Real-time model of three-dimensional dynamic reattachment using neural networks

Abstract

The reattachment of unsteady separated flow fields is a critical issue in the determination of both helicopter and wind-turbine blade performance as well as for poststall maneuvers in aircraft. To fully understand this process and to enable control, numerical techniques that provide real-time models of the reattachment process over a broad parameter range must be realized. This article describes real-time models, using neural networks, for the dynamic reattachment of three-dimensional unsteady separated flow fields. The results indicate that the neural network model accurately predicts the dynamic reattachment process to within 5% of the experimental data across the parameter space bounded by nondimensional pitch rates a of 0.01 and 0.20. However, the error was substantially larger for an a+ of 0.02. Analyses indicate that the parameter space is governed by two different sets of flow physics that transition at roughly an a of 0.03. As such, the results show that neural network models can be used not only to detect changes in the flow physics, but for defining areas within the parameter space where additional experimental characterization would be useful. Further, the results indicate that the flow field wing interactions are three dimensional, however, the spanwise effects of the three dimensionality are subdued relative to dynamic stall.