Development of 2-step regression model for reconstruction of ambient air PM mass: Theoretical, and machine learning approaches

Particulate matter (PM) mass closure, also known as mass reconstruction (RM), establishes a balance between gravimetrically measured PM mass (MM) and the sum of its analyzed chemical components. It aims to reduce sampling artifacts, account unmeasured mass, and ensure measurements reliability. However, it is laborious, time consuming and incapable to capture non-linear atmospheric complexity. To address this, we have developed a novel machine learning based Two Step Regression model designed to compute non-linear and linear components sequentially subject to mass closure constraints. Model is developed from speciation dataset measured in coal-mining area of Singrauli—Sonebhadra, India. Its robustness and chemical plausibility have been ensured via step-wise hyper-parameter tuning and explainable AI (SHAP) respectively. The Extreme gradient boosting—Ridge regression (XGB-RR) with intermediate conditional constraint has performed better relative to others with low RMSE, and high R 2 . The optimized XGB-RR model is applied for PM 10 and PM 2.5 dataset at two distinct elevations. The model yields linearly resolved RM having strong correlation (≥ 0.97) with MM and chemically explainable components having maximum deviation (~ 3%) from the actual measurements. The model offers scalable, frequent and interpretable alternative for offline and online source apportionment (SA) studies.