Machine learning reconstruction of large-scale motions and inclination angles in stratified atmospheric surface layer Available to Purchase

This study develops a hierarchy of extreme gradient boosting (XGBoost)-based models to predict and interpret large-scale coherent structures in the atmospheric surface layer (ASL), utilizing 63 h of high-resolution field data collected during the Surface Layer Turbulence and Environmental Science Test campaign. The target variable is the filtered streamwise velocity at the outer-layer scale, and the models leverage dynamic and thermodynamic inputs sampled at one or two vertical levels. Results show that dual-height input configurations substantially improve predictive performance at elevated heights by capturing the vertical coherence of wall-attached eddies. In addition to accurate reconstruction of velocity fluctuations, the models recover the structural inclination angles of dominant turbulent motions, with predicted values closely matching two-point correlation estimates and empirical stability-dependent scaling. Notably, feature importance (FI) analysis reveals that the most influential input components align along the dominant inclination angle of coherent structures. This directional consistency is quantitatively validated by comparing FI-derived contours with traditional two-point correlation maps. Together, these findings demonstrate that XGBoost not only enables skillful prediction of large-scale ASL dynamics but also encapsulates physically meaningful structure orientation through interpretable metrics. The integration of machine learning with structural turbulence theory provides a robust framework for wind profile prediction, structure-aware modeling, and data-driven parameterization of wall-bounded flows.