Active learning for efficient data-driven aerodynamic modeling in spaceplane design

Abstract

Data-driven surrogate model is extensively applied in the aerodynamics prediction for flight vehicle design. However, for three-dimensional problems, modeling costs hugely in acquiring adequate data. Adaptive and efficient sampling represents a promising approach, so we employ Active Learning (AL) to assess the sampling space. An aerodynamics prediction framework based on the Loss for Active Learning (LLAL) is proposed for capturing the sample insufficient space. The LLAL-based method refines the model by estimating the biases between the ground truths and the predictions as a measure of the high-value samples, then selecting top-K samples for infilling. We first validate our proposed method using an analytical benchmark two-dimensional function, followed by its application to aerodynamics prediction of spaceplane with the number of 60 and the deformation dimensions of 18. Through using both the Latin Hypercube Sampling and LLAL-based methods to infill samples, we observed the R2 of lift-to-drag ratio improves from 0.82 to 0.85. The AL method can enhance the accuracy of models with a limited number of samples, thereby reducing sampling costs and improving the efficiency of aerodynamic design.