Dynamic Estimation of PM2.5 Penetration and Removal Rates Using Physics-Informed Neural Networks for Indoor Air Quality Management

Abstract

Quantifying indoor air pollutant dynamics is crucial for assessing exposure risks and optimizing ventilation strategies. This study advances previous research by developing a Physics-Informed Neural Network (PINN) model that dynamically estimates ventilation rate, penetration factor, and particulate removal rate in real time. The model integrates space operation factors (e.g., occupancy, window/door status, air purifier and air conditioner use) and meteorological variables (e.g., temperature, humidity, wind conditions, and outdoor PM₂.₅ levels) to predict indoor PM₂.₅ behavior without assuming static coefficients.

A key contribution of this study is the application of SHapley Additive exPlanations (SHAP) to quantitatively analyze the influence of each variable. The results indicate that outdoor humidity, window opening, and occupancy significantly impact the penetration factor, while air purifier operation, occupancy, and window opening play major roles in particulate removal. Notably, this study identifies a previously unreported effect: occupancy enhances removal rates due to particle inhalation, allowing for a direct estimation of personal exposure. Specifically, the mass flow rate of PM₂.₅ inhaled per occupant is approximately 10 times the indoor PM₂.₅ concentration (μg/hour). This approach refines traditional exposure assessments by quantifying PM₂.₅ uptake per person.

While the model is currently specific to a single measured space, it provides a practical tool for real-time air quality management. Future research will focus on expanding its applicability through long-term data collection across diverse environments and integrating reinforcement learning to optimize air quality control strategies. This study lays the groundwork for adaptive ventilation management, balancing air quality improvements with energy efficiency.