Efficient global multidisciplinary optimization based on surrogate models

Abstract

In order to solve MultiDisciplinary Optimization problems involving computationally expensive disciplinary solvers, we propose a surrogate based approach where each disciplinary solver is replaced by an inexpensive surrogate model. As an approximation error is introduced with the use of surrogate models, a probabilistic framework is considered to manage the uncertainty propagation error from the MultiDisciplinary Analysis to the objective function of the optimization problem. This leads to the construction of an adaptive surrogate-based approach, called EGMDO for E_cient Global Multidisciplinary Optimization, to solve MDO problems. The approach is based on the adaptive enrichment of the disciplinary surrogate models in areas where the optimum is likely to be. Two applications are considered with a signi_cant reduction in terms of number of calls to the expensive disciplinary solvers: a reduction by up to a factor of 30 is observed on the Sellar test problem and in a preliminary aircraft design test case.