Interpretable machine learning for occupant-specific PM2.5 exposure assessment in higher education buildings

Outdoor-origin fine particulate matter (PM2.5) poses significant health risks in Higher Education Institution (HEI) buildings, where occupants spend extended periods across diverse functional spaces. This study develops a scalable framework for indoor PM2.5 exposure assessment by coupling CONTAM–EnergyPlus co-simulation with machine learning metamodels and SHapley Additive exPlanations (SHAP)-based interpretability. An Extreme Gradient Boosting (XGBoost) metamodel trained on 2,729 zones across five UK HEI buildings achieved high predictive accuracy (R ≈ 0.95; R2 > 0.90 on held-out data). SHAP analysis, representing what appears to be the first such application in HEI indoor air quality assessment using a physics-driven metamodel, identified building airtightness (Q50) as the dominant exposure driver, followed by infiltration air change rate (ACHINF) and indoor–outdoor temperature difference (ΔT). Microenvironmental modelling indicated indicative pronounced exposure heterogeneity among occupant groups within the adopted assumption space: offices dominated staff exposure while educational facilities drove student exposure. Improving airtightness from baseline to Q50 = 3 m3/h·m2 reduced population-weighted exposure by up to 32.3%; however, approximately 88% of zones still exceeded the WHO 2021 annual guideline of 5 μg/m3. These findings demonstrate that envelope improvements alone are insufficient for WHO compliance and must be complemented by integrated mechanical ventilation and filtration strategies, alongside urban-scale policies such as Clean Air Zones and emissions control measures that reduce outdoor PM2.5 at source. The framework provides a transparent, physics-grounded basis for screening HEI building stocks and prioritising evidence-based air quality interventions.