MACHINE LEARNING CONTROL OF 3D LIQUID FILM

The rise of machine learning has made it a crucial element in the field of fluid dynamics, and it seems very promising indeed when applied to flow control problems. This work focuses on the development of ML algorithms that can be used in the hot-dip galvanizing process to regulate the thickness of the liquid zinc film, reduce instabilities, and obtain a smoother surface, which can be critical to the performance of the final product for industrial applications. In this study, we seek to enhance the simulation of a physical problem by implementing and testing several algorithms: Bayesian Optimization (BO), Lipchiz global optimization (LIPO), and Reinforcement Learning algorithm (RL). These algorithms are applied within a simulated environment called BLEW (Boundary Layer Wiping). Our approach involves an initial analysis of the experimental setup, where we assess various configurations of jets and sensors. Additionally, we focus on optimizing the models' hyperparameters and defining appropriate actions and a cost function to improve learning performance. The results are visualized through learning curves, followed by a comparative analysis of the different algorithms employed.