Over the past decade, global aerosol size distributions have undergone substantial changes, yet accurately capturing these dynamics through satellite-based fine-mode (fAOD) and coarse-mode (cAOD) particles remains challenging. Here, we introduce a novel deep learning model, the Spatial-Temporal Pre-trained Transformer (SPT), designed to enhance the accuracy of fAOD and cAOD retrievals by independently capturing spatial and temporal features from satellite data. Leveraging this approach, we generate a 23-year (2001–2023) global SPT-derived fAOD and cAOD dataset at 500 nm, featuring daily temporal resolution and a 0.5° spatial resolution, providing a valuable resource for atmospheric and climate studies. The SPT fAOD and cAOD are uploaded as Geotiff format, stretched from -90° to 90° latitude and from -180° to 180° longitude.