arxiv: 2512.09333 · v1 · submitted 2025-12-10 · 💻 cs.LG

Improved Physics-Driven Neural Network to Solve Inverse Scattering Problems

Yutong Du , Zicheng Liu , Bo Wu , Jingwei Kou , Hang Li , Changyou Li , Yali Zong , Bo Qi This is my paper

Pith reviewed 2026-05-16 23:48 UTC · model grok-4.3

classification 💻 cs.LG

keywords inverse scattering problemsphysics-driven neural networkselectromagnetic reconstructionactivation functionstransfer learningdynamic subregioncomputational efficiency

0 comments

The pith

An improved physics-driven neural network with GLOW activation reconstructs electromagnetic scatterers more accurately and efficiently.

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

This paper develops an improved physics-driven neural network framework for solving electromagnetic inverse scattering problems. It introduces a Gaussian-localized oscillation-suppressing window activation function to stabilize convergence in a lightweight architecture. A dynamic scatter subregion identification strategy adaptively refines the computational domain to reduce cost and prevent missed detections. Transfer learning integrates the method with practical data while preserving physical interpretability from iterative algorithms. Numerical simulations and experiments show the approach delivers higher reconstruction accuracy, robustness, and efficiency than existing state-of-the-art solvers.

Core claim

The IPDNN framework uses a new GLOW activation function and a dynamic subregion strategy together with transfer learning to combine the physical constraints of iterative solvers with the speed of neural networks, yielding superior accuracy and lower computational cost in reconstructing scatterers from scattered electromagnetic fields.

What carries the argument

The GLOW activation function, which suppresses oscillations while localizing Gaussian windows to stabilize the physics-driven network training, paired with the dynamic scatter subregion strategy that adaptively identifies and refines relevant computational domains.

If this is right

Enables real-time inference for inverse scattering while retaining physical consistency.
Reduces overall computational cost by focusing resources only on active subregions.
Extends applicability to practical scenarios through transfer learning without retraining from scratch.
Improves robustness against noise compared with prior neural and iterative methods.

Where Pith is reading between the lines

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

The lightweight architecture may support deployment on embedded hardware for field-deployable imaging systems.
Similar activation and domain-adaptation ideas could transfer to acoustic or optical inverse problems.
Further gains might come from hybridizing the network with additional wave-equation constraints.

Load-bearing premise

The GLOW activation and dynamic subregion strategy will generalize to complex unseen real-world scattering cases without introducing instability or missed detections.

What would settle it

A set of experimental measurements from a complex noisy scattering target outside the training distribution where the network either misses a scatterer or fails to converge.

Figures

Figures reproduced from arXiv: 2512.09333 by Bo Qi, Bo Wu, Changyou Li, Hang Li, Jingwei Kou, Yali Zong, Yutong Du, Zicheng Liu.

**Figure 2.** Figure 2: Variation of GLOW function value when (a) fixing [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: The schematic diagram of the applied neural network architecture. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: The flow chart for the dynamic scatter subregion identification method. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Based on the experimental dataset (a) “dielTM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: The convergence curve corresponding to the imaging example of “FoamDieInt” in Figure 5. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: Imaging examples with the experimental dataset “FoamDielInt” by varying the hyperparameter [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 8.** Figure 8: The convergence curves indicate that with GLOW activation function, the convergence rate is insensitive [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

**Figure 9.** Figure 9: Dynamic scatter subregion identification and imaging results of the four representative scatterers. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

**Figure 10.** Figure 10: Test of noise robustness of the proposed IPDNN sovers versus the classical SOM and a data-driven [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

**Figure 11.** Figure 11: The intermediate imaging results during iteration with (denoted by “Pretrained”) or without (denoted [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗

**Figure 12.** Figure 12: Predicted relative-permittivity distribution from the solver that pretrained based on the dataset [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗

read the original abstract

This paper presents an improved physics-driven neural network (IPDNN) framework for solving electromagnetic inverse scattering problems (ISPs). A new Gaussian-localized oscillation-suppressing window (GLOW) activation function is introduced to stabilize convergence and enable a lightweight yet accurate network architecture. A dynamic scatter subregion identification strategy is further developed to adaptively refine the computational domain, preventing missed detections and reducing computational cost. Moreover, transfer learning is incorporated to extend the solver's applicability to practical scenarios, integrating the physical interpretability of iterative algorithms with the real-time inference capability of neural networks. Numerical simulations and experimental results demonstrate that the proposed solver achieves superior reconstruction accuracy, robustness, and efficiency compared with existing state-of-the-art methods.

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.

Desk Editor's Note private letter to a colleague

The GLOW activation and dynamic subregion strategy add something concrete to physics-informed networks for inverse scattering, though the performance gains still need quantitative backing.

read the letter

The main takeaway is that the authors have added a Gaussian-localized oscillation-suppressing window activation function and a dynamic way to pick out scatter subregions, which could make physics-driven networks more practical for inverse scattering. These changes are presented as extensions that stabilize training and cut down on unnecessary computation by focusing the domain. Adding transfer learning to move from simulations to experiments is a logical move to improve real-world use. The paper does a good job explaining how these pieces integrate the physical model with neural network speed. The idea of preventing missed detections through adaptive refinement makes sense for imaging applications where scatterers can be sparse. Where it falls short is the lack of concrete evidence for the superiority claims. The abstract mentions numerical simulations and experimental results showing better accuracy, robustness, and efficiency, but without metrics, baselines, or ablation studies, it's difficult to see how much of an improvement there really is. The dynamic subregion strategy might work well on the tested cases, but if its identification rules don't generalize when scatterer numbers or noise change, the robustness part could be overstated. This paper is for researchers already working on neural solvers for electromagnetic inverse problems. They might get value from testing the new activation or the subregion method in their setups. It deserves peer review because the core ideas are specific and build on solid prior work in the area, even though the results section will need strengthening to support the conclusions.

Referee Report

2 major / 2 minor

Summary. This paper presents an improved physics-driven neural network (IPDNN) framework for electromagnetic inverse scattering problems. It introduces a new Gaussian-localized oscillation-suppressing window (GLOW) activation function to stabilize convergence in a lightweight architecture, a dynamic scatter subregion identification strategy to adaptively refine the computational domain and avoid missed detections, and transfer learning to extend applicability to practical scenarios. The central claim, supported by numerical simulations and experiments, is that the solver achieves superior reconstruction accuracy, robustness, and efficiency compared to existing state-of-the-art methods.

Significance. If the central claims hold, the work could meaningfully advance real-time electromagnetic imaging by integrating the physical consistency of iterative solvers with the speed of neural inference. The GLOW activation and dynamic subregion approach, if shown to generalize, would offer concrete tools for reducing computational cost while maintaining accuracy in inverse scattering applications such as microwave imaging.

major comments (2)

[Abstract] Abstract: the claim of 'superior reconstruction accuracy, robustness, and efficiency' is asserted without any quantitative metrics (e.g., RMSE, SSIM, runtime comparisons), error bars, specific baseline implementations, or ablation results, preventing assessment of whether the improvements are load-bearing or marginal.
[Method (dynamic subregion strategy)] Dynamic subregion identification strategy (described in the method section): the adaptive refinement relies on decision thresholds and refinement criteria tuned on the training distribution; no experiments demonstrate stability under distribution shift (different scatterer counts, contrasts, or noise statistics). Under-segmentation would supply an incomplete domain to the network and directly falsify the robustness component of the headline claim.

minor comments (2)

[Method] The explicit mathematical definition and hyper-parameters of the GLOW activation function should be stated as an equation rather than described only in prose.
[Numerical results] Figure captions for the reconstruction results should include the exact quantitative error metrics and the precise SOTA baselines being compared.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, providing clarifications and indicating revisions made to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of 'superior reconstruction accuracy, robustness, and efficiency' is asserted without any quantitative metrics (e.g., RMSE, SSIM, runtime comparisons), error bars, specific baseline implementations, or ablation results, preventing assessment of whether the improvements are load-bearing or marginal.

Authors: We agree that the abstract would benefit from explicit quantitative support to allow readers to evaluate the claims directly. In the revised manuscript, we have updated the abstract to include specific metrics such as average RMSE reductions, SSIM scores, runtime comparisons against named baselines (e.g., standard PDNN and other iterative solvers), and references to error bars and ablation studies reported in the results section. revision: yes
Referee: [Method (dynamic subregion strategy)] Dynamic subregion identification strategy (described in the method section): the adaptive refinement relies on decision thresholds and refinement criteria tuned on the training distribution; no experiments demonstrate stability under distribution shift (different scatterer counts, contrasts, or noise statistics). Under-segmentation would supply an incomplete domain to the network and directly falsify the robustness component of the headline claim.

Authors: We thank the referee for raising this important concern about generalization. The decision thresholds are derived from electromagnetic scattering physics rather than being purely empirical, which provides a degree of robustness. Nevertheless, to directly address distribution shift, we have added new experiments in the revised manuscript that test the subregion strategy on cases with varying scatterer counts, contrasts, and noise levels outside the original training distribution, confirming that under-segmentation is avoided and reconstruction quality is maintained. revision: yes

Circularity Check

0 steps flagged

Minor self-citation but independent new components; no reduction to inputs by construction

full rationale

The paper introduces GLOW activation and dynamic subregion strategy as novel elements within a physics-driven NN framework for inverse scattering. These additions are presented as independent improvements rather than self-definitional or fitted-input predictions. Central superiority claims rest on numerical simulations and experiments, which constitute external validation. Any self-citations are not load-bearing for the core derivation, keeping circularity minimal and non-central.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Framework rests on standard electromagnetic scattering physics being correctly encoded in the network loss; new components are the GLOW function and subregion strategy, which lack independent verification beyond the reported tests.

axioms (1)

domain assumption Electromagnetic wave scattering physics is accurately captured by the neural network loss function.
Standard assumption for physics-informed neural networks in inverse scattering.

invented entities (1)

GLOW activation function no independent evidence
purpose: Stabilize convergence by suppressing oscillations in network outputs.
Newly proposed in this work with no external validation cited.

pith-pipeline@v0.9.0 · 5431 in / 1155 out tokens · 42511 ms · 2026-05-16T23:48:32.849345+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

[1]

Wiley, Hoboken, NJ, USA, 2018

Xudong Chen.Computational Methods for Electromagnetic Inverse Scattering. Wiley, Hoboken, NJ, USA, 2018

work page 2018
[2]

Electromagnetic modeling of dam- aged single-layer fiber-reinforced laminates.IEEE Trans

Zicheng Liu, Changyou Li, Dominique Lesselier, and Yu Zhong. Electromagnetic modeling of dam- aged single-layer fiber-reinforced laminates.IEEE Trans. Antennas Propag., 65(4):1855–1866, 2017

work page 2017
[3]

A 3-D printed continuous carbon fiber-reinforced composite for highly effective microwave shielding.IEEE Antennas Wirel

Changyou Li, Qian Zhu, Bing Lv, Yichou Huang, and Changying Wu. A 3-D printed continuous carbon fiber-reinforced composite for highly effective microwave shielding.IEEE Antennas Wirel. Propag. Lett., 20(5):758–762, 2021

work page 2021
[4]

Qiqi Dai, Yee Hui Lee, Hai-Han Sun, Genevieve Ow, Mohamed Lokman Mohd Yusof, and Abdulka- dir C. Yucel. 3DInvNet: A deep learning-based 3D ground-penetrating radar data inversion.IEEE Trans. Geosci. Remote Sens., 61:1–16, 2023

work page 2023
[5]

Deep learning inversion of ground-penetrating radar with prior physical velocity field for complex subsurface root system.IEEE Trans

Xiaowei Zhang, Xuan Zhao, Shenghua Lv, Lingfei Xv, Chen Lin, and Jian Wen. Deep learning inversion of ground-penetrating radar with prior physical velocity field for complex subsurface root system.IEEE Trans. Geosci. Remote Sens., 63:1–15, 2025

work page 2025
[6]

Sherif S. Ahmed. Microwave imaging in security — two decades of innovation.IEEE J. Microw., 1(1):191–201, 2021

work page 2021
[7]

Muruganandam and G

V. Muruganandam and G. Rajini. Enhancing airport security: Integrating pso-tam with deep learning for real-time threat assessment. In2024 2nd International Conference on Advances in Computation, Communication and Information Technology (ICAICCIT), volume 1, pages 182–188, 2024

work page 2024
[8]

Torres-Verdin

Guozhong Gao and C. Torres-Verdin. High-order generalized extended Born approximation for electromagnetic scattering.IEEE Trans. Antennas Propag., 54(4):1243–1256, 2006. 14

work page 2006
[9]

Habashy, Ross W

Tarek M. Habashy, Ross W. Groom, and Brian R. Spies. Beyond the Born and Rytov approximations: A nonlinear approach to electromagnetic scattering.J. Geophys. Res. Solid Earth, 98(B2):1759–1775, 1993

work page 1993
[10]

A. J. Devaney. A filtered backpropagation algorithm for diffraction tomography.Ultrason. Imaging, 4(4):336–350, 1982

work page 1982
[11]

An inverse algorithm for velocity reconstruction.Inverse Probl., 14(5):1345, 1998

Tsili Wang and Michael L Oristaglio. An inverse algorithm for velocity reconstruction.Inverse Probl., 14(5):1345, 1998

work page 1998
[12]

A. J. Devaney. Inverse-scattering theory within the Rytov approximation.Opt. Lett., 6(8):374–376, 1981

work page 1981
[13]

Slaney, A.C

M. Slaney, A.C. Kak, and L.E. Larsen. Limitations of imaging with first-order diffraction tomography. IEEE Trans. Microwave Theory Tech., 32(8):860–874, 1984

work page 1984
[14]

Alon Schatzberg and Anthony J. Devaney. Rough surface inverse scattering within the Rytov ap- proximation.J. Opt. Soc. Am. A, 10(5):942–950, 1993

work page 1993
[15]

Y. M. Wang and W. C. Chew. An iterative solution of the two-dimensional electromagnetic inverse scattering problem.Int. J. Imag. Syst. Technol., 1(1):100–108, 1989

work page 1989
[16]

Jin Choi

Sung Chan Jun and U. Jin Choi. Convergence analyses of the Born iterative method and the distorted Born iterative method.Numer. Funct. Anal. Optim., 20(3-4):301–316, 1999

work page 1999
[17]

Chew and Y.M

W.C. Chew and Y.M. Wang. Reconstruction of two-dimensional permittivity distribution using the distorted Born iterative method.IEEE Trans. Med. Imaging, 9(2):218–225, 1990

work page 1990
[18]

Haddadin, S.D

O.S. Haddadin, S.D. Lucas, and E.S. Ebbini. Solution to the inverse scattering problem using a modified distorted Born iterative algorithm. In1995 IEEE Ultrason. Symp. Proc., volume 2, pages 1411–1414, 1995

work page 1995
[19]

A contrast source inversion method.Inverse Probl., 13(6):1607, 1997

Peter M van den Berg and Ralph E Kleinman. A contrast source inversion method.Inverse Probl., 13(6):1607, 1997

work page 1997
[20]

Inversion of experimental multi-frequency data using the contrast source inversion method.Inverse Probl., 17(6):1611, 2001

Richard F Bloemenkamp, Aria Abubakar, and Peter M van den Berg. Inversion of experimental multi-frequency data using the contrast source inversion method.Inverse Probl., 17(6):1611, 2001

work page 2001
[21]

X. Chen. Subspace-based optimization method for inverse scattering problems with an inhomoge- neous background medium.Inverse Probl., 26(7):074007, 2010

work page 2010
[22]

Subspace-based optimization method for solving inverse-scattering problems.IEEE Trans

Xudong Chen. Subspace-based optimization method for solving inverse-scattering problems.IEEE Trans. Geosci. Remote Sens., 48(1):42–49, 2010

work page 2010
[23]

Subspace-based optimization method for inverse scattering problems utilizing phaseless data.IEEE Trans

Li Pan, Yu Zhong, Xudong Chen, and Swee Ping Yeo. Subspace-based optimization method for inverse scattering problems utilizing phaseless data.IEEE Trans. Geosci. Remote Sens., 49(3):981– 987, 2011. 15

work page 2011
[24]

Deep-learning schemes for full-wave nonlinear inverse scattering prob- lems.IEEE Trans

Zhun Wei and Xudong Chen. Deep-learning schemes for full-wave nonlinear inverse scattering prob- lems.IEEE Trans. Geosci. Remote Sens., 57(4):1849–1860, 2019

work page 2019
[25]

Teixeira, Che Liu, Arye Nehorai, and Tie Jun Cui

Lianlin Li, Long Gang Wang, Fernando L. Teixeira, Che Liu, Arye Nehorai, and Tie Jun Cui. Deep- NIS: Deep neural network for nonlinear electromagnetic inverse scattering.IEEE Trans. Antennas Propag., 67(3):1819–1825, 2019

work page 2019
[26]

Physics-inspired convolutional neural network for solving full-wave inverse scattering problems.IEEE Trans

Zhun Wei and Xudong Chen. Physics-inspired convolutional neural network for solving full-wave inverse scattering problems.IEEE Trans. Antennas Propag., 67(9):6138–6148, 2019

work page 2019
[27]

Prasad, and Krishna Agarwal

Zicheng Liu, Mayank Roy, Dilip K. Prasad, and Krishna Agarwal. Physics-guided loss functions improve deep learning performance in inverse scattering.IEEE Trans. Comput. Imaging, 8:236–245, 2022

work page 2022
[28]

SOM-Net: Unrolling the subspace-based optimization for solving full-wave inverse scattering problems.IEEE Trans

Yu Liu, Hao Zhao, Rencheng Song, Xudong Chen, Chang Li, and Xun Chen. SOM-Net: Unrolling the subspace-based optimization for solving full-wave inverse scattering problems.IEEE Trans. Geosci. Remote Sens., 60:1–15, 2022

work page 2022
[29]

Neural Born iterative method for solving inverse scattering problems: 2D cases.IEEE Trans

Tao Shan, Zhichao Lin, Xiaoqian Song, Maokun Li, Fan Yang, and Shenheng Xu. Neural Born iterative method for solving inverse scattering problems: 2D cases.IEEE Trans. Antennas Propag., 71(1):818–829, 2023

work page 2023
[30]

Unrolled convolutional neural network for full-wave inverse scattering.IEEE Trans

Yarui Zhang, Marc Lambert, Aur´ elia Fraysse, and Dominique Lesselier. Unrolled convolutional neural network for full-wave inverse scattering.IEEE Trans. Antennas Propag., 71(1):947–956, 2023

work page 2023
[31]

Quality-factor- inspired deep neural network solver for solving inverse scattering problems.IEEE Trans

Yutong Du, Zicheng Liu, Miao Cao, Zupeng Liang, Yali Zong, and Changyou Li. Quality-factor- inspired deep neural network solver for solving inverse scattering problems.IEEE Trans. Geosci. Remote Sens., 63:1–13, 2025

work page 2025
[32]

Physics-driven neural network for solving electromagnetic inverse scattering problems, 2025

Yutong Du, Zicheng Liu, Bazargul Matkerim, Changyou Li, Yali Zong, Bo Qi, and Jingwei Kou. Physics-driven neural network for solving electromagnetic inverse scattering problems, 2025

work page 2025
[33]

Meaney, K.D

P.M. Meaney, K.D. Paulsen, and T.P. Ryan. Two-dimensional hybrid element image reconstruction for TM illumination.IEEE Trans. Antennas Propag., 43(3):239–247, 1995

work page 1995
[34]

Computational comparison and validation of point spread functions for optical microscopes.IEEE Trans

Zicheng Liu, Yingying Qin, Jean-Claude Tinguely, and Krishna Agarwal. Computational comparison and validation of point spread functions for optical microscopes.IEEE Trans. Comput. Imaging, 2025

work page 2025
[35]

Compressive sensing as applied to inverse problems for imaging: Theory, applications, current trends, and open challenges.IEEE Antennas Propag

Giacomo Oliveri, Marco Salucci, Nicola Anselmi, and Andrea Massa. Compressive sensing as applied to inverse problems for imaging: Theory, applications, current trends, and open challenges.IEEE Antennas Propag. Mag., 59(5):34–46, 2017

work page 2017
[36]

Rudin, Stanley Osher, and Emad Fatemi

Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms.Phys. D, 60(1):259–268, 1992. 16

work page 1992
[37]

M.M. Ney. Method of moments as applied to electromagnetic problems.IEEE Trans. Microw. Theory Tech., 33(10):972–980, 1985

work page 1985
[38]

Strong and weak forms of the method of moments and the coupled dipole method for scattering of time-harmonic electromagnetic fields.Int

Akhlesh Lakhtakia. Strong and weak forms of the method of moments and the coupled dipole method for scattering of time-harmonic electromagnetic fields.Int. J. Mod. Phys. C, 3(2):583–603, 1992

work page 1992
[39]

Gibson.The Method of Moments in Electromagnetics

W.C. Gibson.The Method of Moments in Electromagnetics. Chapman and Hall/CRC, New York, NY, USA, 2021

work page 2021
[40]

Inverse scattering solver based on deep neural network with total variation regularization.IEEE Antennas Wirel

Jie Ma, Zicheng Liu, and Yali Zong. Inverse scattering solver based on deep neural network with total variation regularization.IEEE Antennas Wirel. Propag. Lett., 22(10):2447–2451, 2023

work page 2023
[41]

Special section: Testing inversion algorithms against experimental data.Inverse Probl., 17(6):1565, nov 2001

Kamal Belkebir and Marc Saillard. Special section: Testing inversion algorithms against experimental data.Inverse Probl., 17(6):1565, nov 2001

work page 2001
[42]

Free space experimental scattering database continuation: experimental set-up and measurement precision.Inverse Probl., 21(6):S117, 2005

Jean-Michel Geffrin, Pierre Sabouroux, and Christelle Eyraud. Free space experimental scattering database continuation: experimental set-up and measurement precision.Inverse Probl., 21(6):S117, 2005. 17

work page 2005