Damage prediction of stiffened plates subjected to underwater contact explosion using the machine learning-based method

Abstract

A machine learning-based method is used to predict damage responses of stiffened plates subjected to underwater contact explosion. The data set of damage responses of a stiffened plate with different thicknesses under different charge masses and standoff distances is collected from a detailed finite element model established by the LS-DYNA program. In order to efficiently obtain the optimal machine learning model, influences of the number of hidden layers, the number and distribution of hidden layer neurons on the network performance are investigated. Due to fluid-structure interaction, material and geometric nonlinearities, it is found that the three hidden layers network is required to predict damage responses of the stiffened plates, even for the prediction of the dimension of the damaged domain which has only one target parameter. Meanwhile, the number of multi-hidden layer neurons can be determined by the empirical formulas coupled with the trial-and-error method, and the rarely studied hidden neuron distribution can be further selected based on the quantitative relationship between input features and target parameters. The damage border, the dimension of the damaged domain, and plastic deformations of the undamaged domain predicted by the optimal networks are in good agreement with the LS-DYNA results.