Efficient aerodynamics models. Viscous implementation in Panel Method using neural networks

A modified Panel Method that uses data generated by 2D aerodynamic models is studied. First, three different 2D models have been developed: incompressible, compressible and viscous. The incompressible model uses analytic solutions obtained by applying conformal transforms. The simulation of compressible flow involves the generation of a structured mesh and high-order finite differences. In the third model, Spalart–Allmaras viscosity is included with AnsysFluent. Then, these 2D models are used to generate data sets, one for each aerodynamic model, with which three different neural networks are trained. Once data has been generated, it is processed and then divided into training, validation and test data. The three neural networks are then trained using Batch Gradient Descent method. The optimal network configuration for each data set is obtained using hyperparameter tuning. Finally, a modified Panel Method is implemented using these networks for numerically obtaining the pressure distribution through a 3D wing based on the modeling and simulation of the induced eddies. The main advantage of this algorithm is that it includes viscosity effects near the boundary layer with the computational cost of potential methods. Consequently, it is feasible to assess the validity of different wing configurations within seconds. This is especially useful in the first steps of the design process when a wide range of configurations should be explored