Improving Low-Cloud Fraction Predictions through Machine-Learning Techniques

In this study, machine learning (ML) models (XGBoost and neural network) are built based on a wide range of cloud-controlling meteorological factors to predict low-cloud fraction (LCF) and explore their relationships using an explainable statistical approach. The prediction efficacy of these ML models is compared against traditional LCF predictors (i.e., the estimated inversion strength and the estimated cloud-top entrainment index) and a multilinear regression model. Our findings show a substantial enhancement in the accuracy of six-hourly LCF predictions by ML models regarding the mean squared errors and coefficients of correlation (R) with observations, compared to other predictors and models. Moreover, the R derived from ML models exhibits the weakest dependence on cloud dynamic regimes. The interpretability analysis indicates the distinct contributions of meteorological factors among regions: potential temperature and relative humidity at 850 hPa (θ₈₅₀ and RH₈₅₀), along with lower-tropospheric stability (LTS), emerge as the most skillful factors over the subtropics, while column-integrated precipitable water vapor (PWV), θ₈₅₀, and θ₁₀₀₀ are the most impactful over the mid-latitudes. Besides, ML models most accurately simulate LCF evolution during the stratocumulus-to-cumulus transition, highlighting the limitations of linear models that incorrectly simulate rapid cloud transitions—a prevalent issue in current climate models. This study indicates the potential utility of ML models in parameterizing low-cloud fraction within numerical models regarding predictive accuracy and computational efficiency.