A deep learning approach to predicting unsteady airfoil performance from tiny datasets

The interaction between fluid and wing surface for the generation of aerodynamic forces is a complex nonlinear phenomenon with multiple variables. Obtaining the aerodynamic performance of a body is, therefore, a complex, time-consuming and costly endeavour. Wind tunnel testing, especially for high-speed flows, requires special facilities and custom-built, high-precision body models. On the other hand, high-precision computational fluid dynamics is expensive, and the possibility of real-life mirroring Direct Numerical Simulation for flows of engineering interest remains impractical for the foreseeable future. Especially during the preliminary design phase, exploring the effect of freestream conditions on the aerodynamic performance of a profile requires multiple trials with high time and labour input. In recent years, deep learning has undergone rapid advances in various applications, from computer vision to natural language processing to time series prediction. The inherent ability to extract hierarchical complex dependencies between variables makes neural networks particularly suitable for predicting an airfoil aerodynamic performance. Therefore, in this work, we explore the possibility of predicting the principal harmonics frequency and amplitude of the aerodynamic performance of an airfoil in a high-compressible transonic freestream from near-field data and free streams parameters obtained from numerical simulations. Neural network architectures based on convolutional neural networks are proposed with the goal of learning the data spatio-temporal correlations through the operation of convolution. Compared to similar works, the architectures of the regression models evaluated here are inspired by well-known architectures deployed in the field of image classification and computer vision. The hypothesis is then tested that well-tested computer vision-oriented architectures, due to their feature-extraction capabilities, can be applied to data with different characteristics.