A novel hybrid machine learning and explainable artificial intelligence approaches for improved source rock prediction and hydrocarbon potential in the Mandawa Basin, SE Tanzania

Abstract

The oil and gas industry are heavily dependent on hydrocarbon resources, and the identification of potential source rocks is a critical aspect of exploration. Accurate determination of the hydrocarbon generation potential (S1 and S2), thermal maturity (Tmax), and total organic carbon (TOC) of source rocks is essential for predicting both the quantity and quality of hydrocarbons. Traditional methods for evaluating these parameters involve laboratory analyses of drill to cuttings or core samples, which can be time-consuming, expensive, and sometimes biased. Moreover, these methods may not be feasible in cases where core material is unavailable. Therefore, this study evaluates the efficacy of machine learning (ML) models for source rock prediction and hydrocarbon potential in the Mandawa Basin, Tanzania, highlighting their application in sparse data scenarios and for complementing analytical techniques. The ML models evaluated include Group Method of Data Handling Differential Evolution (GMDH-DE), Group Method of Data Handling (GMDH), and Random Forest (RF). The results show that GMDH-DE outperforms both GMDH and RF in predicting S2, S1, Tmax, and TOC. Specifically, GMDH-DE achieved R² values of 0.9932 (training) and 0.9794 (testing) for TOC prediction, highlighting its superior accuracy and generalization capability. For Tmax prediction, GMDH-DE exhibited R² values of 0.9926 (training) and 0.9802 (testing), indicating a precise fit to the data and a strong correlation between input parameters and Tmax values. Furthermore, GMDH-DE demonstrated excellent performance in predicting S1 and S2 during training/testing, with R² values of 0.9919/0.9822 for S1 and 0.9854/0.9708 for S2, showcasing its potential for assessing hydrocarbon generation capacity. The findings also suggest that the source rocks contain a combination of kerogen types II, mixed II/III and III, capable of generating both oil and gas, across a range from immature to mature thermal maturity stages.

Additionally, as GMDH-DE is a black-box model, an Explainable Artificial Intelligence (XAI) tool was integrated into this study to ensure transparency and interpretability of the model's predictions, thereby enhancing trust and reliability in its application. The XAI tool of Shapley Additive Explanation (SHAP) revealed that the model performed most accurately when wireline log measurement of bulk density (RHOB), deep resistivity (LLD), and sonic travel time (DT) were used as input variables, suggesting that these variables significantly impact the model's output.

The findings suggest that GMDH-DE is a reliable tool for source rock prediction, offering valuable insights into source rock quality, quantity, and kerogen type, especially in scenarios with sparse data or to complement traditional methods. By addressing the challenges of limited analytical data and providing robust corrections, this research contributes to advancing machine learning applications in geosciences, providing a robust methodology for improving source rock prediction and therefore to assess the charge element, which is critical for petroleum systems.