Machine learning eliminates near-surface warm bias in reanalysis and reveals weaker winter surface cooling over Arctic sea ice

The surface energy budget of the Arctic Ocean governs sea ice growth in winter and melt in summer. Understanding the surface energy budget and 2m temperature and correctly representing them in models is a key condition for understanding and projecting Arctic climate change. Direct observations of surface fluxes are scarce, and widely used reanalysis datasets suffer from systematic biases. Here, we train a neural network with observational data to bias-correct ERA5 reanalysis surface fluxes. We achieve substantial reductions in RMSE for hourly values of net shortwave radiation (~40%), downward longwave radiation (~16%) and the total surface energy budget (~55%) as well as 2m temperature (~34%). Our bias-correction eliminates the wintertime warm bias of about 4K in ERA5, reduces wintertime surface cooling by about 50% and dampens summertime surface heating. This revised surface cooling estimate is consistent with independent satellite-observed sea ice growth rates. In contrast to ERA5 fluxes, our bias-corrected data capture the observed clear and cloudy states of the Arctic winter boundary layer and the associated bimodal distribution of net longwave radiation. The bias-corrected data provide an improved baseline for climate model evaluation, climatological and case studies and forcing to drive stand-alone sea ice and ocean models.