Modular surrogate-based optimization framework for expensive computational simulations

Abstract

In practical applications, the use of computational modeling has been industry-wide adopted to speed up product development as well as reduce physical testing costs. Such models of complex or large systems are, however, often computationally expensive, hence solution times of hours or more are not uncommon. Additionally, as these models are typically evaluated using blackbox solvers, the direct study of relations between design parameters renders demanding in terms of computational time and provides poor engineering insight and understanding. To address this, a modular framework integrating computation automation with the use of surrogate-based modeling, optimization and visualization techniques is presented. The framework is built in the Python programming language. Its use is illustrated on a study of the side impact response of a car body using an artificial neural network as a surrogate together with the NSGA-III genetic algorithm for optimization.