Machine Learning for Explanation of Subgrid Convective Precipitation: A Case Study over CONUS Using a Convection-Allowing Model and SHAP Analysis

Abstract Understanding the role of dynamic, thermodynamic, and cloud microphysical parameters governing the occurrence and magnitude of convective precipitation at subgrid scales is crucial for reducing uncertainties in extreme precipitation properties projected by global climate models. This study evaluates the efficacy of extreme gradient boosting decision trees in detecting subgrid convective precipitation features (CPFs) and estimating their extent using convection-allowing simulations based on observations across the contiguous United States (CONUS). Our annual analysis reveals that despite the imbalanced nature of detecting subgrid CPFs, the model achieves an 85% F1 score using key CPF predictors at 100-km grids. Performance metrics for CPF detection and extent estimation show significant seasonal variations corresponding to mesoscale environmental conditions influencing their space–time dynamics. Shapley values from the model indicate that cloud liquid and ice water content are primary predictors in CPF detection, while midlevel vorticity induced by convective heating profiles predominantly influences CPF extent estimation. Significance Statement Understanding future changes in precipitation extremes requires a grasp of small-scale convective processes, which global circulation models currently cannot explain. This study aims to enhance our understanding of where, when, and to what extent machine learning can characterize the occurrence and extent of convective precipitation features. Learning from convective-enabled simulations and ground-based radar data over the contiguous United States, the findings demonstrate that extreme gradient-boosted decision trees can effectively capture these features and explain the relative importance of cloud microphysics, thermodynamics, and dynamic environmental forces that drive the space–time variability of convective precipitation.