Surrogate model for seismic response analysis of structures by combining a deep generative model of earthquake ground motions and multi-task Gaussian process regression

Abstract

Dynamic response analysis using time history earthquake ground motions as input is effective for detailed evaluation of seismic performance of structures. However, response analysis using a large number of ground motion data may be inefficient and difficult from the viewpoint of computational cost. This study proposes a surrogate model for seismic response analysis and an efficient sampling method for sampling critical ground motion time histories. Both methods are based on Generative Adversarial Networks and Gaussian process regression, which enables utilization and management of high-dimensional time series of ground motions. Numerical experiments demonstrated that the surrogate model successfully predicted response analysis results with reasonable accuracy. In addition, by combining the above models with a Markov chain Monte Carlo method, it was possible to efficiently sample ground motion time histories capable of causing significant damage to structures.