Tea-flavored liquor can often represent one of the most favorite alcoholic beverages in recent years. A combination of tea and liquor flavors can be produced after raw material preparation, tea maceration, sugar supplementation, alcoholic fermentation, distillation, aging, and blending. However, its aroma profile is characterized by pronounced complexity and variability. Substantial challenges also remained to accurately assess and then predict the aroma quality. In this study, an explainable machine learning (ML) framework was developed to predict the sensory quality grades of the tea-flavored liquor. The key volatile compounds were identified under quality differentiation. A total of 110 tea-flavored liquor samples were taken after identification. A tasting panel was selected for the standardized protocols of the sensory assessment. Each sample was assigned to one of three predefined quality grades (Grade A, high quality; Grade B, medium quality; Grade C, low quality) after evaluation. The aroma compound profiles of the tea-flavored liquor were then determined using headspace solid phase micro-extraction-gas chromatography-mass spectrometry (HS–SPME–GC–MS). Furthermore, the 13 ML models were benchmarked. The performance of the model was assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC). Shapley Additive exPlanations (SHAP) were employed to quantify the contribution rates of the aroma compounds to the predictions, in order to enhance the model interpretability. According to the sensory evaluation, 28 samples were classified as Grade A, 42 as Grade B, and 40 as Grade C. Among the 32 volatile aroma compounds, 24 exhibited significant differences (P < 0.05) over the three quality grades, including 10 terpenes, 10 esters, and 4 higher alcohols. The samples of the RANK A and RANK B grades also displayed broadly similar aroma profiles, whereas most compounds in the RANK C were presented at significantly lower concentrations. Among the 13 ML models, the Radial Support Vector Machine (Radial SVM) achieved the best performance in the prediction, with an AUC of 0.92 and all accuracy, precision, recall, and F1-score values exceeding 0.8. The SHAP analysis further revealed that the terpenes constituted the largest subgroup among the top 20 most influential compounds, followed by 7 esters and 3 higher alcohols. Key aroma compounds contributed to the prediction, including linalool (floral), anethole (spicy), methyl salicylate (mint-like), nerol (floral), ethyl undecanoate (fruity), and isoamyl alcohol (alcoholic). Linalool, anethole, methyl salicylate, and ethyl undecanoate greatly contributed to the sensory quality of the tea-flavored liquor. While the nerol similarly shared a positive correlation, where its contribution rate followed a complex, nonlinear trend: Its positive influence first increased, then diminished as the concentrations rose. There was a synergistic interaction between esters and other terpene compounds at the lower concentrations. Collectively, the sensory perception of nerol was amplified after interaction. By contrast, the isoamyl alcohol was accumulated to diminish the overall aroma quality of the tea-flavored liquor. In conclusion, an accurate and interpretable ML model can be expected to identify the volatile compounds most critical to quality differentiation, particularly for the sensory quality grading of tea-flavored liquor. These findings can provide a scientific basis for the targeted optimization of the production, quality control, and flavor enhancement in the tea-flavored liquor. Future work can be expected to focus on the interactions among key aroma compounds, in order to enrich the theoretical foundation of the flavor chemistry in alcoholic beverages.

The study aimed to provide technical guidance for process optimization, aging management, and market promotion of wines from specific geographical regions by systematically analyzing the flavor quality evolution of dry red wines during aging and constructing a predictive model for flavor evolution. The experiment focused on Cabernet Sauvignon dry red wines from Turpan, with 81 samples collected across 2-15 years of aging. Conventional physicochemical indicators (e.g., alcohol content, total acidity, volatile acidity) were analyzed using national standard method (GB/T 15038-2006), while flavor characteristics (e.g., appearance, aroma, mouthfeel) were quantified by a trained tasting panel. Sensory data underwent cluster analysis and principal component analysis (PCA), followed by weighted calculation of PCA loading contribution rates for sensory traits. A Gaussian Process Regression (GPR) model was developed using nonlinear least squares fitting to construct a mathematical model for wine flavor evolution. Results showed that the tested wines met national physicochemical standards. Cluster analysis of sensory data revealed three distinct groups (2-6 years, 7-9 years, >9 years) with significant intra-group sensory similarity, indicating stage-specific differentiation of flavor traits across aging. PCA demonstrated that the first two principal components (PC1 and PC2) accounted for 92.9% and 2.3% of total variance, respectively. The "overall impression" exhibited the highest weight on PC1, while astringency and acidity showed minimal weights. The flavor evolution model indicated a 95% confidence interval covering 0-12.40 years of aging. Sensory quality initially increased to a peak at 2.95 years, then gradually declined until 7.77 years, after which maximum degradation occurred. Model outputs showed that wines aged 2-4 years achieved high scores in appearance, aroma, and mouthfeel; aged 5-8 years displayed slow degradation in mouthfeel and aftertaste but retained basic quality; wines exceeding 9 years of aging exhibited intensified oxidation and significant sensory decline. The integration of PCA and Gaussian process regression (GPR) provided a novel framework to capture the nonlinear evolution of sensory quality, overcoming the limitations of traditional linear models. The hybrid approach identified critical inflection points and quantified the contributions of core sensory attributes to overall quality. The model’s ability to predict aging-related changes in flavor profiles offers a data-driven tool for optimizing production consistency and storage protocols in wine manufacturing. Notably, sensory variability among same-vintage samples highlighted the need for improved standardization in grape selection, maceration control, and oak aging practices. The study emphasized that expanding the sample size and refining process parameters could enhance the model’s reliability and reduce vintage-to-vintage inconsistencies. The PCA-GPR model demonstrated its scientific and industrial relevance by bridging the gap between sensory science and data-driven modeling. By constructing a region-specific sensory weight system through principal component analysis, the regional adaptability of the evaluation model was significantly improved. Furthermore, GPR through kernel function design, can flexibly capture the nonlinear stage evolution characteristics of flavor quality and wine age, and utilize the covariance matrix to dynamically assess the uncertainty of flavor decay in products of different vintages, providing a probabilistic decision-making basis for storage risk management. Overall, the research provides a foundation for developing region-specific quality assurance systems, enabling wineries to balance aging duration with economic feasibility and improve product value through precise flavor management.

Dry red wines can be sourced from the Ningxia and Xinjiang regions of China. This study aims to characterize the taste quality, according to the conventional and easily measurable polyphenol series indices. The UV-Vis spectrophotometer was used to detect the conventional polyphenols in wine. A predictive model was constructed to evaluate the astringency and aftertaste of dry red wine using the polyphenol indices. A total of 50 (Cabernet Sauvignon) dry red wines were selected as the research objects (20 from Ningxia and 30 from Xinjiang, each from different small producing areas). The total acidity, volatile acidity, and pH of the wines were tested to evaluate the acidity, according to GB/T 15038-2006. The antioxidant activity of wine was determined by DPPH scavenging activity. Total phenols were determined with the UV-Vis spectrophotometer using the Folin-Ciocalteu method. Flavanol content was determined by p-DMACA hydrochloric acid. The contents of tartrate (A₃₂₀), flavonol (A₃₆₀), and total anthocyanins (A₅₂₀) in the wines were determined at different absorbance levels. The aftertaste and astringent sensation were evaluated using the quantitative sensory description in Panel Check software. Correlation analysis (CA) was used to analyze the interaction between the indices of the wine polyphenol series, and linear discriminant analysis (LDA) was to easily distinguish different regions. The modeling samples were analyzed with the principal component analysis. The taste quality of wine samples was characterized and predicted with partial least squares regression. The CA results indicated that there were significant differences (P<0.05) in the effects of total phenols, anthocyanins, flavanols, and flavonols on the astringency and aftertaste of the dry red wines. However, there was a minimal impact of tartrate on the aftertaste. Furthermore, the total acids and pH had no significant (P>0.05) effect on the astringency of the wine. Taking the Yongning region of Ningxia as an example, LDA easily differentiated dry red wine samples from the Yongning of Ningxia and the Xinjiang region using polyphenol series indices. Two discriminant functions generated by LDA were used to explain 80.8% and 19.2% of the variance, respectively. The accurate discriminant rate of 80.0% was obtained after classification. While there was the correct discriminant rate of 62% after the cross-validation. The sensory evaluation showed that the astringency scores were 6.18 and 5.70, in Ningxia and Xinjiang wine samples, respectively, while the aftertaste scores were 5.37 and 4.85, respectively. Therefore, the total phenols, anthocyanins, flavanols, flavonols, and DPPH contributed significantly to the astringency and aftertaste of wine. Tartrate mostly influenced the wine astringency, whereas the total acids and pH dominated the wine aftertaste. Polyphenol indices with cost-saving UV-Vis spectrophotometers can be expected to better assess the taste quality of dry red wine. The sample size can be increased to promote the taste quality of dry red wine for the accuracy and stability of the model.

The purpose of this study was to develop a rapid non-destructive detection on of polysaccharides in the dry red wine by the attenuating total reflection-Fourier transform infrared spectroscopy (ATR-FTIR) with chemometric technologies. 100 dry red wines were also collected from the Xinjiang production area of western China. Alcohol precipitation was utilized to extract the wine polysaccharides from these test materials. After that, the polysaccharide powder was obtained after vacuum freeze-drying. Then, the polysaccharides were decomposed into the monosaccharides with the trifluoroacetic acid. The monosaccharides were quantitatively characterized by high high-performance liquid chromatography (HPLC-PDA). As such, the different types of polysaccharides were identified from the wine samples. The monosaccharide concentration was calculated to compare their respective characteristic structures in the molar ratio, including total soluble polysaccharides (TSP), mannosan protein (MP), arabinose-galactose-rich polysaccharides (PRAG), rhamnogalacturonic acid glycan type II (RG-II), homogalacturonized glycan (HG), and dextran (GL). The mid-infrared spectra of wine was were also collected by ATR-FTIR. The spectral preprocessing was carried out by standard normal transform (SNV) and multivariate scattering correction (MSC). The competitive adaptive reweighting algorithm (CARS) was followed for band screening. Finally, the partial least squares regression (PLSR) and backpropagation neural network (BPNN) were combined to simulate, predict and evaluate the indicators. The spectral characteristic information in the 1 900-900 cm^-1 band was screened to determine the content of several polysaccharide substances that were measured by HPLC-PDA. The results indicate that the content of various polysaccharide varied greatly among the test liquor samples, with the TSP content of (859.41±293.65) mg/L, MP (208.08±78.42) mg/L, PRAG (418.30±140.00) mg/L, RG-II (113.17±55.11) mg/L, GL (95.46±62.10) mg/L, and HG (24.41±55.86) mg/L. The content of several polysaccharides in the test wine samples was also verified using the linear and nonlinear correction. The ATR-FTIR model shared the a better prediction on the content of several polysaccharides in wine. The PLSR model showed the better performance than the BPNN. The coefficient of determination (R_c²) values of the PLSR model between the characteristic bands and the content of polysaccharides (TSP, MP, PRAG, RG-II, and GL) were 0.98, 0.96, 0.92, 0.99, 0.98, respectively. The coefficient of determination (R_p²) values were 0.85, 0.92, 0.83, 0.83, 0.84, respectively. The relative analysis errors (RPDc) in the training set were 6.50, 5.31, 3.62, 9.10, and 7.86, respectively. The relative analysis errors (RPD_P) of the prediction set were 2.68, 3.99, 2.44, 2.52, and 2.37, respectively. Therefore, the ATR-FTIR can be expected to detect the polysaccharides in dry red wine. The content of polysaccharides can be accurately predicted in the TSP, MP, PRAG, RG-II, and GL, according to the spectral characteristic band of 1 900-900 cm^-1. The finding can provide the application potential for the rapid and nondestructive detection of polysaccharides in dry red wine.