Failure probability under uncertain surrogate model predictions

In current engineering practice, surrogate models are increasingly applied for the substitution of computationally demanding numerical models in the context of predicting failure probabilities. However, while computationally very efficient, these models in general do not predict the same model responses as compared to the real numerical model. As such, when these surrogate models are applied in the context of estimating the failure probability of the structure, this estimate is uncertain and affected by the precision of these surrogates. Two commonly applied surrogate models directly provide, next to a nominal response, also an estimate of the uncertainty that is attributed to this response: Kriging and Interval Predictor Model. The goal of this paper is to objectively compare these two methods in terms of deterministic accuracy, the conservatism of the estimate of the uncertainty and the computational cost that is needed to construct and evaluate the surrogate model.

Jaar van publicatie:2018

Toegankelijkheid:Open