Abstract: Accurate identification of precipitation types is crucial for the simulation of water-energy balance, hydrological cycle, and cryospheric processes. To enhance the accuracy of precipitation phase recognition over China, this study explores the performance of five machine learning methods within temperature ranges where traditional approaches struggle to accurately differentiate precipitation phases (-7℃，10℃) based on the dataset of daily surface climate variables of China. The results show that the MLP model performs well in identifying precipitation types across China, achieving an overall accuracy of 88.1% on the test set, a 4.4% improvement over traditional methods. In the Tibetan Plateau, the accuracy reached 91.4%, an increase of 12.8% compared to traditional methods. The MLP model notably improves the identification of rain and sleet, especially in areas with frequent rainfall and sleet in eastern Tibetan Plateau and southern China. The performance of MLP model is influenced by sample distribution, achieving lower identification errors for rainfall in South China and snowfall in North China. Analysis shows that wet-bulb temperature, which incorporates information on temperature, humidity and pressure, has the greatest impact on MLP model’s performance. Precipitation amount also significantly affects the probability of sleet classification. Additionally, the MLP model successfully reproduces the observed dependence of precipitation types on elevation and relative humidity. Specifically, it captures the increase in the temperature threshold for precipitation phase discrimination with elevation, as well as the higher probability of sleet occurrence with increasing humidity, demonstrating its ability to represent key physical characteristics with a certain degree of interpretability. Overall, MLP-based machine learning method significantly improves the accuracy of precipitation type identification across China, facilitating imporved understanding and modeling of hydrological and cryospheric processes.