Study of the Influence of the Initial a Priori Training Dataset Size in the Efficiency and Convergence of Surrogate-Based Evolutionary Optimization

The development of an automatic geometry optimization tool for efficient aerodynamic shape design, supported by Computational Fluid Dynamic (CFD) methods is nowadays an attractive research field, as can be observed from the increasing number of scientific publications during the last years. Surrogate-based global optimization methods have demonstrated a huge potential to reduce the actual number of CFD runs, and therefore drastically speed-up the design process. Nevertheless, surrogates need initial high fidelity data sets to be built and to reach a proper accuracy. This work presents a study on the influence of the initial training dataset size in the proposed approach behavior. This approach is based on the use of Support Vector Machines (SVMs) as the surrogate model for estimating the objective function, in combination with an Evolutionary Algorithm (EA) and an adaptive sampling technique focused on optimization called the Intelligent Estimation Search with Sequential Learning (IES-SL). Several number of training points have been fixed to check the convergence, the accuracy and the objective function reached by the method.