A physics-guided Deep Learning Model for SMB of Central West Greenland

Direct observations of the surface mass balance (SMB) over the Greenland Ice Sheet are currently limited to basin-scale estimates and sparse in situ measurements, making a comprehensive observational assessment challenging. Therefore, RCMs remain the best option for producing spatially and temporally continuous SMB estimates, but models show substantial regional differences, even under present-day conditions. While satellite observations provide extensive information about surface processes, a fully satellite-based SMB product remains to be seen.

In this study, we present SMBnet, a deep learning model that estimates SMB over central West Greenland by integrating satellite observations, reanalysis data, and simple physical constraints. SMBnet combines satellite-derived ice surface temperature and albedo with ERA5 reanalysis data, ice velocity, and GRACE/-FO mass anomalies to produce spatially and temporally continuous SMB estimates. The model employs a U-Net architecture and is trained using multiple loss terms that enforce consistency with observations and incorporate physical knowledge of accumulation and ablation processes. Although applied here to central West Greenland, the approach shows clear potential for extending it to the entire ice sheet, offering a computationally efficient, observation-driven complement to traditional RCMs for estimating present-day SMB.