Minimizing computational data requirements for multi-element airfoils using neural networks

Artificial neural networks were used to fill in a design space of computati onal data to optimize the flap position for maximum lift for a multi-elemen t airfoil. Multiple-input, single-output networks were trained using NASA A mes Research Center's variation of the Levenberg-Marquardt algorithm. The c omputational data set was generated using an incompressible Navier-Stokes a lgorithm with the Spalart-Allmaras turbulence model. An empirically-based c riteria, designated the "pressure difference rule," was applied to the trai ning set because numerical inaccuracies for the computational method were i dentified near maximum lift. The neural networks were trained with only thr ee values of the inputs: flap deflection, gap, and overlap at various angle s of attack. The entire computational data set was thus sparse, and yet by using only 52-70% of the computed data, the trained neural networks predict ed the aerodynamic coefficients within 1.7% of the maximum lift coefficient . In addition, a genetic algorithm and a gradient-based optimizer were inte grated with the neural networks to optimize the high-lift rigging. This new optimization process had a higher fidelity and a reduction in CPU time whe n compared with an optimization procedure that excluded the genetic algorit hm.