<p>Substantial advances in PM_2.5 monitoring and modeling have improved our understanding of air pollution, however, the absence of long-term, spatially continuous datasets has limited insights into historical pollution dynamics across diverse environments in China. This study reconstructed monthly PM_2.5 concentrations from 1981 to 2023 by integrating MERRA-2 aerosols, ERA5 meteorology, and surface observations within a unified machine learning (ML) framework. Extreme Gradient Boosting (XGB) outperformed other algorithms and was used for national-scale reconstruction. Sparse monitoring coverage, particularly in northern and northwestern China, was mitigated through a virtual observation network constructed from MERRA-2 grid cells strongly correlated with ground measurements. Emission influences were incorporated through surface gaseous pollutants (SO₂, NO₂, CO), improving representation of chemically driven variability. The reconstructed 43-year dataset captures seasonal cycles, spatial gradients, and multi-decadal trends consistent with independent observations. SHAP interpretation reveals spatio-temporal contrasts in PM_2.5 drivers, meteorology dominates in dust-prone northwestern China and densely populated eastern regions, whereas emissions exert stronger control in cleaner, high-elevation areas such as Tibet. Nationally, temperature emerges as the dominant predictor. In eastern and southern China, winter temperature inversions lead to pollution accumulation, while concentrations drop substantially when temperatures exceed 20°C. This reconstruction fills observational gaps and extends records over four decades, providing a robust basis for understanding emission-meteorology interactions and supporting targeted air quality management and health policy planning at regional and national scales.</p> Graphical Abstract <p></p> <p>This visual summary serves as a pivotal entry point into the research, offering a concise overview of the study methodologies and principal findings on PM_2.5 estimation and its drivers across China. The graphical abstract integrates reanalysis products (MERRA-2 aerosols, ERA5 meteorology), ground observations, and machine learning (XGB, RF, SVM, LR) to generate monthly PM_2.5 estimates from 1981 to 2023. Among the tested models, XGB demonstrated the best performance (R² = 0.96-0.98 against independent datasets for 1981-2023 and 2015-2023), with minimal bias across regions. The central map illustrates regional variations in dominant feature-drivers, highlighting temperature and SO₂ as the most influential predictors of PM_2.5. Time-series comparisons show consistency between modelled anomalies, satellite reanalysis (1981-2023), and ground observations (2015-2023), capturing both long-term increases before 2010 and subsequent declines driven by emission control policies. SHAP feature importance analysis reveals that in eastern and southern China, temperature interacts with anthropogenic emissions (SO₂, NO₂) to amplify wintertime PM_2.5 pollution under inversion-prone conditions. Collectively, the graphical abstract underscores the strength of machine learning in resolving spatiotemporal PM_2.5 patterns, identifying region-specific drivers, and informing targeted air-quality management strategies in China.</p>