A deep learning model for predicting daily PM_2.5 concentration in response to emission reduction

Air pollution remains a leading global health threat, with fine particulate matters (PM_2.5) causing millions of premature deaths annually. Chemical transport models (CTMs) are essential for estimating how emission controls improve air quality but are computationally intensive. Here, we present CleanAir, a deep learning model that simulates daily PM_2.5 concentration and its chemical composition in response to emission reductions at a 36-kilometer horizontal resolution. Built on a residual symmetric three-dimensional U-Net architecture, CleanAir can estimate 365-day PM_2.5 concentration over China within 10 seconds on a graphics processing unit or 160 seconds on a central processing unit-three to four orders of magnitude faster than CTMs. Results from CleanAir agree well with those from a Community Multiscale Air Quality (CMAQ) model for both PM_2.5 concentration and emission-induced changes. Trained on 2416 emission scenarios from the CMAQ model, CleanAir generalizes well across unseen meteorology and emissions. With fast simulation capability, CleanAir enables extensive evaluation for short-term emission control measures and long-term mitigation pathways, leading to more responsive decision-making.