Risk-Adaptive Learning of Seismic Response using Multi-Fidelity Analysis

Abstract

Performance-based earthquake engineering often requires a large number of sophisticated nonlinear time-history analyses and is therefore demanding both with regard to computing resources and technical expertise. We develop a risk-adaptive statistical learning method based on multi-fidelity analysis that enables engineers to conservatively predict structural response using only low-fidelity analyses such as Pushover analyses. Using a structural model of a 35-story building in California and a training data set consisting of nonlinear time-history and pushover analyses for 160 ground motions, we accurately and conservatively predict maximum story drift ratio, top-story drift ratio, and normalized base shear under the effect of 40 ground motions not seen during the training.