Damage prediction of ship cabins subjected to underwater contact explosion by deep neural network with grid search algorithm

The grid search (GS) algorithm combined with new input features obtained by physical equations are adopted to automatically obtain the optimal deep neural network (DNN) structure to predict damage responses of ship cabins subjected to underwater contact explosion. Databases of fractured domain and plastic deformations of cabins under different charge masses, stand-off distances, and attack angles are collected by LS-DYNA. Influences of activation functions, numbers of hidden layers, optimization ranges and distributions of neurons in hidden layers on prediction accuracy and optimization efficiency are investigated by GS algorithm. It is found that the optimal numbers of hidden layers for predicting fractured domain and plastic deformations considering both prediction accuracy and optimization efficiency are five and four, respectively, which indicates the widely used three should be extended. Then, the optimization range of neurons in multiple hidden layers can be directly defined by empirical formulas for the single hidden layer neurons without using the inefficient trial-and-error method, and the uniform distribution of hidden neurons can be used without considering non-uniform distribution. Although new input features have negligible influence on fractured domain prediction, they can significantly reduce differences between the maximum and minimum MAE of plastic deformations, which further greatly improves optimization efficiency.