Examination of Cloud Phase Partition and Controlling Factors Based on In Situ and Satellite Observations and Machine Learning Approaches

Abstract Cloud phase partition among liquid, mixed, and ice phases was investigated using in situ and satellite observations. A large in situ observation dataset was compiled from 11 flight campaigns covering a near global extent (87°N–75°S, 128°E–37°W). Random forest ensemble models were used to quantify the effects of key controlling factors on cloud phase partition, including temperature ( T ), relative humidity with respect to ice (RHi), vertical velocity ( w ), and aerosol number concentrations (Na) for larger and smaller particles. Our results show that using RHi or T as a single predictor significantly enhanced the prediction accuracy of total cloud occurrences and cloud phase partition, respectively. The inclusion of Na predictors provided a slight but consistent improvement in the F1 score. The feature importance analysis revealed that Na predictors were used by the model nodes at a frequency comparable to that of w . The models successfully captured key physical features, such as the increase in ice phase frequency with decreasing temperature and the peak in-cloud frequency near ice saturation. They also reproduced the main biases in satellite products, such as the overestimation of mixed phase and ice phase by CloudSat and radar–lidar (DARDAR), respectively. These results demonstrate the utility of machine learning for quantifying the contributions of complex factors governing cloud phase partition and for diagnosing systematic biases in observation datasets.