Multi-source data fusion empowered by machine learning enables accurate discrimination of Maotai-flavor Baijiu production techniques.

Abstract

The brewing process-particularly the distinctions among Kunsha (KSJ), Suisha (SSJ), and Fansha (FSJ)-critically governs the final quality and flavor characteristics of Maotai-Flavor Baijiu (MFB), underscoring the necessity for robust authentication methodologies. To address this challenge, this study developed an integrated analytical framework combining multi-source flavor profiling with advanced machine learning techniques. A total of 80 samples representing the three processes were comprehensively characterized using electronic nose (e-nose), electronic tongue (e-tongue), GC-FID, and HS-SPME-GC-MS. Significant disparities in volatile profiles and taste attributes were observed, with e-nose sensors exhibiting high sensitivity to nitrogen oxides and e-tongue analysis identifying sourness as a key discriminative attribute. Chromatographic analyses further identified 45 key aroma-active compounds (ROAV ≥ 1). Nine machine learning classifiers were constructed and systematically refined through Bayesian hyperparameter optimization. The CatBoost model demonstrated superior performance, achieving a classification accuracy of 97.0 %. Interpretability analysis via the SHAP framework revealed that compounds such as ethyl acetate and 2,3,5,6-tetramethylpyrazine served as core features driving process discrimination, consistent with established principles of brewing chemistry. These findings indicate that the integration of multi-source data and optimized machine learning offers a robust, data-driven strategy for process traceability and quality control in MFB production.