Long short-term memory networks as emulators for finite element models of nonlinear structural dynamic systems: part 2—results and perspectives

Abstract Building on the theoretical developments presented in Part 1, this companion paper (Part 2) implements the Recursive Averaged Multistep Sequence-to-Sequence (RAMSS) framework to address the high computational cost of nonlinear time history analysis (NLTHA) of high-fidelity finite element (FE) models of structural systems. This implementation enables efficient seismic analysis and seismic performance assessment and supports the development of surrogate modeling and digital twinning for complex structural systems. Built upon a Long Short-Term Memory (LSTM) architecture, the LSTM–RAMSS framework is a data-driven machine learning (ML) surrogate model that improves predictive performance through an optimized convolutional autoencoder (CAE)-based seismic ground motion selection process, a dilation strategy, and an intermediate averaging mechanism during prediction. To quantify the effect of data selection on predictive accuracy, earthquake ground motions pre-sorted into engineering-based intensity measure ranges and selected via the CAE-based method are compared with those selected through a randomly sampled (RS) process when evaluating LSTM-RAMSS performance. The predictive accuracy and adaptability of LSTM-RAMSS are evaluated across six structural engineering benchmark problems to assess its robustness in modeling nonlinear structural response time histories under earthquake ground motions of varying intensity, frequency content, and duration. This study provides insights into the potential of RAMSS to advance digital twin technology in structural engineering and outlines future research directions and perspectives related to surrogate modeling of nonlinear structural dynamic systems.