Neural network based geometric primitive for airfoil design

A geometric primitive for CAD implementation is presented (Bezier neural network, BNN). It is specifically designed to reproduce geometric shapes with functional requirements such as aerodynamic and hydrodynamic profiles. This primitive can be useful when a known and well defined map between functional requirements and geometric data does not exist, and it have to be deduced by a physical or numerical experimental analysis. BNN gives rise to a typical CAD representation, a Bezier curve, of a functional profile, once the functional parameters are supplied. In BNN the capability of neural network to approximate very complex and non-linear function has been combined with the capability of Bezier functions to describe geometry, in a unique neural network. In this work BNN is used in the representation of aerodynamic profiles starting to their typical functional parameters: lift and drag coefficients, Reynolds number and angle of attack. BNN is tested in reproducing the wing profile of the 4-digit NACA series. The output of BNN is compared with the results of a fluid-dynamic analysis performed by commercial software.