A nonlinear dynamics-informed LSTM network for response prediction of strong earthquake-excited high-rise building structures

Abstract

Long short-term memory (LSTM) networks, which have attracted increasing attention, have emerged as a promising approach for predicting structural dynamic responses, particularly including earthquake-induced time–frequency characteristics. However, due to the multi-modal effects and diverse plasticity development paths in high-rise structures under strong earthquakes, existing LSTM networks fail to explicitly and simultaneously capture such complex dynamic behaviors from a structural dynamics perspective. To address this challenge, this work develops DYNLSTM-Tall, a novel nonlinear dynamics-informed LSTM network, based on two-level mapping relationships derived from the numerical substructure method and a comprehensive time–frequency domain evaluation metric. The architecture of DYNLSTM-Tall and the combined loss function are optimized through ablation experiments across five extended datasets. On this basis, the prediction accuracy and generalization ability of DYNLSTM-Tall is demonstrated by training/validation loss convergence, quantitative evaluation metrics and comparison between predicted vs. actual earthquake response histories of test samples with varying seismic damage states. DYNLSTM-Tall’s superior performance is further validated through comparison with two state-of-the-art LSTM variants across three case high-rise structures, achieving correlation coefficients of 0.95 ∼ 0.98 and accurately identifying up to six lower vibration modes. This research underscores its strong potential as an alternative for seismic risk assessment as well as structural earthquake response prediction.