A deep learning approach to inverse scattering analyses: Recovering interfacial defects in laminated structures

Abstract

We propose a deep learning (DL) regression strategy to solve inverse acoustic scattering problems. This strategy is capable of recovering heterogeneous defect fields in any interface and direction of composite laminates. The training procedure of the neural network (NN) model uses stochastic Gaussian fields as output, which are related to interfacial damage fields of the physical problem. We assume prior knowledge of the material properties of the ultrasound incident field and the elastic layers’ properties. Furthermore, we model the interfaces using the Quasi-Static-Approximation, a method that generates position-dependent interfacial stiffness matrices, composed of a set of uncoupled normal and tangential springs. We validate our approach and evaluate its performance for noisy input data and reduced-order models. Additionally, we also consider model errors at the composite interfaces. The obtained results show that the presented method is promising due to its generalization capability in recovering different defect field profiles, its robustness in relation to noisy data and model errors and its low computational cost, once the DL model is trained, in comparison to traditional inverse approaches (iterative).