Increases in extreme precipitation can increase the intensity and frequency of flooding, which imposes significant social and economic impacts across the United States (US) and globally. Variations in the magnitudes and frequencies of extreme precipitation can be attributed to a high dimension of physical factors across spatial and temporal scales. Studies linking extreme precipitation to its contributing drivers need to address the challenges in preparing comprehensive extreme events datasets and dealing with high dimensionality and strong nonlinearity in the relationships. In this study, we extracted extreme precipitation characteristics (e.g., magnitudes, return periods) in six major US regions including coastal and mountainous areas, together with local, regional, and global factors at the matching temporal resolution, and then integrated dimension reduction, two ensemble tree-based machine learning (ML) methods, namely random forest (RF) and Extreme Gradient Boosting (XGBoost), and neural networks(NN), to identify the key drivers of regional extreme precipitation towards mechanistic understanding. Based on the selected combination of key drivers, we further established the emulators of extreme precipitation at various return periods, which opens the possibility of the future extreme precipitation outlook with the inputs from CMIP climate projections, whose native resolution is not high enough to resolve heavy rainfall.