Neural network-motivated regularity analysis of inverse problems and application to inverse scattering

We derive Lipschitz regularity estimates for an approximate inverse of a general compact operator that depends non-linearly on a vector parameter. The regularity estimates are motivated by the desire to develop neural networks (NN) that compute that approximate inverse and the convergence of the NNs follows from the Lipschitz regularity estimates. Such compact operators arise in inverse wave scattering applications with unbounded domains, and we illustrate our theory by showing that the particular assumptions of our regularity analysis hold for the problem of identifying cracks from far field data. Numerical results using a NN for parameter recovery demonstrate the accuracy, efficiency and robustness of our approach.