Physics-Informed Machine Learning for Prediction of Sea Ice Dynamics Derived from Spaceborne Passive Microwave Data

Spaceborne passive microwave (PMW) images have been used as primary data sources to obtain sea ice concentration (SIC) and sea ice velocity (SIV) information in the polar oceans. Based on the PMW satellite observations and meteorological air temperature and wind data, we develop a fully convolutional neural network to predict daily SIC and SIV. When training this deep learning model, instead of using a fully data-driven approach, we integrate physical knowledge about sea ice dynamics to regulate the prediction results into physically valid values. This physics-informed learning is conducted by including the physics loss function that is independent of the data loss function. Our experiment shows that the physics loss function improves SIC and SIV predictions for most of the Arctic Ocean and winter seasons. The enhancement by the physics loss function appears more substantial when we predict SIV with a small number of training samples.