Response Surface Methods for Efficient Aerodynamic Surrogate Models

摘要

Response surface methods for the approximation of outputs of computer experiments such as the Kriging method often suffer from a lack of accuracy or efficiency. Many computationally expensive samples are needed for the globally correct reproduction of an unknown response. We investigate adaptive sampling strategies, which can automatically identify critical regions of an input-parameter domain and require less samples than traditional one-stage approaches like Latin hypercube designs. Furthermore, we propose a new method which makes use of the assumption that the aerodynamic responses are not of arbitrary structure, but rather related to other instances of a mutual problem class. Both approaches are validated with numerical test cases, showing that they produce more accurate surrogate models using less samples than traditional approaches.