Physics-Informed Neural Network.Applied Computational Electromagnetics Society Journal, 40(2):103–111, feb 2025

    Yuan Zhang, Renxian Li, Huan Tang, Zhuoyuan Shi, Bing Wei, Shuhong Gong, Lixia Yang, and Bing Yan. Electromagnetic Scattering from a Three-Dimensional Object Using Physics-Informed Neural Network.Applied Computational Electromagnetics Society Journal, 40(2):103–111, feb 2025. doi: 10.13052/2025.ACES.J.400203

    work page doi:10.13052/2025.aces.j.400203 2025

  32. [32]

    Journal of Mathematical Physics64(9), 091902 (2023) https://doi.org/10.1063/5

    Yuyao Chen and Luca Dal Negro. Physics-Informed Neural Networks for Imaging and Parameter Retrieval of Photonic Nanostructures from Near-Field Data.APL Photonics, 7(1):010802, 2022. doi: 10.1063/5. 0072969

    work page doi:10.1063/5 2022

  33. [33]

    Y . Deng, K. Fan, B. Jin, J. Malof, and W. J. Padilla. Physics-Informed Learning in Artificial Electromagnetic Materials.Applied Physics Reviews, 12(1):011331, 2025. doi: 10.1063/5.0232675

    work page doi:10.1063/5.0232675 2025

  34. [34]

    Physics-Informed Neural Networks for the Numerical Modeling of Steady-State and Transient Electromagnetic Problems with Discontinuous Media.arXiv preprint arXiv:2406.04380, 2024

    Michel Nohra and Steven Dufour. Physics-Informed Neural Networks for the Numerical Modeling of Steady-State and Transient Electromagnetic Problems with Discontinuous Media.arXiv preprint arXiv:2406.04380, 2024

    work page arXiv 2024

  35. [35]

    Approximating Electromagnetic Fields in Discontinuous Media Using a Single Physics-Informed Neural Network.arXiv preprint arXiv:2407.20833, 2024

    Michel Nohra and Steven Dufour. Approximating Electromagnetic Fields in Discontinuous Media Using a Single Physics-Informed Neural Network.arXiv preprint arXiv:2407.20833, 2024

    work page arXiv 2024

  36. [36]

    Shaviner, Hemanth Chandravamsi, Shimon Pisnoy, Ziv Chen, and Steven H

    Gal G. Shaviner, Hemanth Chandravamsi, Shimon Pisnoy, Ziv Chen, and Steven H. Frankel. PINNs for Solving Unsteady Maxwell’s Equations: Convergence Issues and Comparative Assessment with Compact Schemes.Neural Computing and Applications, 37(29):24103–24122, 2025. doi: 10.1007/ s00521-025-11554-2

    work page 2025

  37. [37]

    Black Hole

    Ziv Chen, Gal G. Shaviner, Hemanth Chandravamsi, Shimon Pisnoy, Steven H. Frankel, and Uzi Pereg. Quantum Physics-Informed Neural Networks for Maxwell’s Equations: Circuit Design, “Black Hole” Barren Plateaus Mitigation, and GPU Acceleration.Quantum Machine Intelligence, 8(1):21, 2026. doi: 10.1007/s42484-026-00365-w

    work page doi:10.1007/s42484-026-00365-w 2026

  38. [38]

    Evolutionary Optimization of Physics-Informed Neural Networks: Evo-PINN Frontiers and Opportunities

    Jian Cheng Wong, Abhishek Gupta, Chin Chun Ooi, Pao-Hsiung Chiu, Jiao Liu, and Yew-Soon Ong. Evolutionary Optimization of Physics-Informed Neural Networks: Evo-PINN Frontiers and Opportunities. IEEE Computational Intelligence Magazine, 21(1):16–36, 2026. doi: 10.1109/MCI.2025.3607749. 12

    work page doi:10.1109/mci.2025.3607749 2026

  39. [39]

    Mixed Finite Elements inR 3.Numerische Mathematik, 35(3):315–341, 1980

    Jean-Claude Nédélec. Mixed Finite Elements inR 3.Numerische Mathematik, 35(3):315–341, 1980

    work page 1980

  40. [40]

    Hybridizable Discontinuous Galerkin Methods for the Time-Harmonic Maxwell Equations.Journal of Computational Physics, 230(19):7151– 7175, 2011

    Ngoc Cuong Nguyen, Jaime Peraire, and Bernardo Cockburn. Hybridizable Discontinuous Galerkin Methods for the Time-Harmonic Maxwell Equations.Journal of Computational Physics, 230(19):7151– 7175, 2011

    work page 2011

  41. [41]

    Arnold, Richard S

    Douglas N. Arnold, Richard S. Falk, and Ragnar Winther. Finite Element Exterior Calculus, Homological Techniques, and Applications.Acta Numerica, 15:1–155, 2006

    work page 2006

  42. [42]

    Sweeping Preconditioner for the Helmholtz Equation: Hierarchical Matrix Representation.Communications on Pure and Applied Mathematics, 64(5):697–735, 2011

    Björn Engquist and Lexing Ying. Sweeping Preconditioner for the Helmholtz Equation: Hierarchical Matrix Representation.Communications on Pure and Applied Mathematics, 64(5):697–735, 2011. doi: 10.1002/cpa.20358

    work page doi:10.1002/cpa.20358 2011

  43. [43]

    Birkhäuser Basel, Basel, 2003

    Wolfgang Bangerth and Rolf Rannacher.Adaptive Finite Element Methods for Differential Equations. Birkhäuser Basel, Basel, 2003

    work page 2003

  44. [44]

    Logg, K.-A

    Anders Logg, Kent-Andre Mardal, and Garth Wells, editors.Automated Solution of Differential Equa- tions by the Finite Element Method: The FEniCS Book, volume 84 ofLecture Notes in Computa- tional Science and Engineering. Springer Berlin, Heidelberg, 2012. ISBN 978-3-642-23099-8. doi: 10.1007/978-3-642-23099-8

    work page doi:10.1007/978-3-642-23099-8 2012

  45. [45]

    Lu , author P

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

    work page doi:10.1038/s42256-021-00302-5 2021

  46. [46]

    Stuart, and Anima Anandkumar

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew M. Stuart, and Anima Anandkumar. Fourier Neural Operator for Parametric Partial Differential Equations. InInternational Conference on Learning Representations (ICLR), 2021. URL https:// openreview.net/forum?id=c8P9NQVtmnO

    work page 2021

  47. [47]

    Physics-Informed Neural Operator for Learning Partial Differential Equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Aziz- zadenesheli, and Anima Anandkumar. Physics-Informed Neural Operator for Learning Partial Differential Equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024. doi: 10.1145/3648506

    work page doi:10.1145/3648506 2024

  48. [48]

    Neural Operators for Accelerating Scientific Simulations and Design.Nature Reviews Physics, 6:320–328, 2024

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandkumar. Neural Operators for Accelerating Scientific Simulations and Design.Nature Reviews Physics, 6:320–328, 2024. doi: 10.1038/s42254-024-00712-5

    work page doi:10.1038/s42254-024-00712-5 2024

  49. [49]

    Chen, Duane Boning, and David Z

    Jiaqi Gu, Zhengqi Gao, Chenghao Feng, Hanqing Zhu, Ray T. Chen, Duane Boning, and David Z. Pan. NeurOLight: A Physics-Agnostic Neural Operator Enabling Parametric Pho- tonic Device Simulation. InAdvances in Neural Information Processing Systems, vol- ume 35, 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/ 5ddfb189c022a317ff1c72e66390...

    work page 2022

  50. [50]

    Chen, Jiaqi Gu, and David Z

    Hanqing Zhu, Wenyan Cong, Guojin Chen, Shupeng Ning, Ray T. Chen, Jiaqi Gu, and David Z. Pan. PACE: Pacing Operator Learning to Accurate Optical Field Simulation for Complicated Photonic Devices. InAdvances in Neural Information Processing Sys- tems, volume 37, 2024. URL https://papers.nips.cc/paper_files/paper/2024/hash/ 7cc16e8635e6f27c295355bd214ef8d8-...

    work page 2024

  51. [51]

    Neural Operator-Based Surrogate Solver for Free-Form Electromagnetic Inverse Design.ACS Photonics, 10(5):1547–1557, 2023

    Yannick Augenstein, Taavi Repän, and Carsten Rockstuhl. Neural Operator-Based Surrogate Solver for Free-Form Electromagnetic Inverse Design.ACS Photonics, 10(5):1547–1557, 2023. doi: 10.1021/ acsphotonics.3c00156

    work page 2023

  52. [52]

    Sooyoung Oh, EungKyu Lee, and Sun K. Hong. Physics-Informed Deep Operator Networks for Modeling 2D Time-Domain Electromagnetic Wave Propagation in Various Media.iScience, 29(3):115076, 2026. doi: 10.1016/j.isci.2026.115076

    work page doi:10.1016/j.isci.2026.115076 2026

  53. [53]

    Accelerated Time-Domain Simulation of Complex Photonic Structures with a Data-Aware Fourier Neural Operator.Optics Communications, 604:132838,

    Zaifan Wu, Yue You, Xian Zhou, and Fang Zhang. Accelerated Time-Domain Simulation of Complex Photonic Structures with a Data-Aware Fourier Neural Operator.Optics Communications, 604:132838,

    work page

  54. [54]

    doi: 10.1016/j.optcom.2025.132838

    work page doi:10.1016/j.optcom.2025.132838 2025

  55. [55]

    Physics-Constrained Machine Learning for Electrodynamics Without Gauge Ambiguity Based on Fourier Transformed Maxwell’s Equations.Scientific Reports, 14: 14809, 2024

    Christopher Leon and Alexander Scheinker. Physics-Constrained Machine Learning for Electrodynamics Without Gauge Ambiguity Based on Fourier Transformed Maxwell’s Equations.Scientific Reports, 14: 14809, 2024. doi: 10.1038/s41598-024-65650-9

    work page doi:10.1038/s41598-024-65650-9 2024

  56. [56]

    Wilson, Yuheng Chen, Alexander V

    Vaishnavi Iyer, Blake A. Wilson, Yuheng Chen, Alexander V . Kildishev, Vladimir M. Shalaev, and Alexandra Boltasseva. Deep learning in photonic device development: nuances and opportunities.npj Nanophotonics, 3:5, 2026. doi: 10.1038/s44310-025-00097-y. 13

    work page doi:10.1038/s44310-025-00097-y 2026

  57. [57]

    Lawrence

    Mauricio Álvarez, David Luengo, and Neil D. Lawrence. Latent Force Models. InProceedings of the Twelfth International Conference on Artificial Intelligence and Statistics, volume 5 ofProceedings of Machine Learning Research, pages 9–16. PMLR, 2009. URL https://proceedings.mlr.press/v5/ alvarez09a.html

    work page 2009

  58. [58]

    Learning Divergence-Free and Curl-Free Vector Fields with Matrix-Valued Kernels

    Ives Macêdo and Rener Castro. Learning Divergence-Free and Curl-Free Vector Fields with Matrix-Valued Kernels. Technical report, Instituto de Matemática Pura e Aplicada (IMPA), 2010

    work page 2010

  59. [59]

    Schön, and Simo Särkkä

    Arno Solin, Manon Kok, Niklas Wahlström, Thomas B. Schön, and Simo Särkkä. Modeling and Inter- polation of the Ambient Magnetic Field by Gaussian Processes.IEEE Transactions on Robotics, 34(4): 1112–1127, 2018. doi: 10.1109/TRO.2018.2830323

    work page doi:10.1109/tro.2018.2830323 2018

  60. [60]

    Carl Jidling, Niklas Wahlström, Adrian Wills, and Thomas B. Schön. Linearly Con- strained Gaussian Processes. InAdvances in Neural Information Processing Systems 30, pages 1215–1224, 2017. URL https://proceedings.neurips.cc/paper/2017/hash/ 71ad16ad2c4d81f348082ff6c4b20768-Abstract.html

    work page 2017

  61. [61]

    Algorithmic Linearly Constrained Gaussian Processes

    Markus Lange-Hegermann. Algorithmic Linearly Constrained Gaussian Processes. InAdvances in Neural Information Processing Systems 31, pages 2141–2152, 2018. URL https://proceedings.neurips. cc/paper/2018/hash/68b1fbe7f16e4ae3024973f12f3cb313-Abstract.html

    work page 2018

  62. [62]

    Constraining Gaussian Processes to Systems of Linear Ordinary Differential Equations

    Andreas Besginow and Markus Lange-Hegermann. Constraining Gaussian Processes to Systems of Linear Ordinary Differential Equations. In Sanmi Koyejo, Shakir Mohamed, Anima Agarwal, Danielle Belgrave, Kyunghyun Cho, and Alice Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 29386–29399. Curran Associates, Inc., 2022. URL https:...

    work page 2022

  63. [63]

    Linearly Constrained Gaussian Processes with Boundary Conditions

    Markus Lange-Hegermann. Linearly Constrained Gaussian Processes with Boundary Conditions. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, volume 130 ofProceedings of Machine Learning Research, pages 1090–1098. PMLR, 2021. URL https: //proceedings.mlr.press/v130/lange-hegermann21a.html

    work page 2021

  64. [64]

    On Boundary Conditions Parametrized by Analytic Functions

    Markus Lange-Hegermann and Daniel Robertz. On Boundary Conditions Parametrized by Analytic Functions. In François Boulier, Matthew England, Timur M. Sadykov, and Evgenii V . V orozhtsov, editors, Computer Algebra in Scientific Computing, volume 13366 ofLecture Notes in Computer Science, pages 225–245. Springer, Cham, 2022. doi: 10.1007/978-3-031-14788-3_13

    work page doi:10.1007/978-3-031-14788-3_13 2022

  65. [65]

    David Dalton, Alan Lazarus, Hao Gao, and Dirk Husmeier. Boundary constrained Gaussian processes for robust physics-informed machine learning of linear partial differential equations.Journal of Machine Learning Research, 25(272):1–61, 2024. URLhttp://jmlr.org/papers/v25/23-1508.html

    work page 2024

  66. [67]

    URLhttps://arxiv.org/abs/2411.16663

    work page arXiv

  67. [68]

    Gaussian Process Regression for Inverse Problems in Linear PDEs.IFAC-PapersOnLine, 59(12):196–201, 2025

    Xin Li, Markus Lange-Hegermann, and Bogdan Rai¸ t˘a. Gaussian Process Regression for Inverse Problems in Linear PDEs.IFAC-PapersOnLine, 59(12):196–201, 2025

    work page 2025

  68. [69]

    Lagaris, Aristidis Likas, and Dimitrios I

    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

    work page doi:10.1109/72.712178 1998

  69. [70]

    Towards understanding the effect of leak in Spiking Neural Networks,

    Enrico Schiassi, Roberto Furfaro, Carl Leake, Mario De Florio, Hunter Johnston, and Daniele Mortari. Extreme Theory of Functional Connections: A Fast Physics-Informed Neural Network Method for Solving Ordinary and Partial Differential Equations.Neurocomputing, 457:334–356, 2021. doi: 10.1016/j.neucom. 2021.06.015

    work page doi:10.1016/j.neucom 2021

  70. [71]

    Sukumar and Ankit Srivastava

    N. Sukumar and Ankit Srivastava. Exact Imposition of Boundary Conditions with Distance Functions in Physics-Informed Deep Neural Networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022. doi: 10.1016/j.cma.2021.114333

    work page doi:10.1016/j.cma.2021.114333 2022

  71. [72]

    Ein Gegenstück zum Ritzschen Verfahren

    Erich Trefftz. Ein Gegenstück zum Ritzschen Verfahren. InProceedings of the 2nd International Congress of Applied Mechanics, pages 131–137, Zürich, 1926

    work page 1926

  72. [73]

    Bobaru, M

    Tomi Huttunen, Petri Malinen, and Peter Monk. Solving Maxwell’s Equations Using the Ultra Weak Variational Formulation.Journal of Computational Physics, 223(2):731–758, 2007. doi: 10.1016/j.jcp. 2006.10.016. 14

    work page doi:10.1016/j.jcp 2007

  73. [74]

    A Survey of Trefftz Methods for the Helmholtz Equation

    Ralf Hiptmair, Andrea Moiola, and Ilaria Perugia. A Survey of Trefftz Methods for the Helmholtz Equation. In Gabriel R. Barrenechea, Franco Brezzi, Andrea Cangiani, and Emmanuil H. Georgoulis, editors,Building Bridges: Connections and Challenges in Modern Approaches to Numerical Partial Differential Equations, volume 114 ofLecture Notes in Computational S...

    work page doi:10.1007/978-3-319-41640-3_8 2016

  74. [75]

    Physics-Constrained 3D Convolutional Neural Networks for Electrodynamics.APL Machine Learning, 1(2):026109, 2023

    Alexander Scheinker and Reeju Pokharel. Physics-Constrained 3D Convolutional Neural Networks for Electrodynamics.APL Machine Learning, 1(2):026109, 2023. doi: 10.1063/5.0132433

    work page doi:10.1063/5.0132433 2023

  75. [76]

    Lax.Hyperbolic Partial Differential Equations, volume 14 ofCourant Lecture Notes in Math- ematics

    Peter D. Lax.Hyperbolic Partial Differential Equations, volume 14 ofCourant Lecture Notes in Math- ematics. Courant Institute of Mathematical Sciences and American Mathematical Society, 2006. ISBN 9780821835760. doi: 10.1090/cln/014

    work page doi:10.1090/cln/014 2006

  76. [77]

    Oxford University Press, 2013

    Rüdiger Verfürth.A Posteriori Error Estimation Techniques for Finite Element Methods. Oxford University Press, 2013

    work page 2013

  77. [78]

    A posteriori error estimates for maxwell equations.Mathematics of Computation, 77 (262):633–649, 2008

    Joachim Schöberl. A posteriori error estimates for maxwell equations.Mathematics of Computation, 77 (262):633–649, 2008. doi: 10.1090/S0025-5718-07-02030-3

    work page doi:10.1090/s0025-5718-07-02030-3 2008

  78. [79]

    Chenkai Mao and Jonathan A. Fan. Accurate and scalable deep maxwell solvers.Proceedings of the National Academy of Sciences, 123(18):e2530330123, 2026. doi: 10.1073/pnas.2530330123. URL https://www.pnas.org/doi/abs/10.1073/pnas.2530330123

    work page doi:10.1073/pnas.2530330123 2026

  79. [80]

    Extension and representation of divergence- free vector fields on bounded domains.Mathematical Research Letters, 7(5–6):643–650, 2000

    Tosio Kato, Marius Mitrea, Gustavo Ponce, and Michael Taylor. Extension and representation of divergence- free vector fields on bounded domains.Mathematical Research Letters, 7(5–6):643–650, 2000

    work page 2000

  80. [81]

    Universal Approximation of an Unknown Mapping and Its Derivatives Using Multilayer Feedforward Networks.Neural Networks, 3(5):551–560, 1990

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Universal Approximation of an Unknown Mapping and Its Derivatives Using Multilayer Feedforward Networks.Neural Networks, 3(5):551–560, 1990. 15 A Additional experiments A.1 Sensitivity Analysis and Ablations We evaluateFLASH-MAXwith different network widths and activation functions in Table 3. The resul...

    work page 1990