Surface melting over the Greenland Ice Sheet is a major contributor to rising sea levels. While regional climate models like MAR simulate melt with high accuracy, they are too computationally intensive for fast or large-scale use. To address this, we developed a machine learning model called MAR-IA, trained on MAR data from 1979 to 2023, and tested it using various configurations, including climate variables from the ERA5 dataset.

MAR-IA closely matches MAR’s melt estimates in ideal cases but is less accurate when predicting unfamiliar years. Models using only ERA5 data performed much worse , and even combining datasets offered limited improvement. Results show that using more recent training data boosts accuracy, suggesting that Greenland’s changing climate reduces the usefulness of older records. Key predictors include air temperature, energy fluxes, and especially surface albedo.

We also found that models trained in southwest Greenland perform better when applied to other regions. Overall, our study shows both the potential and the pitfalls of using machine learning to forecast ice melt in a warming world.