A method to improve design reliability using optimal Latin hypercube sampling

Abstract

In the paper an algorithm for design reliability improvement is proposed. Its key part consists in the computation of the correlations between constraint functions and design variables which are subsequently used to find the new design iteration. It is shown that the optimal Latin hypercube (OLH) sampling provides an extremely efficient technique for assessing the values of correlation coefficients. Since finding the large OLH designs is not a trivial task, a study on the OLH generation algorithms was performed.Two algorithms were found to be particularly effective, namely, the columnwise-pairwise algorithm and the genetic algorithm. The presented strategy proves to be especially useful when alternative gradient-based methods cannot be used, which is often the case for computationally expensive problems involving noisy and highly nonlinear responses. The method is best suited for problems where the probability of failure for the initial design is large and the main interest is to find a more reliable design rather than the optimal one in the sense of reliability-based optimization. The method is illustrated with two numerical examples. One model example and one concerning the problem of thin-walled beam crash.