Data-driven equation discovery of a sea ice albedo parametrisation

How can machine learning improve the representation of sea ice in climate models? In this study , Diajeng W. Atmojo and co-authors use symbolic regression, an interpretable machine-learning technique, to develop an improved parametrization of Arctic sea ice albedo based on satellite and reanalysis data. The newly discovered equation identifies key physical drivers of sea ice albedo, including snow depth, surface temperature, sea ice thickness, and air temperature, and reduces mean squared error by approximately 51% compared with the commonly used Parkinson and Washington parametrization. Unlike traditional neural networks, the interpretable equation also enables investigation of regional and seasonal differences in the processes controlling sea ice albedo. Its ability to transfer across distinct sea ice regimes, including the Barents Sea and Central Arctic, highlights its potential to improve our understanding of the Arctic surface energy budget and climate projections.