Toward a new class of hybridized reliability methods designed for efficiency and robustness with large computer models

Abstract

Reliability approaches hold a prominent place in the set of probabilistic uncertainty propagation and quantification tools used in engineering analysis. These methods have seen much rigorous development to date and are in many cases very effective. When used in conjunction with large computational mechanics computer models, however, a new regime is entered that is very different from the ones in which the methods have traditionally been developed and tested. Analysis and examples will be cited here that demonstrate some pitfalls in using classical reliability methods in this regime, and mechanisms will be discussed for avoiding or reducing the pitfalls. Three different aspects of reliability procedures will be considered: 1) the optimization procedure for finding the critical or most-probable point; 2) the limit-state surface approximation step at the MPP; and 3) the integration to determine exceedence probability.