Arctic Sea Ice Prediction Based on Multi-Scale Graph Modeling With Conservation Laws

Abstract

Arctic sea ice prediction is critical for exploring climate change, resource extraction, and shipping route planning. This paper introduces a novel neural network model, Ice Graph Attention neTwork (IceGAT), that is trained to predict sea ice concentration (SIC) from a number of atmospheric, oceanic, and land surface measurements. It is based on two design principles: (a) the complex spatial interactions in weather dynamics are captured via a series of graphs corresponding to different spatial resolutions and (b) the incorporation of the physical conservation laws for moisture and potential vorticity. We devise two main variants with 1 hr and 24 hr temporal resolution and determine the optimal input horizon to be 5 days. IceGAT features leading accuracy (96.7%; +2.4% over the current state-of-the-art) and low inference time (1/4 s, on a single GPU). An online implementation (based on data from ERA5) alongside supplementary videos and our shared code are accessible at: https://lannwei.github.io/IceGAT/.