Leveraging observations to bias-correct the Arctic surface energy budget in reanalysis through machine learning

The Arctic Ocean plays a critical role in global climate dynamics, yet direct observations of its surface energy budget are sparse and largely constrained to individual field campaigns. Current estimates often rely on reanalysis data, notably ERA5, which demonstrates systematic biases in boundary-layer properties and surface fluxes over Arctic sea-ice. In this study, we train an artificial neural network (ANN) to predict surface fluxes observed during MOSAiC, SHEBA, Arctic Ocean 2018 and ARTofMELT expeditions using ERA5 data as input. Data from shorter field campaigns, such as N-ICE, are used for testing our model against unseen data. Our results indicate that ERA5 Arctic surface fluxes exhibit very low correlations with observations and are characterized by large RMSE values. Our predictions demonstrate significant error reductions across key variables: sensible heat flux (~39%), 2m temperature (~39%), net shortwave radiation (~37%), downward longwave radiation (~21%) and net longwave radiation (~17%). Furthermore, we find a higher correlation (~0.58) with observations compared to ERA5 (~0.24) and approximately 50% reductions in RMSE of the hourly total surface energy budget, i.e. the sum of the individual fluxes. We produce a bias-corrected estimate of surface energy fluxes over Arctic sea-ice. We use our bias correction to revise previous estimates of the climatological surface energy budget over Arctic sea ice. We will make the trained weights available to allow for the custom derivation of bias-corrected fluxes in individual case studies and for climate model evaluation.