Chinese baijiu is one of the six most renowned distilled spirits in the world (Qiao et al., 2023). Among them, sauce-flavor baijiu is particularly notable for its rich and diverse aroma and flavor profile, which is closely linked to its intricate, multi-stage fermentation process and extended production cycle (Wang et al., 2024; Wu et al., 2021), but these features also pose challenges for regulating and maintaining stable baijiu yields.

The brewing process of sauce-flavor baijiu involves two rounds of grain feeding, nine rounds of steaming, eight fermentations, and seven distillations(Duan et al., 2022; Jin et al., 2017). Specifically, the jiupei (fermented grains) are first gelatinized by steaming, then subjected to open-air stacking fermentation, followed by approximately 30 days of anaerobic pit fermentation before distillation(Fig. S1). The distilled liquor obtained at each stage serves as the base baijiu for sauce-flavor baijiu production (Li et al., 2024; Yang et al., 2025a). Notably, in the seventh round, only pit fermentation and distillation are conducted, without stacking fermentation. This multi-round steaming and continuous fermentation process, coupled with seasonal environmental fluctuations during the lengthy production cycle, inevitably leads to differences in the microecological structure of jiupei across fermentation rounds (Li et al., 2025a; Wei et al., 2024). However, how these microecological structure differences influence the yield variation and stability of base baijiu across fermentation rounds remains unclear.

Previous studies on sauce-flavor baijiu microbiota have predominantly emphasized the functional attributes of dominant taxa. For example, acid-tolerant Acetilactobacillus jinshanensis—with relative abundance up to 92%—is the primary acid producer during fermentation (Chen et al., 2024), while Pichia kudriavzevii (54.65%) plays key roles in community assembly and metabolic regulation (Zhang et al., 2021). Despite these advances, systematic classification of microbial communities based on abundance—such as always rare taxa (ART) and conditionally rare taxa (CRT)—has received relatively little attention (Dai et al., 2022; Li et al., 2023). Emerging evidence indicates that rare taxa can strengthen microbial interaction networks, enhance environmental resilience, and participate in specialized metabolic pathways, thereby influencing flavor compound formation and potentially contributing to yield stability (Chakraborty et al., 2025; Mu et al., 2024; Yuan et al., 2023). For instance, Zhou et al. demonstrated that the rare anaerobe Clostridium tyrobutyricum produces multiple medium- and short-chain fatty acids and fatty alcohols during pit fermentation (Zhou et al., 2023). Therefore, integrating abundance-based microbial classification into microecological analysis may provide deeper insight into the ecological mechanisms underlying yield variation and support more precise regulation of baijiu fermentation.

In recent years, artificial intelligence (AI) has been increasingly applied in the baijiu brewing industry (Niu et al., 2025; Zhu et al., 2025). Machine learning, with its capabilities in feature extraction and multivariate modeling, enables effective handling of high-dimensional data, complex variable interactions, and latent patterns in the brewing process (Gite et al., 2025; Menichetti et al., 2023; Schreurs et al., 2024). Among these approaches, stacking ensemble learning hierarchically integrates multiple complementary base learners via a meta-learner, yielding superior generalization and robustness over single models when handling nonlinear relationships, feature interactions, and heterogeneous noise (Xiao et al., 2025; Zhang et al., 2025b). Previously, our team successfully addressed challenges in quality classification and prediction of sauce-flavor baijiu using various machine learning algorithms (Li et al., 2024; Li et al., 2025b; Li et al., 2025c; Yang et al., 2025b). However, key challenges remain unresolved, particularly the high complexity of microbial community structures and strong dynamic coupling of process parameters, which continue to hinder stable base baijiu yield in production (Pan et al., 2023; Zhang et al., 2025).

To address these challenges, this study investigated jiupei samples from stacking and pit fermentation stages and their corresponding base baijiu yields across fermentation rounds 1 to 7. Multivariate statistical analyses were conducted to elucidate yield differences among fermentation rounds and their correlations with physicochemical parameters. Based on a microbial abundance classification framework, ecological network models combined with random forest algorithms were employed to identify key microbial taxa associated with yield variation and regulating community assembly and stability. Furthermore, a stacking ensemble model was developed for online yield prediction, with SHAP analysis used to determine critical physicochemical thresholds. This study integrates microecology and machine learning to elucidate the mechanisms underlying baijiu yield variation, providing a scientific basis for yield stabilization and supporting the intelligent transformation of traditional baijiu fermentation practices.