![[2602.13805] Fast Physics-Driven Untrained Network for Highly Nonlinear Inverse Scattering Problems](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.13805] Fast Physics-Driven Untrained Network for Highly Nonlinear Inverse Scattering Problems
Summary
This paper presents a novel Real-Time Physics-Driven Fourier-Spectral solver for electromagnetic inverse scattering, achieving significant speed improvements in reconstruction times.
Why It Matters
The advancements in this research address computational limitations in high-dimensional optimization for inverse scattering problems, enabling faster and more accurate real-time applications in fields such as microwave imaging. This could lead to enhanced capabilities in medical imaging and other technologies reliant on electromagnetic wave analysis.
Key Takeaways
- Introduces a Physics-Driven Fourier-Spectral solver for inverse scattering.
- Achieves a 100-fold speedup over existing untrained neural networks.
- Utilizes a contraction integral equation to handle high-contrast nonlinearity.
- Implements a bridge-suppressing loss to improve boundary accuracy.
- Demonstrates robust performance under noise and antenna uncertainties.
Computer Science > Machine Learning arXiv:2602.13805 (cs) [Submitted on 14 Feb 2026] Title:Fast Physics-Driven Untrained Network for Highly Nonlinear Inverse Scattering Problems Authors:Yutong Du, Zicheng Liu, Yi Huang, Bazargul Matkerim, Bo Qi, Yali Zong, Peixian Han View a PDF of the paper titled Fast Physics-Driven Untrained Network for Highly Nonlinear Inverse Scattering Problems, by Yutong Du and 6 other authors View PDF HTML (experimental) Abstract:Untrained neural networks (UNNs) offer high-fidelity electromagnetic inverse scattering reconstruction but are computationally limited by high-dimensional spatial-domain optimization. We propose a Real-Time Physics-Driven Fourier-Spectral (PDF) solver that achieves sub-second reconstruction through spectral-domain dimensionality reduction. By expanding induced currents using a truncated Fourier basis, the optimization is confined to a compact low-frequency parameter space supported by scattering measurements. The solver integrates a contraction integral equation (CIE) to mitigate high-contrast nonlinearity and a contrast-compensated operator (CCO) to correct spectral-induced attenuation. Furthermore, a bridge-suppressing loss is formulated to enhance boundary sharpness between adjacent scatterers. Numerical and experimental results demonstrate a 100-fold speedup over state-of-the-art UNNs with robust performance under noise and antenna uncertainties, enabling real-time microwave imaging applications. Subjects: Machine Lear...
