An Exploration of Flow Control Using Machine Learning and Computational Fluid Dynamics

Abstract

Although numerous studies relating to computational fluid dynamics and machine learning have been conducted in relation to automotive development, the majority focus on either early development using completed 3D models, or the final testing stages of development, or machine learning accelerated computational fluid dynamic simulations. While this approach is helpful in software development and simulation, it is not easily adaptable to automotive design where the final model is constantly changing and being modified. Consequently, the aim of this study is to propose a method for conducting computational fluid dynamics and machine learning concurrently to accelerate the development process. The proposed method is used to design and improve the aerodynamic efficiency of an object. The approach focuses on developing, implementing, and comparing machine learning models capable of generating optimised three-dimensional objects with the required geometry to direct airflow paths required in applications such as pressure generation, as needed for both active and passive flow control. The study concludes that both decision tree regression and long short-term memory (LSTM) autoencoder models could be used to optimise the aerodynamic efficiency of solid bodies, but that the LSTM autoencoder performs better overall. An undesirable effect of the shape optimisation is an overall reduction in shape size as optimization increases.

https://doi.org/10.59200/ICARTI.2023.017