Hull Form Design Optimization using ROM with Machine Learning and Active Subspace Methods

Abstract

An engineering design using high-fidelity CFD methods requires computationally expensive calculations, as the full order CFD simulation should be applied to all design candidates to resolve detailed flow field information and estimate the design performance, which becomes realistic constraints on time and memory. Alternative methods based on the surrogate models have disadvantages in that the number of objective and constraint functions is limited to a small value. In this study, the dimension of the entire flow field was reduced by using a reduced order model (ROM), while the design space is also reduced by the active subspace method, and mapped back to the full dimension. For optimal shape design, ROM was developed as a POD-GPR method, integrated with Gaussian Process Regression (GPR), a machine learning-based method. The developed ROM method was verified for its accuracy and efficiency by POD analysis and experiments. In the first design problem, the design convergence was compared with the existing surrogate model and vanilla ROM in addition to the ROM applied with ASM. In addition, we added the optimal solutions obtained through optimization iterations to the database and used them to update the ROM. When optimization was carried out with the updated ROM, it showed improved convergence compared to the initial model, and more total resistance was reduced. The second problem is to minimize reverse flow, and based on the overall flow field information predicted by ROM, it was possible to reduce the area where negative shear stress exists and its strength. It was confirmed that the wake distribution flowing into the propeller changed due to this effect. Further analysis and research are needed for this result, but it has been shown that ROM can be applied to more diverse design problems.