Comparison of Different Machine Learning Algorithms in the Parameterization of Near-Surface Turbulent Fluxes

The study of non-iterative schemes for near-surface tu rbulent flux parameterization is significant for model forecasting and climate prediction. In this paper, four machine learning algorithms (Random Forest, CatBoost, XGBoost and LightGBM) are used to design non-iterative schemes for near-surface turbulent flux parameterization. Through comparison, we find that the results obtained by XGBoost and LightGBM exhibit significant fluctuations, sensitivity to minor data changes, and poor stability, with lower accuracy compared to the BH&H scheme, making them unsuitable to design the parameterization model. While CatBoost shows less fluctuation, its results still possess minor irregularities compared to the Random Forest model. The results of Random Forest are smooth, stable, and highly accurate, making it the most suitable algorithm among the four algorithms for application in the parameterization of near-surface turbulent fluxes.