An improved physical information network for forecasting the motion response of ice floes under waves

Abstract

Physics-informed neural networks (PINNs) have increasingly become a key intelligent technology for solving partial differential equations. Nevertheless, for simulating the dynamic response of ice floes to waves, researchers often still resort to traditional numerical methods and empirical formulas. The limitations of these methods include extended computational durations and challenges in precisely conforming to physical principles. To effectively overcome these challenges and achieve efficient and accurate prediction of sea ice motion response, this study proposes an improved PINN method for the longitudinal motion response of sea ice under regular wave action. The approach features two principal innovations: first, a neural network loss function module tailored to the ice motion response equations, and second, an attention mechanism focused on temporal sequences and wave data. Through case studies of sea ice motion under three different wave conditions, this research validates the effectiveness of the improved PINN method. A comparison between the network's training and testing outcomes with experimental figures reveals significant consistency, affirming the method's robustness and accuracy. The application of this study demonstrates the potential for accurately predicting the dynamic response of sea ice in complex sea conditions, providing significant technical support and new research directions for future studies.