MFFSOM-Net: Multi-Frequency Fusion Subspace Optimization Network for Inverse Scattering
Year: 2026
Type: article
Abstract: This letter proposes a multi-frequency fusion subspace-optimization network, termed MFFSOM-Net, for full-wave electromagnetic inverse scattering. Unlike single-frequency learning-based schemes that update the contrast from limited spectral information, the proposed network performs simultaneous multi-frequency induced-current learning within a physics-unrolled architecture. At each stage, the induced currents at different frequencies are restored by frequency-specific current restoration blocks, while the corresponding total fields and permittivity maps are derived through the electric-field integral equation. The multi-frequency permittivity estimates are then fused and fed into the next unrolled stage, enabling stable information propagation between current restoration and contrast reconstruction. This design combines the nonlinear representation capability of neural networks with explicit electromagnetic consistency constraints. Numerical tests on MNIST, polygon, and Austria profiles, together with experimental Fresnel data, demonstrate that MFFSOM-Net provides sharper boundaries, fewer artifacts, and improved robustness compared with benchmark methods.
Authors: Haonan Wang, Zicheng Liu, Bo Qi, Yali Zong, Peixian Han
Institution: Northwestern Polytechnical University
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204090794