Journals Proceedings

Abstract

The modern engineering design process often employs computer simulations to evaluate candidate designs. This setup transforms the design process into an optimization problem which involves a computationally-expensive black-box function, namely, which lacks an analytic expression and where each function evaluation requires large computational resources. An additional challenge in such settings is that simulation runs may consistently fail for some specific candidate designs, but with the reason for failure being unknown. To effectively handle such challenging problems this paper proposes an engineering optimization framework which incorporates a classifier whose goal is to predict if a candidate design is likely to result in a failed simulation run. This prediction is then used to dynamically divert the optimization search away from such designs, without sending them to the simulation. Numerical experiments based on an airfoil optimization problem show the effectiveness of the proposed approach. (Abstract)