Cavitation nuisance identification through machine learning during propeller optimisation

Paper i proceeding, 2022

The marine propeller design process runs under strict time limitations and even if it entails contradicting requirements from different stakeholders and complex physical phenomena, the analysis tools must be very fast. Cavitation nuisance is such a complex phenomenon that is hard to predict accurately from these tools and requires additional evaluation by the blade designer. Thus, a good blade design depends on approximate analysis tools and on the expertise of an experienced blade designer. Therefore, we previously developed an interactive optimisation methodology, where interactive genetic algorithms were utilised for blade design optimisation and cavitation was manually evaluated by the blade designer. However, since blade design involves a large design space, the optimisation requires populations of thousands of individuals, something that makes the manual evaluations by the blade designer very laborious. Accordingly in this study, a machine learning pipeline has been developed with the aim to reduce the number of manual evaluations and classify the cavitation nuisance automatically. Nested-cross validation has been used in order to identify the best classification algorithms combined with the most suitable hyperparameters for three different propellers with both suction and pressure side cavitation. The results have shown that using machine learning can be very beneficial in order to reduce user fatigue in interactive optimisation processes.