DeepDPO: A Self-Supervised Deep Learning Approach Based on the Differentiable Physical Optics Method for 3-D Reconstruction of Perfect Electric Conductors

This communication presents a novel deep learning (DL) framework for solving 3-D inverse scattering problems (ISPs), aimed at reconstructing geometric shapes of perfect electric conductors from scattered electric fields. ISPs are inherently ill-posed and non-linear, challenging both conventional and DL methods. To address this, we propose DeepDPO, a self-supervised DL approach that integrates the reliability of conventional model-based methods with the efficiency of learning-based approaches. The framework employs a probabilistic auto-decoder neural network (PADNN) for continuous geometric representation using a signed distance field and incorporates the mesh-based physical optics (PO) method for forward scattering calculations. The reconstructed model is obtained by decoding the object-specific code through the PADNN and optimizing it by minimizing residuals between computed and actual fields. By compressing complex 3-D geometries to compact codes, the PADNN alleviates ill-posedness, while the PO approximation mitigates non-linearity. Consequently, DeepDPO achieves faster convergence, enhanced accuracy, and eliminates the need for hard-to-obtain electromagnetic data in pre-training. Numerical results demonstrate notable improvements in computational efficiency, accuracy, and the ability to represent complex objects, offering a promising solution for 3-D ISPs with the potential for further integration of physical principles.