An Improvement of a Metamodel-Based Importance Sampling Algorithm for Estimating Small Failure Probabilities

This work presents an improvement to a Monte Carlo-based algorithm of literature for the estimation of small failure probabilities. The original algorithm of literature is based on the implementation of the optimal importance density by a surrogate, kriging-based metamodel approximating the response function which determines the failure by reference with a given threshold. The significant improvement of the algorithm is obtained by a more effective training of the metamodel, and allows for a further decrease of the computational efforts required in the failure probability estimation. The performance of the new algorithm is demonstrated on a few analytic examples of literature.