Machine learning combined with GC-FID for discrimination of different categories of maotai-flavor baijiu
Liang Yang, Chun Xian, Shuai Li, Ye Wang, Xinying Wu, Qingcai Chen, Wenwu Zhao, Cheng Zhao, Xiaobo Li, Junjun He, Renyuan Chen, Chunlin Zhang
Abstract
Maotai-flavor Baijiu, a traditional Chinese liquor produced via solid-state fermentation, exhibits diverse base Baijiu types due to variations in fermentation rounds, styles, and grades. While crucial for flavor complexity, current manual identification methods hinder blending efficiency and quality control. This study employed GC-FID and machine learning to analyze 410 base Baijiu samples. Decision Tree (74.36 %), XGBoost (92.9 %), and Random Forest (62.3 %) emerged as optimal classifiers for fermentation rounds, typical styles, and Chuntian grades, respectively. SHAP analysis revealed: (1) esters as primary markers for fermentation rounds, (2) ester-trimethylbutanol combinations for grade differentiation, and (3) multi-compound signatures (butyric acid, tetramethylpyrazine, 2-butanol et al.,) for style discrimination. Notably, marker compounds' flavor properties - beyond mere concentration - critically influenced their discriminative power, as evidenced by correlations with nine sensory dimensions.