Adaptive Surrogate Model for Failure Probability Estimation

Abstract

Estimating failure probability becomes a fundamental task in many complex engineering designs and optimizations. Yet, evaluation of failure probability via direct sampling from a given system can be computationally expensive and sometimes impossible. Although the construction of a response surface/surrogate could reduce such computational cost, reliance on its sampling alone may still yield an erroneous estimate of the failure probability. In this paper, we employ generalized polynomial chaos and develop an adaptive method whose surrogate model evolves with the additional data sampled from the underlying system as the iteration proceeds. It is more flexible by not requiring an accurate surrogate model in priori. Via three distinct numerical examples and one practical problem on a spintronic device, we demonstrate that our novel scheme provides an efficient tool to estimate system failure probability.