Efficient Representation of New Ship Dynamics Through Reuse of Existing Maneuvering Models

Abstract

Reliable ship maneuvering models are essential for simulation, control, and autonomy, but constructing ship-specific models remains costly and time-consuming due to dedicated experiments and parameter identification. Meanwhile, many validated maneuvering models already exist from previous projects, yet their use is typically limited to the ships for which they were developed. This study proposes a framework for predicting the dynamics of a new ship by blending multiple trusted maneuvering models rather than identifying a new model from scratch. A neural network computes state-dependent blending weights over a dictionary of existing ship models, allowing the target dynamics to be represented as a convex combination of validated responses. The proposed method is evaluated through simulation studies and full-scale experiments using the research vessel RV Gunnerus. Results show that prediction accuracy improves rapidly when the dictionary provides sufficient coverage of the target dynamics, while additional models yield only marginal benefits. When the target dynamics lie outside the covered response space, the method converges toward the closest available model rather than producing unconstrained extrapolations. These findings suggest that validated maneuvering models can be reused as engineering assets and that new ship dynamics can be represented efficiently through blending existing models. Note to Practitioners—Developing a trustworthy maneuvering model for a new ship often requires costly experiments, expert knowledge, and extensive system identification. As a result, constructing a dedicated model for every new vessel is not always practical. This work proposes a different perspective: instead of building a new model from scratch, a new ship can be represented as a blend of previously validated ship models. The proposed method uses a neural network to combine existing maneuvering models in a state-dependent manner, enabling prediction of new ship motions while reusing accumulated modeling assets. A key practical insight is that prediction performance depends primarily on the coverage of the model library rather than on selecting a single best model. Consequently, organizations that already possess multiple validated maneuvering models can systematically leverage these assets for future projects. The method is most effective when the target ship lies within the range of behaviors represented in the library; outside this range, predictions naturally converge toward the closest available model. Overall, the framework provides a practical and interpretable approach for reusing existing ship models and reducing modeling effort for new vessels.