A Machine Learning based Expert System for Optimizing CFD Solver Parameters

Summary

Computational Fluid Dynamics is a viable tool in the field of aerodynamics enabling to reduce time, effort and budget required for experimental testing. Although powerful and established for various years, it remains a complex tool calling for experienced users to ensure consistent high-quality results. This complexity primarily stems from the underlying model, namely the Navier-Stokes equations typically combined with a set of equations resolving the effects of turbulence. Additionally, to obtain accurate high-fidelity result appropriate meshes are required. As a consequence, a substantial number of parameters needs to be selected carefully and the quality of a result often highly depends on individual knowledge and experience of a user. Hence, a strong desire exists to reduce the number of input parameters without causing a loss of accuracy and efficiency. Such reduction of parameters might be viewed as a prerequisite to CFD as a tool in process chains for multidisciplinary applications where typically no user interaction is possible. In this article we propose a machine-learned Expert System for CFD to provide guidance for users in selecting optimal or at least near optimal parameter combinations. The proposed Expert System is divided into two macro steps, the surrogate model and a genetic algorithm to determine from the surrogate model the parameters. Numerical examples are presented to demonstrate the approach.

Abstract

Full Paper