Knowledge-Driven Adaptive Sequential Modeling and Prediction of Aerodynamic Characteristics

In the early stage of aircraft design, it is of great significance to make rapid and accurate evaluation of aerodynamic characteristics. Compared with the traditional high-cost or inaccurate evaluation methods, such as wind tunnel test, CFD simulation and engineering calculation, surrogate model has the advantage of low cost and high speed, but its prediction accuracy depends on the training samples, which is obtained from expensive evaluation. The traditional one-stage sampling method is blind, which may cause unnecessary evaluations. Adaptive sequential sampling can effectively reduce the number of sample points, but it may lead to neglect of complex regions and reduce the confidence of the overall prediction. Therefore, we proposed a novel sequential method which reweights the sampling point utility by experts’ prior knowledge of vague understanding for complex/simple regions. By comparing with Latin Hypercube and general adaptive sequential sampling method, the effectiveness of the method is verified on vertical force increment and axial force increment. It is proved that our method can obviously reduce the number of sampling points which provides a basis of aerodynamic characteristics prediction. Moreover, it improves the accuracy of complex regions with only a slight negative effect on simple regions.