Seismic reliability assessment of base-isolated steel structures via Autoregressive Neural Network and Pearson distribution method

Abstract

In seismic structural engineering, different methods are used to evaluate performance. Simplified approaches provide conservative estimates, while advanced analyses achieve higher precision at the cost of significant computational effort. Nonlinear Time History Analysis (NTHA) remains the most reliable method, but its high computational demand has led many researchers to propose simplified models, often resulting in conservative outcomes. This study proposes an Artificial Intelligence (AI)-based method to approximate NTHA. An Autoregressive Neural Network (ARNN) is developed to generate complete time-history responses of structures with minimal error relative to NTHA. Using ground motion data and the first three fundamental periods as inputs, the ARNN replicates NTHA responses with high accuracy. Unlike conventional surrogate models that predict only peak responses, the ARNN produces the entire response history. The ARNN is further integrated with a moment-based reliability framework employing the four-moment Pearson distribution (4M-Pearson), enabling efficient and accurate seismic reliability assessment. A three-story base-isolated steel structure is analyzed as a case study. Results demonstrate that the proposed ARNN achieves high precision in predicting both structural time-history responses and seismic reliability.