Using neural networks to predict vortex-panel analyses : a feasibility study

Abstract

This thesis studies the feasibility of using neural networks to "learn" the vortex panel method. This study is motivated by the desire for the rapid and accurate prediction of fluid flows during the preliminary design of engineering systems, where traditional computational fluid dynamics (CFD) are too computationally costly. The results show that a two-layer neural network can estimate the pressure coefficient and elements in the vortex-panel influence-coefficient matrix. However, when the neural-network-predicted influence-coefficient matrix is used to estimate the pressure coefficients, the results are in poor agreement with the baseline prediction, although general trends are captured.