Adaptative Kriging for Reliability Analysis of Sequential Simulation Models of Fatigue-Loaded Systems

Abstract

Sequential numerical model chains are often used in industrial analyses. For example, structural durability analysis involves first finite element simulations followed by fatigue postprocessing. The prohibitive computational cost of these entire numerical chains generally makes reliability assessment unfeasible unless strong simplifications are made. A new active learning Kriging method for sequential models (AK-SM) is proposed here to overcome the computational burden. AK-SM introduces a novel enrichment strategy within the well-known Active Kriging Monte Carlo Simulation framework, leveraging the sequential nature of the performance function. An imputation criterion based on a local functional decomposition and Kriging prediction variance is designed to selectively bypass costly evaluations of the first models and prioritize the more affordable postprocessor. The analysis of an analytical creep-fatigue interaction problem first, and of a modal transient finite-element fatigue of a bracket then, show that AK-SM is particularly suited to sequential numerical chains by achieving significant computational savings while maintaining a high accuracy in the failure probability estimation.