Deep Learning-based Metamodeling Technique for Nonlinear Seismic Response Quantification

Abstract

Classical physics-based numerical techniques such as finite element method (FEM) usually takes a huge computational resource and time for simulation-based uncertainties in structural analysis. Especially, it is expensive when nonlinear time history analysis is involved. The metamodeling technique becomes an alternative with an ability to predict the time history response of both elastic and inelastic structural system in a data-driven fashion. Various machine learning and deep learning-based techniques have been attempted to have their limitations in handling huge and sequential data while it comes to predict the whole response time history given stochastic ground motion acceleration. In this regard, Long Short Term Memory (LSTM) approach is found to be useful. The present study used LSTM deep learning network based metamodeling approach for nonlinear seismic response history approximation. The novelty of the approach lies in its capability of capturing record to record variability even for an inelastic structure. This metamodel can work with the desired level of accuracy with very limited data. The proposed approach has shown satisfactory results to approximate seismic response of a nonlinear single degree of freedom system.