Reliability analysis of high-dimensional models using low-rank tensor approximations

TL;DR

This paper explores the use of low-rank tensor approximations as an efficient meta-modeling technique for high-dimensional reliability analysis, demonstrating improved tail probability estimation over traditional methods in complex engineering models.

Contribution

It introduces low-rank tensor approximations as a scalable and effective meta-modeling approach for high-dimensional reliability analysis, outperforming sparse polynomial chaos in key applications.

Findings

Low-rank tensor approximations outperform polynomial chaos in tail probability estimation.

The method scales linearly with input dimension, making it suitable for high-dimensional problems.

Provides a full probabilistic description of the model response, enabling diverse statistical analyses.

Abstract

Engineering and applied sciences use models of increasing complexity to simulate the behaviour of manufactured and physical systems. Propagation of uncertainties from the input to a response quantity of interest through such models may become intractable in cases when a single simulation is time demanding. Particularly challenging is the reliability analysis of systems represented by computationally costly models, because of the large number of model evaluations that are typically required to estimate small probabilities of failure. In this paper, we demonstrate the potential of a newly emerged meta-modelling technique known as low-rank tensor approximations to address this limitation. This technique is especially promising for high-dimensional problems because: (i) the number of unknowns in the generic functional form of the meta-model grows only linearly with the input dimension and (ii) such approximations can be constructed by relying on a series of minimization problems of small size independent of the input dimension. In example applications involving finite-element models pertinent to structural mechanics and heat conduction, low-rank tensor approximations built with polynomial bases are found to outperform the popular sparse polynomial chaos expansions in the estimation of tail probabilities when small experimental designs are used. It should be emphasized that contrary to methods particularly targeted to reliability analysis, the meta-modelling approach also provides a full probabilistic description of the model response, which can be used to estimate any statistical measure of interest.