A machine learning-driven interpretative framework for reconstructing hydrocarbon evolution in hybrid petroleum systems

The genetic identification of hydrocarbons in complex hybrid petroleum systems remains challenging due to overlapping geochemical signatures caused by multi-source inputs and superimposed geological processes. Traditional biomarker-based methodologies often struggle to decouple these nonlinear interactions, leading to interpretive uncertainties in source correlation, thermal maturity assessment, and secondary alteration characterization. This study introduces an unsupervised machine learning framework leveraging manifold learning to resolve these challenges within the hybrid petroleum system of the eastern Junggar Basin. We employed Uniform Manifold Approximation and Projection (UMAP) to analyze high-dimensional molecular fingerprints of hydrocarbons. This approach allowed us to systematically disentangle the genetic signals influenced by multiple factors, including source material, thermal evolution, mixing, biodegradation, and migration-induced phase fractionation. Results identify two primary oil families: Permian-derived and Jurassic-sourced oils, each exhibiting unique evolutionary pathways shaped by differential thermal maturation and post-generation alterations. Spatial mapping of these genetic types reveals systematic trends in hydrocarbon accumulation, highlighting preferential migration pathways and high-potential exploration targets. This workflow not only advances the interpretation of hybrid petroleum systems but also establishes a transferable framework for optimizing exploration strategies in geochemically complex basins. The integration of machine learning with petroleum geochemistry provides a promising pathway to reconcile multi-proxy datasets, reduce interpretive subjectivity, and enhance predictive accuracy in hydrocarbon genetic studies.