CFD–ML Approach for Predicting Breathing-Zone PM_2.5 in Indoor Cooking Environments

Yonsei University, Department of Architecture & Architectural Engineering, 03722 Seoul, South Korea

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Cooking activities generate high concentrations of PM2.5, which degrades indoor air quality and increase occupant exposure. Although Computational Fluid Dynamics (CFD) can accurately resolve airflow and pollutant distributions in such environments, the high computational cost makes it impractical for evaluating multiple ventilation conditions. This study proposes a CFD-machine learning (CFD-ML) approach to rapidly and accurately predict breathing-zone PM2.5 concentrations in a residential kitchen under varying supply airflow rates. A total of 49 CFD simulations were performed for different supply airflow conditions during cooking, and the breathing-zone PM2.5 concentration of the occupant was extracted from each case to construct a training dataset. Using this CFD-derived dataset, three regression models—Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Network (DNN)—were trained and compared in terms of their ability to predict breathing-zone PM_2.5 as a function of ventilation conditions. Model performance was evaluated using R², MAE, RMSE, and CVRMSE. All models reproduced the decreasing trend of breathing-zone PM_2.5 with increasing supply airflow; however, their prediction accuracy differed. In the test dataset, the DNN exhibited the highest performance, achieving an R² of 0.9906 and a CVRMSE of 2.65%, closely matching the CFD results across the entire concentration range. In contrast, the SVM models exhibited larger prediction errors—up to approximately 13.8%—particularly in extreme concentration ranges conditions. The average computational cost of the machine learning models was reduced by approximately 98.7% compared with CFD, demonstrating their capability to evaluate multiple ventilation cases with substantially improved efficiency. These results indicate that a DNN trained on CFD-derived data can effectively predict breathing-zone PM_2.5 in kitchen environments and provide a foundation for exposure-oriented ventilation strategy assessment and ventilation control development.