Advanced source rock characterization integrating pyrolysis, petrophysical logs, and machine learning in the unconventional Cane Creek reservoir, Utah

Abstract

The Cane Creek clastic interval of the Paradox Formation in the northern Paradox Basin, presents significant challenges and opportunities for hydrocarbon exploration, particularly in its complex lithologic and structural formations. Through the combination of source rock analysis, petrophysical data and machine learning methods, this innovative predictive model aims to identify highly productive hot-spots in moderate fractured dolomitic intervals. The petrophysical logs of gamma-ray, resistivity, density, porosity are processed including a rigorous log quality control, statistical pre-processing, and lithology controls. Results from laboratory pyrolysis measurements indicated TOC values up to 14.5 wt%, with thermal maturity levels assessed through Tmax values spanning from 431 °C to 484 °C. These findings suggest variable maturity and highlight zones conducive to condensate and gas production. Key findings include total organic carbon (TOC) content across various clastic intervals, notably clastics 2, 3, 10 and 21 (Cane Creek), which exhibited significant hydrocarbon generation potential. This study integrates machine learning models, with legacy petrophysical logs to improve source rock characterization. The GPR model demonstrated superior predictive capabilities compared to other ML models such as Regression Tree Models (RTM), Support Vector Machines (SVM), and Neural Networks (NN). While GPR achieved an R-squared of 0.43 and an RMSE of 1.00 during validation, other models showed lower R-squared values and higher errors, underscoring GPR’s reliability for complex geological data interpretation. The integration of petrophysical, geochemical, and mineralogical data identified fractured dolomitic intervals as key targets, supporting enhanced hydrocarbon migration and accumulation and identification of potential hydrocarbon production zones. Furthermore, the ML predictive model is adapted to other wells in the region that lack source rock analysis.