A Physics-Based Deep Learning to Extend Born Approximation Validity to Strong Scatterers

In this study, we present a novel approach to address nonweak scattering problems by integrating deep learning (DL) into the Born series. Typically, the first-order Born approximation (BA) is limited to cases where the contrast between the scatterer and the background medium is exceptionally low. While higher-order terms in the Born series can be used for higher contrasts, convergence issues may arise due to highly oscillatory factors in Green's function. To overcome this limitation, we introduce a physics-based DL method inspired by the Born series, which effectively predicts the distribution of the complex electromagnetic field. The proposed series assures convergence when scattering occurs from high-contrast objects, due to the use of a learning-based forward operator. Exploiting the physics-based nature of our model, we adopt a simple convolutional neural network (CNN) architecture, requiring significantly fewer training data. Our results demonstrate very good generalization capabilities of the proposed approach, showcasing its ability to handle unseen background fields and profiles. We deem this innovative series as an extension of the Born series that can be effectively employed in highly nonlinear problems.