Particulate matter with aerodynamic diameter less than 2.5 μm (PM2.5) increases mortality and morbidity.1,2 PM2.5 is composed of a mixture of chemical components that vary across space and time.3 Due to limited hyperlocal data availability, less is known about health risks of PM2.5 components, their US-wide exposure disparities, or which species are driving the biggest intra-urban changes in PM2.5 mass. Here, we developed the first national super-learned models across the US for hyperlocal estimation of annual mean elemental carbon (EC), ammonium (NH4+), nitrate (NO3-), organic carbon (OC), and sulfate (SO42-) concentrations across 3,535 urban areas at a 50-m spatial resolution, and at a 1-km resolution for non-urban areas from 2000 to 2019. Using ensembles of machine learning models and ~82 billion predictions across 20 years, hyperlocal super-learned PM2.5 components are now available for further research. We found remarkable spatiotemporal intra-urban and inter-urban variabilities in PM2.5 components. We anticipate our work to be a critical milestone for conducting new studies on individual and combined health risks of PM2.5 components, environmental justice analysis, or understanding fine-scale spatiotemporal variabilities of PM2.5 composition. Urban planners and regulators may also use these predictions for selecting locations of new daycares, schools, nursing homes, or air-quality monitors.