Probabilistic Deep Learning Solutions to Electromagnetic Inverse Scattering Problems Using Conditional Renormalization Group Flow

Deep learning (DL) has demonstrated great potential in solving nonlinear electromagnetic inverse scattering (EMIS) problems. However, the lack of uncertainty quantification and reliability assessment in the predictions remains an unresolved challenge. Overcoming this challenge is of critical importance in practice because of the presence of uncertainties and noise in the measurement data and the ill-posedness of EMIS problems. Here, we propose a conditional renormalization and generation-flow (cRG-Flow) network, a reversible hierarchical network, and develop a novel probabilistic DL strategy to quantify the uncertainty in DL solutions of EMIS problems. The proposed strategy provides two other important advantages over the existing DL approaches for solving EMIS problems. First, due to the reversible property of cRG-Flow, both the forward EM scattering computation and its inverse procedure can be considered simultaneously in the network and, as a result, physical constraints on the DL solutions can be easily imposed. Second, the interpretability of DL solutions is facilitated due to the utilization of RG-Flow. Illustrative examples are provided to demonstrate the performance gain in terms of reconstruction quality and robustness. The proposed method may pave a new avenue for addressing the EMIS problems based on DL schemes with improved reliability and interpretability.