Quantitative mining of wine aroma literature data for specification reuse

Wine aroma datasets are abundant yet hard to reuse due to dramatic cross-study differences in analytical platforms, quantification protocols and reporting formats. Quantitative volatile data published between 1998 and 2026 were curated and normalized for compound identities, concentrations, analytical metadata and odor thresholds. A total of 103 valid studies were screened from more than 5000 publications, and the dataset was constructed with records of 792 volatile compounds, 38 grape cultivars and samples from 22 countries. An initial panel of 31 exploratory candidate varietal markers was selected from the intersection of PCA, PLS-DA and random forest screening. A proof-of-concept evaluation used HS-SPME-GC-MS and sensory analysis of 24 commercial Cabernet Sauvignon and Marselan wines. The commercial-wine PLS-DA model yielded R²Y(cum) = 0.971 and Q²(cum) = 0.947, with an internal ROC-AUC of 0.840; no independent external test set was available. A framework for reusing literature-based wine aroma data was explored in this work. Incomplete metadata, inconsistent quantification approaches, missing-value handling and unbalanced varietal and geographical coverage remain limitations requiring independent validation under harmonized analytical conditions.