Machine learning-enhanced fragility curves: Advancing reliability and safety of structures in seismic risk assessment

Fragility curves are essential in seismic risk assessment and performance-based design in structural engineering. The most accurate method to create these curves is through extensive Non-linear Time History Analysis (NLTHA) at various seismic intensities, assessing reliability across different PGAs. However, traditional fragility curves, constrained by computational costs, often oversimplified. This research introduces an innovative Autoregressive Neural Network (ARNN) for predicting structures’ time-history response during earthquakes, enabling more efficient fragility curve generation through cost-effective Monte Carlo Simulation (MCS). The ARNN’s unique input layer, which includes modal analysis to extract structural periods, windowed earthquake data, and structural responses, enables the handling of multiple structural parameters. Additionally, ARNN allows a single time history record to be partitioned into multiple training data sets, enhancing the efficiency of the machine learning. Differing from traditional fragility curves, this approach considers uncertainties in both ground motion and structural components, requiring 10–20 NLTHA records for ground motion alone and 125 to 300 records when considering both uncertainties. This methodology’s effectiveness is demonstrated through three numerical examples, including a nonlinear column, a damper-equipped structure, and a base-isolated building, significantly enhancing structural reliability and safety in seismic evaluations.