Hybrid numerical–ML integrated model for hourly high spatial resolution PM_2.5 forecasting up to 120 hours

Persistent PM_2.5 heavy pollution events driven by rapid urbanization and economic expansion bring severe public health risks, which creates an urgent demand for reliable spatiotemporal hourly pollutant forecasting. Traditional prediction schemes relying solely on chemical transport models or data-driven learning tools suffer from limited spatial resolution and unsatisfactory prediction accuracy. This study constructs a coupled hybrid forecasting framework in China that integrates GEOS-CF atmospheric composition simulation outputs (coupled with GEOS-Chem chemical modules) and multi-source high-resolution spatiotemporal covariates via machine learning algorithms. The system outputs continuous gap-free PM_2.5 gridded predictions at 1 km spatial resolution with a maximum 120-hour forecast window. Strict time-based train–test separation is implemented to avoid data leakage and artificially inflated evaluation metrics. Our framework, rigorously validated under realistic forecasting conditions using a temporally independent dataset, demonstrates substantial improvements over the raw GEOS-CF forecasts. For the first 24 hours, it achieves an R of 0.63 (vs. 0.29) and an RMSE of 30.17 μg/m³ (vs. 99.48 μg/m³). Regional validation in two typical polluted areas further confirms the model's effectiveness: in the Fenwei Plain, RMSE decreased from 102.21 μg/m³ (GEOS-CF) to 38.96 μg/m³ over the full 120 hours; in the Beijing–Tianjin–Hebei region, RMSE decreased from 117.88 μg/m³ to 39.38 μg/m³ over the full 120 hours. This computationally efficient framework, requiring only standard computing resources, offers a scalable solution for operational air quality forecasting and informed public health interventions.