Nozzle Shape Optimization based on Machine Learning using Higher Order Neural Networks

Abstract

In this contribution, a methodology of plane nozzle shape optimization based on machine learning is introduced. In contrast to standard deep neural network, the proposed neural network is built using higher order neural units. Polynomial structures together with various activation functions are employed as approximators of strongly nonlinear Navier-Stokes equations which govern the flow. Shape of well-known NASA nozzle is chosen as initial geometry which is approximated with 5-th order Bezier curve. Different geometrical shapes, derived from the initial geometry, are employed in order to obtain training data set. Thus, the task consists of multi-variable optimization with defined cost function as a targets which are calculated by means of computational fluid dynamics (CFD) performed on fully structured meshes. The goal of this optimization is obtain geometry which meets desired conditions at the outlet of the nozzle e.g., flow field uniformity, specified flow regime etc. Finally, performance of different approximators is compared and best candidates of optimization are validated through CFD calculation.