NEURAL NETWORK ALGORITHM FOR NONLINEAR STRUCTURAL SEISMIC RESPONSE

A new method based on the long short-term memory (LSTM) neural network model is proposed for calculating seismic responses of nonlinear structures. It adopts a unidirectional multilayer stacked LSTM architecture and recursively calculates structural responses using a sliding time window. Updated accuracy evaluation indexes are suggested to take into account the sensitivity difference in different response amplitude ranges and avoid the phase sensitivity issue of traditional ones. The new method is validated by multi-layer frame structures subjected to measured ground motions, the principles for selecting network hyperparameters are given, and the generalization ability of the method for different conditions is discussed. The results indicate that the LSTM model achieves good computational accuracy and is robust to various type of ground motions. Due to the cloud deployment feature of the neural network model, the new method is expected to contribute to application scenarios where traditional numerical methods are limited, such as rapid simulation of seismic response in urban areas.