Neural network prediction of three-dimensional unsteady separated flowfields

Abstract

Unsteady surface pressures were measured on a wing pitching beyond static stall. Surface pressure measurements confirmed that the pitching wing generated a rapidly evolving, three-dimensional unsteady surface pressure field. Using these data, both linear and nonlinear neural networks were developed. A novel quasilinear activation function enabled extraction of a linear equation system from the weight matrices of the linear network. This equation set was used to predict unsteady surface pressures and unsteady aerodynamic loads. Neural network predictions were compared directly to measured surface pressures and aerodynamic loads. The neural network accurately predicted both temporal and spatial variations for the unsteady separated flowfield as well as for the aerodynamic loads. Consistent results were obtained using either the linear or nonlinear neural network. In addition, fluid mechanics modeled by the linear equation set were consistent with established vorticity dynamics principles.