A Physics‐Guided Machine Learning Hybrid Parameterization of Near‐Surface Turbulent Fluxes in the Arctic Stable Boundary Layer From MOSAiC Observations

The transfer of momentum and sensible heat across the air‐ice interface shapes Arctic climate, sea‐ice evolution, and weather predictability. Most weather and climate models estimate these fluxes using bulk formulas based on Monin‐Obukhov similarity theory (MOST), yet MOST‐based relationships can deteriorate under strongly stable stratification and intermittent turbulence common over sea ice during the polar night. In this work, a full year of ten‐minute data from the meteorological tower setup during the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition is examined. We evaluate the advanced Arctic bulk parameterization of Gryanik et al. (2020, 2021, https://doi.org/10.1175/jas‐d‐19‐0255.1 , https://doi.org/10.1029/2021ms002590 ) (hereafter, the Gryanik scheme ) against MOSAiC observations, with emphasis on regimes where performance degrades: very strong stability (, especially ), strong winds (), and the autumn freezing period over the MOSAiC floe (24 October–31 December 2019). Over the full stable boundary layer data set, the Gryanik scheme achieves RMSEs of for momentum flux () and for sensible‐heat flux (), but systematic regime‐dependent biases emerge, including substantial momentum underprediction and errors in that can reach order under very stable or strong‐wind conditions. To address these deficiencies while preserving the MOST‐based bulk structure, we develop a regime‐aware Gryanik‐ML Hybrid that applies three least‐squares boosting (LSBoost) residual‐correction models for the , , and freezing period regimes using only routine near‐surface predictors. When applied year‐round, the hybrid improves overall skill, reducing RMSE to for and for , and provides a practical pathway to reduce systematic biases in sea‐ice flux parameterizations without sacrificing physical transparency.