Emulating Greenland Ice Sheet Surface Melt Using Graph Neural Networks

As global mean temperatures exceeded the 1.5 °C threshold in 2024, the urgency to better quantify the impacts of global warming, including sea level rise contributions from polar ice sheets, has intensified. The Greenland Ice Sheet (GrIS) has experienced significant mass loss over recent decades, primarily driven by surface melting, a process expected to accelerate under continued warming. Surface melt is influenced by a combination of factors and complex interactions between atmosphere and ice sheet surface, but simulating these processes using coupled climate models is computationally expensive and often impractical.

In this study, we develop a graph neural network (GNN) as an emulator for GrIS surface melt, trained on output from the Community Earth System Model version 2 (CESM2), which explicitly calculates surface melt through a downscaled surface energy balance framework. GNNs are uniquely suited to this task, as they capture spatial and relational dependencies across the ice sheet, enabling the emulator to reproduce spatially resolved melt fields and identify the influence of key atmospheric patterns.

We will first evaluate the emulator’s performance in replicating CESM2 simulated melt under different climatic conditions and employ explainability techniques to identify the relative importance of key atmospheric patterns in driving surface melt. This work aims to demonstrate the utility of machine learning emulators in enhancing our understanding of GrIS surface melt dynamics and advancing projections of sea level rise under future climate scenarios.