Physics-Embedded Deep Convolutional Network: A Novel Approach for Prediction of Sea Ice Concentration and Motion

Reliable prediction of short-term Arctic sea ice variation is crucial for ensuring the safety of navigation on Arctic shipping routes. While deep-learning models have demonstrated potential in improving the accuracy of sea ice predictions, many data-driven approaches focus solely on individual aspects of sea ice without considering the interrelationships and underlying physical laws governing various sea ice factors. To address this limitation, we introduce a dual-task prediction model that simultaneously targets sea ice concentration (SIC) and sea ice motion (SIM). Our approach incorporates a novel loss function that enforces dynamic constraints derived from the sea ice control equation, ensuring that predictions of both SIC and SIM are consistent with physical dynamics. We conduct comprehensive comparative experiments to identify the optimal model structure for predicting SIC and SIM. Our findings reveal that a dual-task branching architecture is particularly effective for this purpose, with a post-decoder branch network structure exhibiting the best performance in predicting both SIC and SIM. By integrating the sea ice dynamics equation into the loss function, our models demonstrate enhanced alignment with physical laws, leading to improved predictability and accuracy in SIC and SIM prediction.