Abstract
The inverse medium problem, inherently ill-posed and nonlinear, presentssignificant computational challenges. This study introduces a novel approach byintegrating a Neumann series structure within a neural network framework toeffectively handle multiparameter inputs. Experiments demonstrate that ourmethodology not only accelerates computations but also significantly enhancesgeneralization performance, even with varying scattering properties and noisydata. The robustness and adaptability of our framework provide crucial insightsand methodologies, extending its applicability to a broad spectrum ofscattering problems. These advancements mark a significant step forward in thefield, offering a scalable solution to traditionally complex inverse problems.