Auto regressive neural network-driven reliability optimization in base-isolated building design

• ARNN is developed to predict response of base-isolated structures. • ARNN utilizes only a selected subset instead of the entire history as input. • Training set covers ground motions and isolation properties. • Dissipated energy is derived from ARNN displacement and Bouc-Wen force. • Efficient and accurate computation in RBDO is developed. As structural vibration control technologies advance, the use of base isolation in buildings has become more widespread, making optimal design increasingly important. This study employs time history analysis with stochastic considerations and uses a heuristic algorithm to optimize base isolation. The goal is to maximize seismic energy dissipation without exceeding displacement or force limits. To reduce the high computational cost of time history analysis, an Auto Regressive Neural Network (ARNN) is introduced to generate complete response histories. Trained on various ground motions and Nonlinear Time History Analysis (NLTHA) data, the ARNN produces full response histories, unlike prior methods focused solely on peak responses. A Single Degree of Freedom (SDOF) example demonstrates ARNN accuracy, and a four-story base-isolated building is used for a Reliability-Based Design Optimization (RBDO) case study. In the RBDO, Peak Ground Acceleration (PGA), earthquake frequency, and bearing properties are treated as random variables. The ARNN's efficiency enables high-accuracy RBDO with NLTHA-level precision and acceptable computation time. The supporting source codes of proposed ARNN are available at https://github.com/johnthedy/Auto-Regressive-Neural-Network .