TINN: Time-step integration neural network for dynamic analysis of nonlinear structure systems subjected to seismic excitation

Abstract

Abstract Dynamic analysis is an essential method for evaluating the seismic response of structures and predicting their seismic damage. The traditional time-step integration methods are computationally time-consuming and sensitive to structural complexity. This study proposes a novel time-step integration neural network (TINN) to perform dynamic analysis for nonlinear structures. TINN is composed of a fully connected feed-forward network and a residual network (ResNet) to simulate the calculation procedure of a single time step in the time-step integration method, and the recursive mechanism is applied to calculate the time history of structural responses. In addition, several techniques are applied to reduce the cumulative error due to the recursive mechanism to improve the accuracy and stability of TINN. TINN could accurately predict the seismic responses of all degrees of freedom (DOF) of the structures, including displacement, velocity, acceleration, and internal force. In particular, the efficiency of TINN doesn't decrease with the increase of structural DOF and the degree of nonlinearity, and it has no restrictions on the amplitude and duration time of the ground motion records. The results on three numerical cases show that TINN performs well in predicting the structural seismic responses with very limited training data available.