Rapid indoor airflow prediction using a hybrid residual learning regression model

Ventilation performance and resulting indoor airflow patterns influence indoor air quality, thermal comfort, and energy consumption, particularly in densely occupied spaces such as classrooms. Nonetheless, accurate airflow prediction remains a challenge. Computational Fluid Dynamics (CFD) provides detailed predictions but is computationally intensive, while machine-learning (ML) models, though faster, operate as black boxes and are limited to the geometry used for training. To address this, we propose a physics-guided ML model termed Residual Learning Regression (RLR), which integrates Multivariate Linear Regression (MLR) with Extreme Gradient Boosting (XGB) residual correction. This hybrid model was trained on 149 validated classroom CFD simulations spanning the effect of diffuser geometry, air change rate, inlet temperature, and occupancy level on airflow patterns. From the RLR, four equations were derived to predict the spatially averaged values of airflow mixing, ventilation effectiveness, temperature, and velocity at the breathing level. Further, geometry-based correction factors were introduced, extending equations’ applicability to room volumes of 25–532 m 3 . Results show that the RLR improves baseline MLR accuracy by about 10 % and achieves airflow prediction comparable to XGB, while maintaining interpretability. Independent experimental validation and uncertainty analysis showed deviations within credible bounds, confirming the model’s robustness. The developed RLR equations offer a scalable and reliable alternative to CFD and black-box ML, bridging the gap between high-fidelity modeling and rapid ventilation assessment. Importantly, its transparent structure and predictive accuracy highlight the RLR model’s potential to inform future ventilation design standards and guidelines, supporting healthier and energy-efficient buildings.