Deep learning for airfoil aerodynamic-electromagnetic coupling optimization with random forest

Abstract

Reducing the design variable space is crucial in multi-objective airfoil profile optimization to improve optimization efficiency and reduce computational costs. Based on random forest and deep neural networks (DNNs), this work performs range reduction on ten design variables obtained through a fourth-order class shape transformation parameterization method for subsonic airfoil profiles. Three aerodynamic performance objectives (lift coefficient, drag coefficient, and lift-to-drag ratio) are evaluated using the Reynolds-averaged Navier–Stokes equations, and two radar stealth performance objectives (horizontal and vertical polarization radar cross sections) are assessed through the method of moments. By combining a DNN architecture with an improved regression prediction capability, predictive models are trained for mapping design variables to design objectives. The prediction errors are below 3% for the aerodynamic performance design objectives and below 1% for the stealth performance design objectives. The particle swarm optimization algorithm provides optimized airfoil profiles for three scenarios. First is a higher lift coefficient with a lower radar cross section. Second is a lower radar cross section. Third is a higher lift coefficient. Increasing the airfoil curvature and reducing the maximum thickness improves the lift coefficient by 386 counts and reduces the drag coefficient by 17 counts. By curving the airfoil leading edge, the radar cross section for the transverse electric and transverse magnetic polarizations decreased by 2.78 and 2.09 dBsm, respectively.