Abstract: To investigate the formation mechanism of tobacco rod closed pressure drop and establish a predictive model, this study systematically analyzed the closed pressure drop of 34 types of samples using a tobacco rod with a circumference of 20 mm using a combined approach of CT 3D reconstruction, physical models, and the Gradient Boosting Decision Tree (GBDT) algorithm. Data quality was enhanced through preprocessing using a physical model, and the predictive performances of four types of models (traditional physical model, basic theoretical model, empirical-physical hybrid model, and machine learning model) were systematically compared. The results indicated that, 1)Among the three physical models, the empirical-physical hybrid model achieved the best predictive performance. It provided a basis for data cleaning and feature screening, enabling the identification and removal of abnormal data with residuals exceeding 100 Pa (approximately 2.6 times the median absolute deviation), thereby effectively improving the physical plausibility of the dataset. 2)The GBDT model achieved the optimal predictive performance on the independent test set (MAE = 38.58 Pa, R²= 0.85, MAPE = 7.08%), significantly outperforming all types of physical models. 3)The equivalent pore diameter extracted via 3D reconstruction is the decisive factor affecting closed pressure drop (feature importance: 76.73%). Removing this feature increased the prediction error by 78.6%. 4)The equivalent pore diameter is significantly influenced by the interaction between tobacco cut width and leaf position, enabling directional regulation of pressure drop through process adjustments. Although only validated on medium cigarettes, the "physically-guided and data-driven" integrated framework proposed in this study provides a new method for the closed-loop optimization of the tobacco industry from raw material to processing, and offers insights for the analysis of fluid flow in porous media.