Study on draw resistance of tobacco-shred packed rods based on structural parameters using machine learning--颗粒学报PARTICUOLOGY

a College of Automotive and Energy Engineering, Tongji University, Shanghai, 201804, China
b Shanghai Tobacco Group Co., Ltd., Technology Center, Shanghai, 200082, China

10.1016/j.partic.2026.08.025

Volume 118, November 2026, Pages 75-87

Received 2 July 2026, Revised 20 August 2026, Accepted 26 August 2026, Available online 26 August 2026, Version of Record 2 September 2026.

E-mail: gaonaiping@tongji.edu.cn

Highlights

• CT-guided machine learning improved prediction accuracy and identified key resistance factors.

• Pore size distribution alone enabled accurate draw resistance prediction.

• Porosity, average pore diameter, and the pore size distribution ratio Q₉₀/Q₁₀ were the dominant predictors.

• Adding Q₉₀/Q₁₀ improved permeability model performance by up to 26.6%.

Abstract

The cigarette rod is a typical porous medium, and its complex internal pore structure has a significant influence on draw resistance. Conventional permeability models, such as the Kozeny–Carman (K–C) model, primarily rely on bulk structural parameters and idealized assumptions. They have limited ability to characterize the heterogeneous pore networks and pore size distributions of cigarette rods, making it difficult to accurately predict draw resistance. To accurately predict cigarette draw resistance, this study used computed tomography (CT) to reconstruct the three-dimensional (3D) pore structure of cigarette rods. A machine learning (ML) model was then developed based on the extracted pore structure parameters, followed by interpretability analysis to reveal their relationship with draw resistance. First, draw resistance was measured for 12 types of cigarette samples. Then, 12 cigarettes from each category were scanned by CT to reconstruct the 3D pore structure and extract porosity, specific surface area (SSA), tortuosity, pore-scale structure parameters, including average pore diameter and pore length-to-diameter ratio, and pore size distribution parameters. Based on these data, a random forest (RF) model was developed to map pore structure features to draw resistance, and feature importance analysis with ablation studies was used to evaluate the effect of different feature combinations on prediction accuracy. The results show that, compared with the traditional K–C model, the RF method significantly improved the prediction accuracy of cigarette draw resistance. After introducing pore-scale structure parameters, the model performance was further improved. The R² of the selected multi-parameter model reached 0.925. Further interpretability analysis indicated that average pore diameter, porosity, and the pore size distribution parameter Q₉₀/Q₁₀ were the key structural factors affecting draw resistance. Accordingly, incorporating Q₉₀/Q₁₀ into the traditional permeability model effectively improved the prediction accuracy by up to 26.6%. This study reveals the relationship between pore structure and draw resistance using an ML method, providing an effective approach for predicting cigarette draw resistance.

Graphical abstract

Keywords

Draw resistance; Porous media; CT-based three-dimensional reconstruction; Machine learning; Permeability model

文章导读