<p>This study proposes three types of Long Short-Term Memory (LSTM) networks to predict the dynamic time history response of a nonlinear soil-structure interaction system. Three networks, the traditional LSTM network, the Bidirectional LSTM (Bi-LSTM) network, and the attention-based LSTM (Att-LSTM) network, are used in this study. A total of 21 real ground motions are used to construct a dataset: 11 for training the models, 4 for model validation, and 6 for prediction. The network-predicted results are compared with simulated results using the evaluation metrics R-squared, normalized root mean square error, and normalized root mean absolute error. Results show that all three LSTM networks can predict the dynamic displacement and acceleration time-history response of a nonlinear soil-structure system, generally in good agreement with the simulated results. In addition, the trained Att-LSTM model outperforms both LSTM and Bi-LSTM in terms of statistical analysis of three performance metrics, because the attention mechanism could effectively learn temporal dependencies. Performance of the three networks follows the pattern: Att-LSTM &gt; Bi-LSTM &gt; LSTM. Overall, the Att-LSTM network shows great potential for predicting nonlinear soil-structure system time history response, and could be a computationally efficient alternative with good accuracy to traditional numerical simulation.</p>