Accurate short-horizon prediction of six degrees of freedom (6-DOF) vessel motions is essential for autonomous navigation, motion compensation, and operational decision making. Traditional seakeeping models rely on hydrodynamic coefficients that are seldom available for full-scale vessels, while purely data-driven approaches may struggle to maintain physical consistency. This study introduces a hybrid physics–machine learning framework that combines Dynamic Mode Decomposition (DMD), which approximates the vessel’s dominant linear drift dynamics, with a causal Temporal Convolutional Network (TCN) that learns nonlinear residual corrections from a 12-hour historical window of environmental, geometric, and motion features. DMD provides an interpretable surrogate of the vessel dynamics through its eigenvalues, growth rates, and mode shapes, serving as a data-derived linear transfer operator. The TCN predicts only the residual departure from this structured baseline, ensuring a stable and causal forecasting architecture. Evaluation on full-scale field data shows that the hybrid model improves prediction accuracy for heave and achieves performance comparable to DMD for surge, while underperforming in sway, roll, pitch, and yaw due to the limited observability of key physical drivers at hourly resolution. These results highlight both the strengths and limitations of residual learning when important nonlinear forcing mechanisms and control inputs are unmeasured. Overall, the study demonstrates that hybrid physics–machine learning approaches provide valuable interpretability and diagnostic insight, even when data limitations are constrained. The framework offers a principled foundation for incorporating additional physical inputs, higher-frequency measurements, and physics-informed architectures in future work on operational ship-motion forecasting.
Chekure et al. (Thu,) studied this question.