CFD-based Surrogate Model for Predicting CO

University classrooms often experience elevated CO2 concentrations due to high occupant density, which can adversely affect attentiveness, induce drowsiness, and impair learning performance. Conventional ventilation control systems typically rely on CO2 measurements from a single fixed sensor, which fails to capture spatial variations arising from differences in occupant number and seating distribution. Although Computational Fluid Dynamics (CFD) can provide detailed spatial predictions of indoor airflow and contaminant dispersion, its high computational cost and long simulation times limit its applicability for realtime evaluation under diverse occupancy conditions. To address this limitation, this study develops a Kriging-based surrogate model for efficient prediction of breathing-zone CO2 concentration distributions. CFD simulations were performed for 23 scenarios combining ventilation rates (150, 250, and 400 CMH), occupant numbers (4, 8, 12, 16, and 20), and seating patterns (Front, Back, and Random). Seat-level breathing-zone CO2 concentrations were extracted from CFD results and used to construct the training dataset. The surrogate model is based on Gaussian Process Regression (GPR) and employs a combination of Constant, Matern 2.5 and White kernels to capture spatial correlation structures. Model performance was evaluated using the coefficient of determination (R2) and mean absolute error (MAE). The developed surrogate model successfully reproduced CFD results across various occupancy conditions and predicted breathing-zone CO2 distributions with substantially reduced computation time. Owing to its high accuracy and low computational demand, the proposed model demonstrates strong potential as a predictive tool for real-time ventilation control and demand-controlled ventilation strategies under dynamic occupancy conditions.