Deep Tensorized Fourier Neural Operator for Electromagnetic Inverse Scattering Problems

This paper proposes an innovative deep learning framework, DeepTFNO, designed to efficiently address the problem of electromagnetic inverse scattering (ISP). The framework draws inspiration from the design principles of DeepONet and Tensor Fourier Neural Operator (TFNO), constructing a multimodal feature extraction structure composed of a trunk network and branch networks, with feature fusion enabling effective information integration. The trunk network leverages TFNO to capture the spatial dependencies of the global permittivity distribution, ensuring high accuracy in external scattering reconstruction. Meanwhile, the branch networks utilize dual-channel TFNO to process the real and imaginary components of the scattered fields, effectively extracting internal multiple scattering information. Furthermore, the framework incorporates a Fusion Attention Module to dynamically adjust feature weights prior to fusion, enhancing the flexibility and adaptability of information integration. Experimental results on datasets with varying complexity demonstrate that DeepTFNO outperforms existing methods in terms of reconstruction accuracy and generalization capability, providing an efficient and reliable solution for ISP.