Flavoromics and machine learning for baijiu flavor and quality prediction

Subjective sensory evaluation dominates the Chinese baijiu industry but lacks objectivity and throughput. We developed a UPLC-TQMS method quantifying 130 key flavor compounds across eight chemical classes within 13 min per run. Leveraging this analytical protocol, we constructed a comprehensive Baijiu flavor database comprising 2048 samples covering all 12 standard flavor types and three quality grades of Nongxiang base liquor. Applying Random Forest algorithms, we discriminated among flavor types with accuracy >0.93, whereas XGBoost performed optimally in quality grading with accuracy of 0.90. Moreover, validation with an independent dataset confirmed the model's reliability and generalization capacity. SHAP analysis identified key chemical markers driving flavor differentiation and quality distinctions, underscoring the strength and interpretability of our approach in decoding complex flavor matrices. This work establishes a large-scale, standardized digital flavor library for Baijiu, providing a robust framework for intelligent quality control and industry standardization.