2025S-OS8-2 A Study on Identifying Ship Maneuvering Models Using a Physics-Informed Neural Network with the MMG Model

抄録

This paper presents a method for identifying ship maneuvering models using a Physics-Informed Neural Network (PINN) that integrates the MMG model as physical knowledge. Autonomous ships have the potential to enhance safety and efficiency in maritime operations, but their realization highly depends on the accuracy of ship maneuvering models. Physics-based mathematical models, such as the MMG model, provide valuable physical insights and strong generalization capabilities. However, their accuracy often falls short under real-world conditions. In contrast, pure data-driven models can accurately predict maneuvering motion based on training data, but struggle to generalize performance beyond it. To address these limitations, this paper combines both approaches by leveraging a PINN framework. Physical knowledge, such as motion equations, is incorporated into the loss function of the PINN, which is treated as physics-based loss as distinguished from data-driven loss. The generalization performance of the PINN-based model can be improved compared to pure data-driven models thanks to the contribution of physics-based loss to the training process. This study utilizes the MMG model as the physical knowledge. Experimental validation using KVLCC2 ship model data demonstrates that the developed PINN-based model fits the actual maneuvering motion better than the MMG model and shows improved generalization performance compared to pure data-driven models. Furthermore, results indicate that the balance between data-driven and physics-based losses determines the fitting and generalization performances of the model.