Chapter 24 - Overview and comparison of reliability analysis techniques based on multifidelity Gaussian processes

Abstract

Reliability analysis in the case of complex systems is often a challenging task. Such a study requires repeated calls to computationally intensive numerical solvers. These simulation codes aim at modeling the system behavior. However, assessing the reliability of a system with respect to uncertainties is an essential step in the design process. Many methods have been investigated in the literature, allowing to estimate failure probabilities in the case of expensive high-fidelity solvers. Among them, the use of surrogate models built by active learning is a popular technique. As these methods are based on a potentially large number of evaluations of the high-fidelity codes, their associated computational cost can still be unaffordable. Thus, the use of lower fidelity solvers in the construction of surrogate models can be a relevant way to further reduce this computational cost. In this chapter, it is proposed to study different reliability analysis methods based on multifidelity surrogate models built using active learning. The performance of the considered approaches is evaluated with respect to two axes: the construction of the multifidelity surrogate models and the enrichment criterion. Different combinations between surrogate models and criteria are evaluated on four test cases. First, two analytical problems presenting different types of relationship between the fidelity solvers (linear and nonlinear) are investigated. Then, two simplified-physics problems relative to hydraulic and aerospace engineering are studied. Eventually, the identified most efficient approach is applied on an aerospace industrial test case relative to the aerodynamic analysis of a strut-braced wing configuration.