Support vectors machines for the estimation of probability of failure: Multifidelity classifiers built from a posteriori discretization error estimators. (May 2023)

• Record Type: Journal Article
• Authors: Mell, Ludovic; Rey, Valentine; Schoefs, Franck
• Is Part Of: Structural safety. Volume 102(2023)
• Publication Date: 2023-05
• Subjects: Reliability -- Probability of failure -- Support vector machines -- Finite element method -- Discretization error; Structural stability -- Periodicals; Safety factor in engineering -- Periodicals; Reliability (Engineering) -- Periodicals; Constructions -- Stabilité -- Périodiques; Coefficient de sécurité en ingénierie -- Périodiques; Fiabilité -- Périodiques; 620.86
• DOI: 10.1016/j.strusafe.2023.102321
• Languages: English
• ISSNs: 0167-4730

Abstract

Abstract: This article proposes two algorithms to compute the probability of failure using Support Vector Machine (SVM) classifiers and Monte Carlo estimator in the context of structural mechanics. The observations used to build the classifiers are obtained from calls to a finite element solver which introduces discretization error. By exploiting guaranteed discretization error estimators, the proposed methodology aims at computing an estimation of the probability of failure not polluted by this discretization error. The first algorithm builds two classifiers in parallel to separate the guaranteed fail population, the guaranteed safe population and the uncertain population. It enables to compute an upper bound and a lower bound of the exact probability of failure. The second algorithm only uses observations whose status (fail or safe) is guaranteed by the error estimators. It results in multi-fidelity SVM-based meta-model as observations computed on different mesh sizes can be used. Those two algorithms are illustrated on two-dimensional mechanical examples for different Monte Carlo populations. Highlights: Population for which classification is polluted by discretization error is exhibited Multi-fidelity support vector machine classifier is built from different meshes Bounds on the exact probability of failure are derived.