Multimodal prediction model and parameter design for kitchen ceiling make-up air system

Existing kitchen make-up air system designs often overlook specific kitchen conditions and data-driven predictions. To address this, a dataset of 546 entries was created via experiments and CFD simulations. A novel XMS framework was developed, integrating XGBoost’s feature interaction with an MLP’s nonlinear mapping via stacking. This model optimizes the inlet area and airflow paths by predicting the temperature around the human body and the concentration of PM2.5 particles. The XMS algorithm reduced RMSE by 19.3% and improved R2 by 5.8% compared to polynomial regression, and outperformed a standalone MLP. For a 7.521 m2 kitchen with a 0.183 m³/s exhaust, an inlet area of 0.008–0.014 m2 and an airflow path of 5.9–6.9 m effectively maintained a body temperature above 290 K with low breathing-zone PM2.5. This study aids in selecting optimal ceiling make-up air systems for diverse kitchens.