Historical long-term PM_2.5 records with high temporal resolution are essential but lacking for research and environmental management. Here, we reconstruct site-based and gridded PM_2.5 datasets at 6-hour intervals from 1960 to 2020 that combine visibility, meteorological data, and emissions based on a machine learning model with extracted spatial features. These two PM_2.5 datasets will lay the foundation of research studies associated with air pollution, climate change, and aerosol reanalysis.