Multi-fidelity Bayesian optimization of wind assisted vessel hull for resistance

Abstract

Wind assisted vessel hull optimization is more expensive than traditional vessels. This thesis tests if multi fidelity modeling may be a solution to reduce the computational cost of optimizing wind assistedvessel hulls. A Bayesian optimization framework was coded that can use both single fidelity and multi fidelity data. A surrogate test problem of a 100 meter long wind assisted vessel was created based on CFD data and an empirical model. The first experiment compared the performance of the single and multi fidelity algorithms. The second experiment evaluated the impact of a better low fidelity by using weighted averages of the CFD and empirical as the low fidelity model. In the third experiment, the resulting vessels from optimizations with CFD and empirical models were compared to validate the necessity of including CFD in the optimization of wind assisted vessels. The results of experiment one show that multi fidelity is only slightly faster, but has a larger spread meaning the slowest optimization takes longer. Experiment two showed that a better low fidelity model drastically reduces the number of iterations needed to perform an optimization. Experiment three showed that the vessels optimized using CFD differ significantly from those optimized with empirical methods. Thus, multi fidelity modeling provided very limited improvements for the optimization of a wind assisted vessel when using empirical models and CFD. Significant improvements are possible with better low fidelity models. Lastly, empirical models used for wind assisted vessels are have limited usability for optimization.