Time-dependent performance of large-scale dome structures subjected to earthquakes using a machine learning-based evaluation method

The global time-dependent performance analysis of structural systems has aroused the interest of researchers in recent years. The conventional technique using finite element analysis is very computationally intensive in evaluating the time-dependent performance of structures with multiple variables. A new hybrid method by integrating a surrogate model with the subset simulation is proposed to evaluate the time-dependent performance of large-scale complex structures subjected to earthquakes. The surrogate model is constructed based on the multilayer perceptron network using a small dataset and the subset simulation is used to estimate the small failure probability. Large-scale dome structures under earthquakes are used to verify the proposed method. The results show that surrogate models have a high accuracy in estimating extreme structural deformations when compared to the high-fidelity results using the finite element method. The hybrid method is efficient and accurate for the global time-dependent reliability analysis. The global sensitivity analysis for the dome structures is also conducted and the importance of variables over the life cycle is evaluated using the surrogate models. The numerical examples demonstrate that the method is highly efficient and accurate in quantifying uncertainty in the time-dependent performance of large-scale complex structures.