Assessment of a CFD-Based Machine Learning Approach on Turbulent Flow Approximation

Abstract

Computational fluid dynamics (CFD) simulation is usually a computationally expensive, memory demanding, and time consuming iterative process. These drawbacks limit the use of CFD, especially when either spatiotemporal scales or geometry complexity increases. This paper presents the preliminary results from the assessment of an approximation model for predicting non-uniform steady turbulent flows in a 3D domain, utilizing deep learning (DL) algorithms. In particular, the artificial neural network (ANN) approach uses most important variables data from currently CFD simulation results to link multi-variable input spaces (e.g. input speed and direction, geometry configuration) with multi-variable output space (e.g. velocity magnitude, pressure gradient) to obtain an efficient and accurate approximation of the entire velocity field for given input flow field characteristics. The results demonstrated higher computational speed with a similar accuracy using DL algorithms versus CFD simulation. This integrated approach can provide immediate feedback for real-time design iterations for the entire computational domain at the early stages of design. Hence, designers and engineers can easily generate immense amounts of design alternatives without facing the time-consuming task of evaluation and selection.