Logging-based prediction of organic geochemical parameters in oil shale during thermal evolution using the XGBoost algorithm

Oil shale in large basins undergoes multiple evolutionary stages, limiting the applicability of a single logging-based prediction model. This study focuses on the oil shale of the Qingshankou Formation in the Songliao Basin, using gamma ray (GR), deep resistivity (LLD), acoustic travel time (DT), neutron porosity (CNL), density (DEN), and depth data as input features. The XGBoost algorithm is employed to develop predictive models for total organic carbon (TOC) content, free hydrocarbon (S1), pyrolyzable hydrocarbon (S2), and maximum pyrolysis peak temperature (Tmax). TOC predictions are further stratified for low-maturity, mature, and high-maturity oil shale intervals. The results show that S2 achieves the highest prediction accuracy (R2 = 0.91), due to its strong correlation with hydrogen index (HI) driven by thermal evolution. TOC prediction accuracy (R2= 0.75) is influenced by combined changes in porosity and organic matter evolution. Tmax prediction (R2 = 0.74) depends mainly on depth and CNL. S1 correlates weakly with all well logs, yielding the lowest accuracy (R2= 0.29). Shale maturity plays a critical role in determining the reliability of TOC prediction models. Low-maturity oil shale exhibits the best TOC accuracy (R2= 0.83), as wellpreserved organic matter and high porosity correlate closely with LLD, DT, CNL, and DEN. In mature oil shale, retained hydrocarbon and reduced porosity weaken logging signals, lowering accuracy to R2 = 0.63. In high-maturity intervals, hydrocarbon expulsion and porosity rebound improve accuracy (R2 = 0.69). Our approach provides a cost-effective, continuous method for evaluating lacustrine oil shale resources. It is particularly applicable to the evaluation of uncored wells.