This paper is the result of oil and gas exploration engineering.

Objective Source rocks are the fundamental material basis for hydrocarbon accumulation. However, due to the limited number of drilling core samples, geochemical analyses alone are often insufficient to provide a systematic evaluation of source rock characteristics throughout the entire well interval. Therefore, utilizing well-log data combined with artificial intelligence to conduct continuous quantitative evaluation of total organic carbon (TOC) in source rocks and source rock prediction is of great significance for a comprehensive and in-depth analysis of hydrocarbon accumulation conditions.

Methods This paper comprehensively investigated the geological characteristics of source rocks, as well as logging-based evaluation and intelligent prediction methods for total organic carbon (TOC), using geochemical analysis data of rock samples and logging data from key wells in the Permian Pusige Formation of the Southwestern Tarim Basin.

Results The source rocks of the Pusige Formation are mainly composed of dark mudstones, with TOC values ranging from 0.01% to 20.6% (average 1.13%). The hydrocarbon generation potential (Pg) varies from 0.02 mg/g to 23.31 mg/g (average 2.76 mg/g). The kerogen types are predominantly Type Ⅱ₂ and Type Ⅲ, while vitrinite reflectance (R_o) values are mainly distributed between 0.55% to 2.55%, indicating a mature stage of thermal evolution. Overall, the source rocks are classified as fair to good source rocks. Due to the influence of argillaceous limestones and other lithologies with abnormally high resistivity, the conventional ΔlgR method exhibits poor applicability. To overcome this limitation, this study employed gamma ray (GR), deep resistivity (M2Rx), compensated neutron porosity (CNC), and compensated density (DEN) logging curves to establish a quantitative TOC prediction model using multiple linear regression and random forest machine learning methods. The random forest model performed significantly better than the multiple regression model, enabling continuous quantitative evaluation of source rock TOC in single-wells of the study area.

Conclusions The Permian Pusige Formation in the Southwestern Tarim Depression develops fair-to-good hydrocarbon source rocks. Compared with the ΔlgR method and multiple regression algorithm, the random forest algorithm can achieve more accurate continuous quantitative prediction of total organic carbon (TOC) content of source rocks along the full well interval in the study area. The Pusige Formation within the research block is characterized by sufficient hydrocarbon-generating material foundation, favorable reservoir-cap rock assemblages and advantageous structural migration and accumulation conditions, demonstrating considerable hydrocarbon exploration potential.