Predicting study on aroma quality prediction model for tea-flavored liquor using explainable machine learning method

Tea-flavored liquor represents a distinctive category of alcoholic beverages produced through raw material preparation, tea maceration, sugar supplementation, alcoholic fermentation, distillation, aging, and blending, resulting in a distinctive combination of tea and liquor flavors. However, its aroma profile is characterized by pronounced complexity and variability, and the accurate assessment and prediction of aroma quality continue to pose substantial challenges. To address these gaps, this study developed an explainable machine learning (ML) framework for predicting the sensory quality grades of tea-flavored liquor and identified the key volatile compounds underlying quality differentiation. A total of 110 tea-flavored liquor samples were evaluated by a professional tasting panel following standardized sensory assessment protocols. Based on the panel’s evaluation, each sample was assigned to one of three predefined quality grades (Grade A, high quality; Grade B, medium quality; Grade C, low quality). The main aroma compound profiles of tea-flavored liquor were determined using headspace solid phase micro-extraction-gas chromatography-mass spectrometry (HS–SPME–GC–MS). Thirteen ML models were benchmarked, and model performance was assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC). To enhance interpretability, Shapley Additive exPlanations (SHAP) were employed to quantify the contribution of aroma compounds to model predictions. According to the sensory evaluation results, 28 samples were classified as Grade A, 42 as Grade B, and 40 as Grade C. Among the 32 major volatile aroma compounds, 24 exhibited significant differences (P<0.05) across the three quality grades, including 10 terpenes, 10 esters, and 4 higher alcohols. Samples graded as RANK A and RANK B displayed broadly similar aroma profiles, whereas most compounds in RANK C were present at significantly lower concentrations. Among the 13 ML models tested, the Radial Support Vector Machine (Radial SVM) achieved the best predictive performance, with an AUC of 0.92 and accuracy, precision, recall, and F1-score values all exceeding 0.8. SHAP analysis further revealed that terpenes constituted the largest subgroup among the top 20 most influential compounds, followed by 7 esters and 3 higher alcohols. Key aroma compounds contributing to the model prediction included linalool (floral), anethole (spicy), methyl salicylate (mint-like), nerol (floral), ethyl undecanoate (fruity), and isoamyl alcohol (alcoholic). Linalool, anethole, methyl salicylate, and ethyl undecanoate were found to exert a positive contribution to the sensory quality of tea-flavored liquor. While nerol similarly showed a positive effect, its contribution followed a complex, nonlinear trend: its positive influence first increased, then diminished as concentrations rose. This unique pattern may be hypothesized to result from a synergistic effect between esters and other terpene compounds at lower concentrations, which collectively amplified the sensory perception of nerol. In contrast, the accumulation of isoamyl alcohol appeared to diminish the overall aroma quality of tea-flavored liquor. In conclusion, this study established an accurate and interpretable ML model for sensory quality grading of tea-flavored liquor and identified volatile compounds most critical to quality differentiation. These findings provide a scientific basis for targeted optimization of production processes, quality control, and flavor enhancement in tea-flavored liquor. Future work will focus on exploring the interactions among key aroma compounds to further enrich the theoretical foundation of flavor chemistry in alcoholic beverages.