Abstract: “When estimating low damage probability of a building to the ten raise to the power minus six level, even with subset MCMC, response analysis of over 150,000 times is necessary. It requires much CPU time and therefore unrealistic. In this paper, once after reducing the number of response analysis without using subset method, we create a n dimensional regression hypersurface (a deep learning regression model) that predicts low damage probability of a building. This enables quantitative estimation with dramatically less effort.”

Keywords: seismic low damage probability, deep learning, importance analysis, gradient boosting decision tree, Keras, TensorFlow