Generative physics-informed machine learning for modeling indoor air quality and its impact on student health and performance

This study presents a novel hybrid framework combining experimental testing, statistical analysis, physics-based modeling, and interpretable machine learning (ML) to assess indoor air quality (IAQ) and its impact on student health and academic performance in healthcare and laboratory environments. A comparative feature importance analysis using Spearman, Kendall methods, and SHAP values identified particulate matter (PM₁₀) as the most critical pollutant, with location, noise, and thermal factors playing secondary roles. Empirical assessments revealed concerning IAQ variations-Lab 187E (Biological Prep) exhibited elevated PM_2.5 (12.4 μg/m³), while Lab 290D (Chemical Storage) showed high TVOC (312 ppb) and Virus Index (3). At Bilborough Medical Centre, CO₂ peaked at 2819 ppm (exceeding NHS limits), and TVOC reached 2618 ppb, highlighting ventilation deficiencies. The PDE-XGBoost model outperformed PDE-RF (MAE: 0.4548 vs. 0.6652; R²: 0.989 vs. 0.976), demonstrating superior predictive accuracy for IAQ parameters. Partial Dependence Plots (PDPs) revealed TVOC and noise as dominant linear predictors of PM₂.5, while temperature exhibited a nonlinear effect. SHAP analysis underscored PM₁₀ direct physiological impact and location's indirect role in pollutant dispersion. The integration of statistical and ML methods provided actionable insights: PM₁₀ mitigation, spatial optimization, and VOC control are critical for healthier environments.