Application of the deep neural network to predict dynamic responses of stiffened plates subjected to near-field underwater explosion

Abstract

For quick prediction of nonlinear elastic-plastic responses of stiffened plates subjected to near-field underwater explosion (UNDEX), a machine learning algorithm was adopted, and two deep neural network (DNN) models composed of three hidden layers were trained based on data collected by a detailed finite element (FE) model. Firstly, a DNN model with the charge mass, the stand-off distance, and thicknesses of the plate and stiffeners as input variables was constructed to predict final plastic deformations of the stiffened plate, and influences of the optimizer, activation function, and under fitting and over fitting on the results were discussed. Then, a more complicated DNN model, which adopts the plastic deformations as another input variable and contains more neurons in hidden layers, was constructed to predict non-uniformly distributed effective plastic strain nephograms. Finally, the first DNN model with more neurons was used to predict the nonlinear displacement-time curves of the plate. By comparing with results obtained by LS-DYNA program, it is found that DNN can accurately estimate dynamic responses of the stiffened plate considering fluid-structure coupling effects, geometric and material nonlinearities. More importantly, the DNN models can capture the strain concentration at the boundary and the oscillation of the displacement-time curves.