Metamodeling-Based Reliability Analysis of Structures Under Stochastic Dynamic Loads with Special Emphasis to Earthquake

Abstract

Monte Carlo Simulation (MCS)-based structural reliability analysis (SRA) approach allows more realistic safety assessment of structures. However, it involves large number of dynamic analyses of structure making it computationally challenging. Metamodeling technique is found to be useful in this regard. The effectiveness of various metamodeling approaches, namely, the polynomial response surface methods, artificial neural network, kriging and support vector regression are demonstrated to approximate nonlinear dynamic response and SRA under stochastic dynamic loads with special emphasis to earthquake loads. In doing so, the metamodel is constructed directly for response approximation for each realization of the stochastic load in a selected bin. The approach does not require to assume structural response apriorias necessary in case of conventional dual response surface approach without additional computational burden. Once the metamodels are obtained, the MCS can be easily performed by obtaining random sample of input parameters and selecting metamodels randomly from the suit of metamodels. The random selection of metamodel implicitly considers the random nature of earthquake to closely follow the conventional notion to reflect the record-wise variation of ground motion. The effectiveness of the proposed direct response approximation approach in nonlinear dynamic response approximation and subsequent estimate of reliability by the various metamodels are demonstrated numerically.