Data-driven machine learning model for propeller design: Applying supervised regression algorithms to predict propeller blade geometry

Abstract

The high demand for innovative and high performance propellers for marine vessels motivates propeller designers to strive for a more efficient and less resource costly design process. To create a more robust process, propeller designers look towards potential resource optimization tools in the form of calculators, optimization al- gorithms and more. In recent years the topic of machine learning have made it’s way into the propeller design field as a multifunctional tool that can be suited for many different tasks as long as the provided data exists. This thesis aimed to use supervised machine learning regression models to generate the first geometric blade profile that lay the starting basis for all propeller design processes. The methodology included the creation of a data set based on high performing controllable pitch propellers, where geometric blade profiles were used to shape the in and output data for the models. Random forest, XGBoost, Neural network and support vector regression models were all used in the machine learning pipeline to make predictions on the blade geometry. Suitable performance metrics and cross- validation was plotted and used to ensure accuracy in the predictions while also providing interpretability to the models. The results from this study indicated that the highest performing model could explain 70% of the data-set variance. While not bad the results could be improved and it is believed that a lack of data quantity is the main factor holding the machine learning models back. More research and different approaches to the topic is needed to fully explore the possibilities of machine learning within the propeller designing field. This study provides valuable insight into the potential uses of machine learning and the specific performance of a set of models for smaller data-sets, resulting in a usable prediction tool that contribute to the propeller design process.