The heterogeneous viscosity distribution of biodegraded heavy oil poses significant challenges for reservoir management. While molecular markers (biomarkers) reflect biodegradation intensity, existing models fail to establish quantitative correlations between biomarker signatures and viscosity due to multicollinearity in high-dimensional geochemical data. This study develops an integrated machine learning framework to decode biomarker-viscosity relationships in the Songliao Basin heavy oils. Our dual-phase methodology combines ridge regression for multicollinearity mitigation with a feedforward neural network (FFNN) to capture nonlinear interactions. Key biomarkers were identified through geochemical analysis of 17 heavy oil samples spanning PM0-PM6 biodegradation levels. The hybrid model achieved exceptional prediction accuracy (R² = 0.99996, RMSE = 3.39) through L2-regularized feature selection and neural network optimization, outperforming standalone FFNN models (cross-validation R² improvement from 0.032 to 0.99996). Reverse prediction experiments validated biomarker response patterns, even in severely biodegraded oils. The advanced machine learning model proposed in this study is applicable to predict the viscosity of heavy oil and its biomarkers, thereby improving reservoir management strategies. Additionally, this study contributes a new perspective on characterizing and managing the reservoirs of various geological backgrounds and origins, not just biodegradable heavy oil reservoirs.

Residual coal-fire gas vents form steep thermal and geochemical gradients that create sharply distinct microbial habitats over short distances. We compared the microbial communities of two microhabitats at a residual vent near Kuitun City, Xinjiang: BS1, a gas-emitting fracture at 64 °C, and BS2, adjacent cooler sediment at 43 °C. BS1 had pH 6.8 and total sulfur 5.86 g/kg, whereas BS2 had pH 8.4 and total sulfur 1.90 g/kg. Illumina sequencing of the 16S rRNA gene V3–V4 region yielded 52,327 and 64,212 merged read pairs retained for downstream analyses for BS1 and BS2, respectively. BS1 contained fewer observed amplicon sequence variants (ASVs; 524 versus 1366) and lower Shannon diversity (5.626 versus 6.588). Chloroflexota dominated both communities but was more abundant in BS1 (43.83% versus 36.33%), whereas BS2 contained larger Proteobacteria and Bacteroidota fractions. BS1 was enriched in Pyrinomonas, unclassified Chloroflexota, unclassified Anaerolineae, and Thermorudis, while BS2 was enriched in Thermogemmatispora, Massilia, and Arthrobacter. At this vent, the hot fracture therefore supported a lower-diversity, Chloroflexota-enriched community, whereas cooler sediment supported a richer and more compositionally diverse community. Together with the mineralogical observations, these results reveal sharp site-scale partitioning of bacterial assemblages across a coupled thermal-pH-sulfur gradient.