Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Study on a Fast Solver for Combined Field Integral Equations of 3D Conducting Bodies Based on Graph Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GraphSolver predicts surface current densities on 3D conducting bodies with a single feed-forward graph network, matching method-of-moments radar cross sections while training far faster than a physics-informed graph solver.

desk verdict GraphSolver is a reproducible, faster-to-train GNN surrogate for CFIE surface currents, but the 'solver' claim overreaches because every test is interpolation inside the training distribution and the reported accuracy is 2–3x worse than PhiGRL despite the text saying it's comparable. read the letter →

arxiv 2501.09923 v1 pith:ZX6GW55W submitted 2025-01-17 cs.LG cs.AIcs.NAmath.NA

Tao Shan , Xin Zhang , Di Wu This is my paper · ORCID
classification cs.LGcs.AIcs.NAmath.NA
keywords graphneuralnetworkcombined-fieldintegralequationsurfacecurrentdensityelectromagneticscatteringmethodofmomentsRao-Wilton-Glissonbasisfunctionsbistaticradarcrosssectionsupervisedlearning
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

GraphSolver is a fully data-driven alternative to the standard numerical route through the combined-field integral equation (CFIE) for 3D conducting bodies. The paper's proposal is to turn a body's triangular mesh into a graph, with one node per triangular patch and an edge wherever two patches share a side, and to train a graph neural network that directly outputs the real and imaginary parts of the x, y, and z components of the surface current density at every node. Training data are generated once by the method of moments, and the network is trained to minimize mean squared error against those reference currents. On spheroids, conical frustums, hexahedrons, missile-shaped targets, and airplane-shaped targets at 300 MHz and 150 MHz, the predicted surface currents and the resulting bistatic radar cross sections closely match the method-of-moments results. Compared with an earlier physics-informed graph residual learning solver, GraphSolver trains in roughly a third to a fifth of the time at somewhat higher error, which the authors argue makes it more suitable when computational resources are limited.

What carries the argument

The load-bearing object is the graph representation of the discretized body: each triangular patch is a node, and two nodes are connected when their triangles share a side, so the graph skeleton is exactly the mesh adjacency that supports current flow between the Rao-Wilton-Glisson basis functions used in the method of moments. On top of this graph, one upsampling fully connected network expands the input features, a graph convolutional layer updates each node's feature by combining its own feature with the mean of its neighbors' features weighted by a trainable kernel of the edge vector, and six downsampling fully connected networks independently output the real and imaginary parts of the x, y, and z components of the surface current density. The whole pipeline is trained end to end with a mean squared error loss against method-of-moments-generated surface currents.

What would settle it

Train GraphSolver exactly as described, then evaluate it on a 3D conducting body from a different geometric family (say a torus or a generic ship hull) at the same 300 MHz and the same $\lambda/10$ mesh; if the predicted surface currents or the resulting bistatic radar cross section error is much larger than the errors reported on the training families, the claim of a general fast solver fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a GNN can solve CFIEs for 3D conducting bodies in a purely supervised, feed-forward way: no residual computation, no iterative correction, and no numerical solver in the loop. The graph is built by making each triangular element a node and connecting nodes whose triangles share a side, so the RWG basis functions that physically carry current between adjacent triangles become the edges over which the network passes messages. Each node's input feature is the node's position plus the incident surface current computed from the illuminating plane wave; the output is the real and imaginary parts of all three Cartesian components of the surface current density at that node. The reported experiments show stable training with negligible overfitting on the basic and missile-shaped targets, and acceptable accuracy on the airplane-shaped targets, where batch normalization was added to handle larger fluctuations.

Load-bearing premise

The claim rests on assuming that a network trained on a few shape families, two frequencies, and one mesh size will solve the combined-field integral equation for arbitrary 3D conducting bodies; the experiments never test a shape, frequency, or mesh density outside that training distribution.

Editorial extensions

If this is right

  • If the trained GraphSolver is correct, online prediction for a new discretization of a seen target family reduces to one feed-forward graph pass, with no impedance-matrix assembly and no iterative solve at inference time.
  • The reported comparison implies that the data-driven solver reaches usable accuracy for radar cross section prediction while spending far less wall-clock training time than the physics-informed residual learning baseline, which matters when GPU hours are limited.
  • Transfer learning from the basic-target model lets the network adapt to the more complex airplane shapes with fewer method-of-moments samples than training from scratch would require.
  • Because the graph representation adapts to different numbers and distributions of triangular elements, the same architecture can in principle handle meshes of different sizes without re-architecting the network.

Reading between the lines

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

  • Beyond the paper, the central claim would be tested most directly by evaluating GraphSolver on a 3D shape family it never saw during training, such as a torus or a generic ship hull, at the same frequency and mesh density; the current experiments only cover targets drawn from the same parametric families used to generate the training data.
  • Beyond the paper, extrapolation in frequency and mesh resolution is untested: the trained models are evaluated at 300 MHz and 150 MHz with a $\lambda/10$ mesh only, so a cheap extension would be to run the same model on the same targets at a new frequency or mesh size and compare radar cross sections.
  • An editorial inference is that the comparison with the physics-informed baseline is on training time and mean squared error; total cost including data generation and end-to-end inference latency would need to be measured before declaring the method suitable for real-time use.
  • One refinement the authors do not pursue is to couple the six output heads through local charge conservation, since the x, y, and z components of a surface current are physically linked; projecting the predictions onto divergence-free currents could improve accuracy on the sharp-cornered airplane targets.
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 / 5 minor

Summary. The paper proposes GraphSolver, a graph neural network that directly predicts surface current densities on 3D conducting (PEC) bodies from their triangular meshes, bypassing the MoM matrix solve. A graph is built by treating each triangular element as a node and connecting adjacent elements; the input features are the incident-field current approximation and node position. The architecture is an upsampling FCN followed by a GCN and six per-component FCNs, trained with MSE loss against MoM-generated currents. Numerical experiments cover basic 3D shapes (spheroids, conical frustums, hexahedrons), missile-shaped targets, and three airplane-shaped targets, with comparisons to the physics-informed PhiGRL method.

Significance. If the claims are validated, a fast data-driven CFIE surrogate would be useful for real-time EM scattering prediction, and the public release of code and trained models is a notable strength for reproducibility. However, the current evidence demonstrates interpolation within the training distribution only: test targets are drawn from the same parametric families, at the same frequencies, and at the same mesh density as the training data. The stated 'comparable computational accuracy' of GraphSolver relative to PhiGRL is also contradicted by the numerical values in Table III, which show roughly 2-3 times larger test MSE. With these gaps, the paper's broader claim of solving CFIEs for 3D conducting bodies is not yet established.

major comments (3)
  1. [Section IV-D, Table III] The statement in Section IV-D that GraphSolver 'maintains comparable computational accuracy' is not supported by Table III. For basic targets, GraphSolver's test MSE is 0.0015 versus PhiGRL's 0.00071; for missile-shaped targets, 0.0035 versus 0.0011; for airplane-shaped targets, 0.014 versus 0.0066. These are approximately 2.1x, 3.2x, and 2.1x larger, respectively. Because this comparison is the basis for claiming suitability in resource-limited scenarios, the claim needs to be either substantially qualified or supported by a statistical equivalence test.
  2. [Section IV (all subsections), Section III-A Eq. (4)] The evaluation is entirely in-distribution: the test targets are drawn from the same parametric families listed in Table II (plus the three fixed airplane shapes), all at fixed frequencies (300 MHz for basic and missile targets, 150 MHz for airplanes) and at a fixed lambda/10 mesh. The input feature vector in Eq. (4) contains only the incident-field current J_inc_s and the node position r; it carries no information about frequency, electrical size, or mesh density. Therefore the network has no mechanism to adapt to different frequencies or mesh resolutions, and no experiment demonstrates such adaptation. The title and abstract claim that GraphSolver solves CFIEs for '3D conducting bodies' generally, which is not supported by the reported experiments. Please add out-of-distribution generalization tests (e.g., different frequencies, mesh densities, or shape families) or restrict the claims to the tested distribution.
  3. [Section IV-C, Figure 7] The airplane-shaped target experiments use only three fixed geometries, and the model is trained via transfer learning from the basic-target model. The manuscript does not state whether the test set for the airplane category consists of the same three geometries with different incident angles or of held-out airplane geometries. If the test set contains the same geometries, the results would demonstrate interpolation over incident angles within three shapes rather than generalization across airplane shapes. Please clarify the train/test split for the airplane datasets and, if possible, evaluate on held-out airplane geometries to support the generalization claim.
minor comments (5)
  1. [Figures 9, 10, and 11 captions] The captions repeatedly state 'real parts of the x, y and y components' and 'imaginary parts of the x, y and y components'; these should read 'x, y, and z components'.
  2. [Section III-A] The graph representation treats each triangular element as a node, while RWG basis functions are defined on edges. The relationship between the per-triangle output current density and the edge-based RWG representation is not explicitly clarified; the claim that RWG functions 'enable the flow of current between nodes' would benefit from a more precise explanation of how the graph encodes the RWG expansion.
  3. [Section IV-A vs. Table III] Section IV-A states that 32,400 data samples are generated with an 80%/20% split, which would give 25,920 training samples, but Table III reports 25,630 training samples for the basic targets. Please reconcile these numbers.
  4. [Throughout] There are several typographical errors: 'misslehead' appears in Table II, 'RWG basises' appears in Section IV-C, and 'the author's knowledge' in Section I should be 'the authors' knowledge'.
  5. [Section V] The conclusion states that GraphSolver is a 'strong candidate for scenarios with limited computational resources,' but the comparison is only with PhiGRL (a physics-informed neural solver). A comparison with a fast iterative numerical solver such as MLFMA, or a discussion of how GraphSolver's online cost compares to such solvers, would strengthen this conclusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the network is a supervised surrogate trained on MoM data, and its held-out evaluation is a standard generalization test, not an equation-level reduction.

full rationale

The paper's derivation chain is: the CFIE (Eq. 1) is discretized with RWG basis functions and converted to a MoM matrix equation (Eqs. 2-3); GraphSolver instead predicts surface current densities directly from a graph representation (Eqs. 4-8); and the network is trained by minimizing MSE against MoM-generated currents (Eq. 9). The prediction target is not defined in terms of the network output; the ground-truth labels come from an independent numerical solver (MoM), and the reported test errors are computed on held-out samples from the same parametric families. This is a standard supervised-learning evaluation, not a case where a fitted parameter is renamed as a prediction or where the output equals the input by construction. The paper does invoke the authors' prior work, PhiGRL [32], for a training-time comparison in Table III, but that comparison is empirical and not load-bearing for the central claim that GraphSolver can approximate CFIE solutions on the tested distributions. The graph convolution operation is attributed to an external reference [46], and no uniqueness theorem or ansatz is smuggled in via self-citation. The absence of out-of-distribution geometry, frequency, and mesh-size tests is a legitimate generalization and correctness concern, but it is not circularity: the in-distribution held-out results remain independent evidence of fitting quality. No specific reduction of the claimed result to its own inputs could be exhibited, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper's contribution is almost entirely a fitted surrogate. The physical equations (CFIE, MoM) are taken as established inputs. The load-bearing assumptions are that the training distribution is representative and that the graph representation and PO-inspired input feature are sufficient. No new physical entities are introduced.

free parameters (2)
  • Trainable network parameters of GraphSolver = Not enumerated
    The functional form of the solver is a neural network; all weights are fit to MoM data. The central claim depends on this fit.
  • Graph convolution kernel width w = Not stated
    The paper specifies GraphConv-[c,d,mean,w] but does not give the value of w; it is a hand-chosen hyperparameter affecting the relative position kernel.
assumptions (4)
  • domain assumption The CFIE (Eq. 1) and its MoM discretization with RWG basis functions produce ground-truth surface currents.
    The training data and the evaluation benchmark are both generated by MoM; the paper does not independently verify MoM accuracy.
  • domain assumption The graph representation using triangle nodes and edges between adjacent triangles preserves sufficient geometric information for current prediction.
    Section III-A asserts this representation is 'concise and informative' but provides no ablation or theoretical justification.
  • domain assumption The incident-field feature Jinc_s(r) = n x Z0 Hinc - n x n x Einc is a sufficient input feature.
    Eq. (4) defines the input node feature; the paper does not test alternatives or analyze the model's sensitivity to this PO-like feature.
  • ad hoc to paper Results on the tested parametric shape families generalize to '3D conducting bodies'.
    The conclusion claims a general fast solver, but the experiments only cover limited shapes at fixed frequencies and mesh sizes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Study on a Fast Solver for Combined Field Integral Equations of 3D Conducting Bodies Based on Graph Neural Networks." pith.science (2026). https://pith.science/paper/ZX6GW55W

@misc{pith2026250109923,
  author       = {Pith},
  title        = {Pith review of: Study on a Fast Solver for Combined Field Integral Equations of 3D Conducting Bodies Based on Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZX6GW55W}},
  note         = {Machine review of arXiv:2501.09923}
}
read the original abstract

In this paper, we present a graph neural networks (GNNs)-based fast solver (GraphSolver) for solving combined field integral equations (CFIEs) of 3D conducting bodies. Rao-Wilton-Glisson (RWG) basis functions are employed to discretely and accurately represent the geometry of 3D conducting bodies. A concise and informative graph representation is then constructed by treating each RWG function as a node in the graph, enabling the flow of current between nodes. With the transformed graphs, GraphSolver is developed to directly predict real and imaginary parts of the x, y and z components of the surface current densities at each node (RWG function). Numerical results demonstrate the efficacy of GraphSolver in solving CFIEs for 3D conducting bodies with varying levels of geometric complexity, including basic 3D targets, missile-shaped targets, and airplane-shaped targets.

Figures

Figures reproduced from arXiv: 2501.09923 by the authors.

Figure 1
Figure 1. Relationship between RWG basis function and surface current density. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Graph representation of RWG basis functions in a 3D conducting body. The 3D conducting body is first discretized into triangular meshes, with each [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. GraphSolver for solving CFIEs of 3D conducting bodies. It consists of one upsampling FCN, one GCN, and six downsampling FCNs. Their detailed [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Schematic of basic 3D targets: spheroid, conical frustum, and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Convergence curve of GraphSolver for solving CFIEs of basic 3D [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Schematic of missile-shaped targets and the convergence curve of [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Schematic of three airplane-shaped targets with their respective lengths [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]
Figure 8
Figure 8. Figure 8: Convergence curve of GraphSolver for solving CFIEs of airplane [PITH_FULL_IMAGE:figures/full_fig_p004_8.png]
Figure 9
Figure 9. Figure 9: Comparison between the surface currents computed by MoM and GraphSolver. (a), (b) and (c) are the results of spheroid, conical frustum and [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 10
Figure 10. Figure 10: Comparison between the surface currents computed by MoM and GraphSolver. (a) and (b) are two different misslehead-shaped targets. In each [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Comparison between the surface currents computed by MoM and GraphSolver. (a), (b) and (c) are the results of three different airplane-shaped [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 43 canonical work pages

  1. [1]

    W. C. Chew, E. Michielssen, J. Song, and J.-M. Jin, Fast and efficient algorithms in computational electromagnetics . Artech House, Inc., 2001

    work page 2001

  2. [2]

    W. C. Gibson, The method of moments in electromagnetics . Chapman and Hall/CRC, 2021

    2021

  3. [3]

    Jin, Theory and computation of electromagnetic fields

    J.-M. Jin, Theory and computation of electromagnetic fields. John Wiley & Sons, 2011

    work page 2011

  4. [4]

    Chew, M.-S

    W. Chew, M.-S. Tong, and H. Bin, Integral equation methods for electromagnetic and elastic waves . Springer Nature, 2022

    work page 2022

  5. [5]

    Pastorino and A

    M. Pastorino and A. Randazzo, Microwave imaging methods and appli- cations. Artech House, 2018

    work page 2018

  6. [6]

    2.5 D forward and inverse modeling for interpreting low- frequency electromagnetic measurements: Geophysics, 73,

    A. Abubakar, T. Habashy, V . Druskin, L. Knizhnerman, and D. Alum- baugh, “2.5 D forward and inverse modeling for interpreting low- frequency electromagnetic measurements: Geophysics, 73,” F165–F177, 2008

    work page 2008

  7. [7]

    Jin, Electromagnetic scattering modelling for quantitative remote sensing

    Y .-Q. Jin, Electromagnetic scattering modelling for quantitative remote sensing. World Scientific, 1993

    work page 1993

  8. [8]

    C. R. Paul, R. C. Scully, and M. A. Steffka, Introduction to electromag- netic compatibility. John Wiley & Sons, 2022

    work page 2022

Show all 47 references
  1. [9]

    Computational elec- tromagnetics: the finite-difference time-domain method,

    A. Taflove, S. C. Hagness, and M. Piket-May, “Computational elec- tromagnetics: the finite-difference time-domain method,” The Electrical Engineering Handbook, vol. 3, no. 629-670, p. 15, 2005

    2005

  2. [10]

    Jin, The finite element method in electromagnetics

    J.-M. Jin, The finite element method in electromagnetics . John Wiley & Sons, 2015

    2015

  3. [11]

    Discontinuous galerkin time-domain methods for multiscale electromagnetic simulations: A review,

    J. Chen and Q. H. Liu, “Discontinuous galerkin time-domain methods for multiscale electromagnetic simulations: A review,” Proceedings of the IEEE, vol. 101, no. 2, pp. 242–254, 2012

    2012

  4. [12]

    The adaptive cross approxima- tion algorithm for accelerated method of moments computations of emc problems,

    K. Zhao, M. N. V ouvakis, and J.-F. Lee, “The adaptive cross approxima- tion algorithm for accelerated method of moments computations of emc problems,” IEEE transactions on electromagnetic compatibility , vol. 47, no. 4, pp. 763–773, 2005

    2005

  5. [13]

    Application of fft and the conjugate gradient method for the solution of electromagnetic radiation from electrically large and small conducting bodies,

    T. Sarkar, E. Arvas, and S. Rao, “Application of fft and the conjugate gradient method for the solution of electromagnetic radiation from electrically large and small conducting bodies,” IEEE Transactions on Antennas and Propagation , vol. 34, no. 5, pp. 635–640, 1986

    1986

  6. [14]

    Rapid solution of integral equations of classical potential theory,

    V . Rokhlin, “Rapid solution of integral equations of classical potential theory,” Journal of computational physics , vol. 60, no. 2, pp. 187–207, 1985

    1985

  7. [15]

    Machine learning in electromagnetics: A review and some perspectives for future research,

    D. Erricolo, P.-Y . Chen, A. Rozhkova, E. Torabi, H. Bagci, A. Shamim, and X. Zhang, “Machine learning in electromagnetics: A review and some perspectives for future research,” in 2019 International Conference on Electromagnetics in Advanced Applications (ICEAA) . IEEE, 2019, ...

    2019

  8. [16]

    DNNs as applied to electromagnetics, antennas, and propagation—A review,

    A. Massa, D. Marcantonio, X. Chen, M. Li, and M. Salucci, “DNNs as applied to electromagnetics, antennas, and propagation—A review,” IEEE Antennas and Wireless Propagation Letters , vol. 18, no. 11, pp. 2225–2229, 2019

    2019

  9. [17]

    Artificial Intelligence: New Frontiers in Real–Time Inverse Scattering and Electromagnetic Imaging,

    M. Salucci, M. Arrebola, T. Shan, and M. Li, “Artificial Intelligence: New Frontiers in Real–Time Inverse Scattering and Electromagnetic Imaging,” IEEE Transactions on Antennas and Propagation , 2022

    2022

  10. [18]

    A review of deep learning approaches for inverse scattering problems (invited review),

    X. Chen, Z. Wei, L. Maokun, P. Rocca et al. , “A review of deep learning approaches for inverse scattering problems (invited review),” ELECTROMAGNETIC WAVES, vol. 167, pp. 67–81, 2020

    2020

  11. [19]

    Machine-learning-based PML for the FDTD method,

    H. M. Yao and L. Jiang, “Machine-learning-based PML for the FDTD method,” IEEE Antennas and Wireless Propagation Letters , vol. 18, no. 1, pp. 192–196, 2018

    2018

  12. [20]

    Study on a fast solver for Poisson’s equation based on deep learning technique,

    T. Shan, W. Tang, X. Dang, M. Li, F. Yang, S. Xu, and J. Wu, “Study on a fast solver for Poisson’s equation based on deep learning technique,” IEEE Transactions on Antennas and Propagation , vol. 68, no. 9, pp. 6725–6733, 2020

    2020

  13. [21]

    Application of Multitask Learning for 2-D Modeling of Magnetotelluric Surveys: TE Case,

    T. Shan, R. Guo, M. Li, F. Yang, S. Xu, and L. Liang, “Application of Multitask Learning for 2-D Modeling of Magnetotelluric Surveys: TE Case,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–9, 2021

    2021

  14. [22]

    Machine-Learning- Based Hybrid Method for the Multilevel Fast Multipole Algorithm,

    J.-J. Sun, S. Sun, Y . P. Chen, L. Jiang, and J. Hu, “Machine-Learning- Based Hybrid Method for the Multilevel Fast Multipole Algorithm,” IEEE Antennas and Wireless Propagation Letters , vol. 19, no. 12, pp. 2177–2181, 2020

    2020

  15. [23]

    A Surrogate Model for the Rapid Evaluation of Electromagnetic-Thermal Effects under Humid Air Conditions,

    H. Zhu, H. Wang, H. Zhang, N. Wang, Q. Ren, Y . Chen, F. Liu, and J. Gao, “A Surrogate Model for the Rapid Evaluation of Electromagnetic-Thermal Effects under Humid Air Conditions,” Elec- tronics, vol. 13, no. 12, p. 2336, 2024

    2024

  16. [24]

    RayProNet: A Neural Point Field Framework for Radio Propagation Modeling in 3D Environments,

    G. Cao and Z. Peng, “RayProNet: A Neural Point Field Framework for Radio Propagation Modeling in 3D Environments,” arXiv preprint arXiv:2406.16907, 2024

    2024 arXiv

  17. [25]

    Multi-Frequency Data Acquisition Model and Hybrid Neu- ral Network for Precise Electromagnetic Wellbore Casing Inspection,

    G. A. Ooi, M. Khater, M. B. Ozakin, T. M. Mostafa, H. Bagci, and S. Ahmed, “Multi-Frequency Data Acquisition Model and Hybrid Neu- ral Network for Precise Electromagnetic Wellbore Casing Inspection,” in Abu Dhabi International Petroleum Exhibition and Conference. SPE, 2022, p....

    2022

  18. [26]

    Learning-based fast electromagnetic scattering solver through generative adversarial network,

    Z. Ma, K. Xu, R. Song, C.-F. Wang, and X. Chen, “Learning-based fast electromagnetic scattering solver through generative adversarial network,” IEEE Transactions on Antennas and Propagation , vol. 69, no. 4, pp. 2194–2208, 2020

    2020

  19. [27]

    Predicting macro basis functions for method of moments scattering problems using deep neural networks,

    C. Key and B. M. Notaro s, “Predicting macro basis functions for method of moments scattering problems using deep neural networks,” IEEE Antennas and Wireless Propagation Letters , vol. 20, no. 7, pp. 1200– 1204, 2021

    2021

  20. [28]

    An AI Predictor: From Point Clouds to Scattered Far Fields for 3-D PEC Targets,

    D.-H. Kong, J.-N. Cao, W.-W. Zhang, W.-C. Huang, X.-Y . He, L. Liu, and M.-Y . Xia, “An AI Predictor: From Point Clouds to Scattered Far Fields for 3-D PEC Targets,” IEEE Transactions on Antennas and Propagation, vol. 72, no. 6, pp. 5179–5190, 2024

    2024

  21. [29]

    Hybrid Physics-Informed Neural Network for the Wave Equation with Unconditionally Stable Time-Stepping,

    S. Qi and C. D. Sarris, “Hybrid Physics-Informed Neural Network for the Wave Equation with Unconditionally Stable Time-Stepping,” IEEE Antennas and Wireless Propagation Letters , 2024

    2024

  22. [30]

    Electromagnetic Modeling Using an FDTD-Equivalent Recurrent Convolution Neural Network: Accurate Computing on a Deep Learning Framework

    L. Guo, M. Li, S. Xu, F. Yang, and L. Liu, “Electromagnetic Modeling Using an FDTD-Equivalent Recurrent Convolution Neural Network: Accurate Computing on a Deep Learning Framework.” IEEE Antennas and Propagation Magazine , 2021

    2021

  23. [31]

    A theory-guided deep neural network for time domain electromagnetic simulation and inversion using a differentiable programming platform,

    Y . Hu, Y . Jin, X. Wu, and J. Chen, “A theory-guided deep neural network for time domain electromagnetic simulation and inversion using a differentiable programming platform,” IEEE Transactions on Antennas and Propagation, vol. 70, no. 1, pp. 767–772, 2021

    2021

  24. [32]

    Physics- informed supervised residual learning for electromagnetic modeling,

    T. Shan, J. Zeng, X. Song, R. Guo, M. Li, F. Yang, and S. Xu, “Physics- informed supervised residual learning for electromagnetic modeling,” IEEE Transactions on Antennas and Propagation , vol. 71, no. 4, pp. 3393–3407, 2023

    2023

  25. [33]

    Solving Combined Field Integral Equations with Physics-informed Graph Residual Learning for EM Scattering of 3D PEC Targets,

    T. Shan, M. Li, F. Yang, and S. Xu, “Solving Combined Field Integral Equations with Physics-informed Graph Residual Learning for EM Scattering of 3D PEC Targets,” IEEE Transactions on Antennas and Propagation, 2023

    2023

  26. [34]

    DeepNIS: Deep neural network for nonlinear electromagnetic inverse scattering,

    L. Li, L. G. Wang, F. L. Teixeira, C. Liu, A. Nehorai, and T. J. Cui, “DeepNIS: Deep neural network for nonlinear electromagnetic inverse scattering,” IEEE Transactions on Antennas and Propagation , vol. 67, no. 3, pp. 1819–1825, 2018. 10

    2018

  27. [35]

    Electromagnetic inverse scattering with perceptual generative adversarial networks,

    R. Song, Y . Huang, K. Xu, X. Ye, C. Li, and X. Chen, “Electromagnetic inverse scattering with perceptual generative adversarial networks,” IEEE Transactions on Computational Imaging , vol. 7, pp. 689–699, 2021

    2021

  28. [36]

    A Multi-branch Deep Learn- ing Architecture for Microwave-Ultrasound Breast Imaging,

    V . Khoshdel, P. Mojabi, and J. LoVetri, “A Multi-branch Deep Learn- ing Architecture for Microwave-Ultrasound Breast Imaging,” in 2023 XXXVth General Assembly and Scientific Symposium of the International Union of Radio Science (URSI GASS) . IEEE, 2023, pp. 1–4

    2023

  29. [37]

    Physics-Informed Supervised Residual Learning for 2-D Inverse Scattering Problems,

    T. Shan, Z. Lin, X. Song, M. Li, F. Yang, and S. Xu, “Physics-Informed Supervised Residual Learning for 2-D Inverse Scattering Problems,” IEEE Transactions on Antennas and Propagation , vol. 71, no. 4, pp. 3746–3751, 2023

    2023

  30. [38]

    3DInvNet: A deep learning-based 3D ground-penetrating radar data inversion,

    Q. Dai, Y . H. Lee, H.-H. Sun, G. Ow, M. L. M. Yusof, and A. C. Yucel, “3DInvNet: A deep learning-based 3D ground-penetrating radar data inversion,” IEEE Transactions on Geoscience and Remote Sensing , vol. 61, pp. 1–16, 2023

    2023

  31. [39]

    Neural born iterative method for solving inverse scattering problems: 2D cases,

    T. Shan, Z. Lin, X. Song, M. Li, F. Yang, and S. Xu, “Neural born iterative method for solving inverse scattering problems: 2D cases,” IEEE Transactions on Antennas and Propagation , vol. 71, no. 1, pp. 818–829, 2022

    2022

  32. [40]

    Unrolled convo- lutional neural network for full-wave inverse scattering,

    Y . Zhang, M. Lambert, A. Fraysse, and D. Lesselier, “Unrolled convo- lutional neural network for full-wave inverse scattering,” IEEE Transac- tions on Antennas and Propagation , vol. 71, no. 1, pp. 947–956, 2022

    2022

  33. [41]

    Phase synthesis of beam-scanning reflectarray antenna based on deep learning technique,

    T. Shan, M. Li, S. Xu, and F. Yang, “Phase synthesis of beam-scanning reflectarray antenna based on deep learning technique,” Progress In Electromagnetics Research, vol. 172, pp. 41–49, 2021

    2021

  34. [42]

    Coding programmable metasurfaces based on deep learning techniques,

    T. Shan, X. Pan, M. Li, S. Xu, and F. Yang, “Coding programmable metasurfaces based on deep learning techniques,” IEEE journal on emerging and selected topics in circuits and systems , vol. 10, no. 1, pp. 114–125, 2020

    2020

  35. [43]

    Real-Time Precision Prediction of 3-D Package Thermal Maps via Image-to-Image Translation,

    M. J. Smith, S. Hwang, V . C. Do Nascimento, Q. Qiu, C.-K. Koh, G. Subbarayan, and D. Jiao, “Real-Time Precision Prediction of 3-D Package Thermal Maps via Image-to-Image Translation,” in 2023 IEEE 32nd Conference on Electrical Performance of Electronic Packaging and Systems (...

    2023

  36. [44]

    Artificial Neural Networks for Microwave Computer-Aided Design: The State of the Art,

    F. Feng, W. Na, J. Jin, J. Zhang, W. Zhang, and Q.-J. Zhang, “Artificial Neural Networks for Microwave Computer-Aided Design: The State of the Art,” IEEE Transactions on Microwave Theory and Techniques , 2022

    2022

  37. [45]

    Fourier neural operator for parametric partial differential equations,

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, “Fourier neural operator for parametric partial differential equations,” arXiv preprint arXiv:2010.08895 , 2020

    2010 arXiv

  38. [46]

    Neural message passing for quantum chemistry,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” 2017. [Online]. Available: https://arxiv.org/abs/1704.01212

    2017 arXiv

  39. [47]

    A comprehensive survey on transfer learning,

    F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, and Q. He, “A comprehensive survey on transfer learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43–76, 2020

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.