Adaptive artificial neural networks for seismic fragility analysis

In seismic probabilistic risk assessment, fragility curves are used to estimate the probability of failure of a structure or its critical components at given values of seismic intensity measures, e.g. the peak ground acceleration. However, the computation of the fragility curves requires a large number of time-consuming mechanical simulations with the finite element method (FEM). To reduce the computational cost, in this paper a statistical metamodel based on artificial neural networks (ANNs) is constructed to replace the FEM model. An adaptive ANNs learning strategy, aimed at prioritizing the data close to the limit state of the structures, is proposed in order to improve the design of experiments for the fragility analysis. The adaptive learning strategy is developed and tested on a nonlinear Takeda oscillator.