Deep Learning Antarctic Ice Dynamics, Charlie Cowen-Breen, UG '22, (2327561)

Predicting where and how Antarctica is fracturing is important for understanding rising sea levels and impacts the lives of millions. Fracturing is hypothesized to occur at locations with low hardness of ice, but hardness data is available only for each square kilometer of Antarctica. In this project, we aim to use recent success in physics-informed neural networks (PINN) to construct higher-resolution maps of ice hardness. PINN’s are models which combine input data (in this case, the hardness data available) with known physical laws (in this case, known ice dynamics) to generate predictions informed by our understanding of physics. A major open question, both in this project and in PINN research more broadly, is how to heavily to weigh the assumed physics against the observational data. Indeed, there are instances in which the known equations may be more reliable than the data, and other instances in which they are less so. Here we use a PINN to construct a 10x resolution blowup of the ice hardness map, and find that the known physical laws should be weighted just 1% as much as observational data for best prediction. This can be interpreted as a demonstration of the importance of correct data to model ice hardness, and serves as an argument for closer observation of Antarctic fracturing.