The construction of a neural network proxy model for ship hull design based on multi-fidelity datasets and the parameter freezing strategy

Abstract

Designing a ship hull to keep the hydrodynamic performance meet the requirements is an indispensable step in the ship's preliminary design. However, in the traditional design process, the design workload is labor-intensive and time-consuming due to the need to estimate a large number of design schemes. To address this issue, this work developed a Neural Network Proxy Model (NNPM) to assist the hydrodynamic resistance prediction for the ship hull design process. The data used in the NNPM construction was supported by two Computational Fluid Dynamic (CFD) methods with different fidelity where the low-fidelity one for tuning and pre-training, and the high-fidelity one for fine-training. To mitigate the risk of overfitting stemming from disparities in data volumes, the parameter freezing strategy is adopted during the high-fidelity dataset-based fine-training. The results obtained from validation numerical experiments indicated that the NNPM exhibited a forecast error of less than 1% when compared to the high-fidelity CFD method. Importantly, this high accuracy is achieved while maintaining a low construction workload, demonstrating the potential of Artificial Intelligence (AI) to predict the hydrodynamic load of hulls in the ship's preliminary design, which can further advance the artificial intelligence-assisted design (AIAD) technology for various marine structures.