ABSTRACT
ABSTRACT: Reliability-based design optimization has received much attention in the last decade since it allows one to combine design optimization techniques with probabilistic performance criteria that take into account the unavoidable uncertainties inherent to engineering. From an industrial point of view though, this approach is still not applicable because the proposed resolution schemes rely upon simplifying assumptions (e.g. the first-order reliability method) that might not hold in practice; or in constrast, they require a large number of simulation runs (e.g. simulationbased reliability methods) which is not compatible with the time-consuming models used in the industry. This latter reason motivated the emergence of surrogate-based strategies. Surrogate models are analytical functions that are much faster to evaluate than the original model (e.g. a finite element model). They are usually built from a set of evaluations of the original model that is called a design of experiments. This paper makes use of a probabilistic surrogate known as kriging and proposes an adaptive refinement technique that gradually builds the design of experiments according to the required accuracy. This technique is then encapsulated in a nested reliabilitybased design optimization loop where the optimizer performs simulation-based reliability and reliability sensitivity analyses. For the sake of efficiency, the simulation technique that is used here is a variance reduction technique known as subset simulation.
