A recurrent neural network model for structural response to underwater shock

Abstract

Simulating the dynamic response of structures to shock involves numerical models that can require significant effort and computational resources to create and solve. This can make accurate naval platform vulnerability analyses involving a large number of simulations impractical and motivates the development of efficient surrogate models for shock and blast effects. This paper presents a novel approach to predicting the response of a floating structure to underwater shock based on machine learning (ML). Velocity time-series for a rigid floating plate subjected to underwater shock and cavitation effects are calculated for a range of charge mass, standoff distance and plate mass per unit area values using a coupled Eulerian–Lagrangian (CEL) numerical model. The computed motions are used to train a recurrent neural network (RNN) to predict the plate response including the effect of reloading due to the cavitation closure pulse. The RNN predicts plate responses from a test set of CEL model instances with a mean value of mean-squared errors between the ML and CEL model velocities, normalized with respect to the velocity ranges, of 1.5 × 10-4.