Physics-Informed Machine Learning for Sea Ice Thickness Prediction

Arctic sea ice thickness is vital for evaluating Arctic sea ice loss. The study aims to investigate Arctic sea ice thickness prediction using PhySIT, a physics-informed machine learning model for sea ice thickness prediction. We use the PhySIT framework to solve the governing partial differential equations relevant to ice thickness distribution with real-world data, with specific assumptions, as constraints in the loss function. Our proposed approach utilizes physics knowledge to guide the prediction process of LSTM and GRU models. We compare our proposed LSTM and GRU models, which integrate physics, against well-known LSTM-based physics-informed model baselines. The results indicate that our PhySIT-based LSTM and GRU models perform better than the baselines. We also incorporated Kolmogorov-Arnold Networks (KAN) to PhySIT to capture nonlinear properties associated with sea ice thickness data. Our results show that the PhySIT using KAN shows better predictive performance. We also introduce a modified layer-wise relevance propagation (LRP) for spatio-temporal data. This method aims to identify the hidden patterns of the data that are most important for the model’s predictions. We use our LRP method on the sea ice thickness data to highlight the features’ contribution to the models’ predictions. This analysis demonstrates that LSTM models are faster at predicting time series data than GRU models. We use our LRP method on the sea ice thickness data to highlight the features’ contribution to the models’ prediction. Our study shows PhySIT framework, using a combination of GRU or LSTM and KAN, is better at predicting sea ice thickness than other recent physics informed models. Our PhySIT framework that designed infusing physics loss using a combination of GRU or LSTM and KAN, is better at predicting sea ice thickness than other models. Our study highlights the potential of incorporating constrained physics into the loss function to enhance a model’s adherence to data-sparse environments. By combining LSTM architectures, physics-informed approaches, and KAN methodologies, we demonstrate the feasibility of developing more effective and simple predictive models for spatio-temporal data.