Integrating microbiome, volatilome, sensory analysis, and temporal deep learning to model flavor formation in Fengxiang Taibai Baijiu

Fengxiang-type Taibai Baijiu (FTB) has an elegant, balanced aroma, yet the mechanisms underlying its flavor formation remain unclear. We integrated fermentation-stage microbiome, volatile, and sensory data across fermentation and aging, and applied a temporal fusion transformer model to resolve dynamic flavor evolution from temporal associations. Leveraging multi-horizon dependencies, we identified consistent time-lagged predictive relationships: microbial dynamics predicted subsequent compound changes, and compound profiles predicted later sensory attributes. During fermentation, Bacillus , Brevibacterium , and Weissella were associated with key precursors, while Pichia kudriavzevii was associated with esterification. Correlation analysis identified ethyl acetate, ethyl hexanoate, isoamyl alcohol, and 2-methylpropanoic acid as core positive contributors across both stages, whereas 3-hydroxy-2-butanone and pentanol negatively correlated with cereal and fermented aromas. During aging, ethyl butanoate and 1-butanol were associated with roasted and Feng-flavor notes. Reconstitution experiments validated these associations for 16 compounds and revealed increased contribution of ethyl 2-methylbutanoate after aging. All associations represent temporal and predictive relationships, not proven causal mechanisms. This study provides a data-driven framework for FTB flavor formation and an AI-based multi-omics approach for traditional fermentations.