A novel hybrid group method of data handling and Levenberg Marquardt model for estimating total organic carbon in source rocks with explainable artificial intelligence

The estimation of Total Organic Carbon (TOC) in source rocks is critical for assessing hydrocarbon potential and source rock quality in petroleum exploration and production. This study presents a novel hybrid model of the Group Method of Data Handling and Levenberg Marquardt (GMDH-LM), which integrates mineralogical and well-log data to improve TOC prediction accuracy. The GMDH-LM model employs non-parametric regression, offering superior adaptability and accuracy in solving non-linear problems, even with small datasets. The results show that the GMDH-LM model achieved a higher coefficient of determination (R 2 ) of 0.9784, lower Root Mean Square Error (RMSE) of 0.1162, Mean Absolute Percentage Error (MAPE) of 0.3722 and less computation time of 0.73 s. Furthermore, the integration of mineral constituents into the model significantly improved the accuracy to R 2 = 0.9983 as compared to where only well log data is used. Additionally, the application of Shapley Additive Explanation (SHAP) analysis presents a key innovation in this work as a novel method in TOC estimation, offering unprecedented transparency and interpretability of the GMDH-LM model by identifying gamma ray (GR), calcite , K-feldspar and deep lateral resistivity (LLD) as the most impactful input variables, thus contributing to the model's trustworthiness, reliability, and overall performance. This interpretability enhances model reliability and supports informed decision-making. The proposed GMDH-LM model represents a significant advancement in hybrid machine learning applications, offering a robust and efficient tool for TOC estimation. Its scalability and precision make it valuable for petroleum reservoir characterization and exploration, with broader implications for optimizing resource management in the oil and gas industry. • Hybrid group method of data handling was proposed to predict total organic carbon. • The GMDH-TOC model was optimized using the modified Levenberg Marquardt algorithm. • SHAP was used to determine the contribution of input data to the model's output. • SHAP identified gamma ray, calcite, and K-feldspar as influential TOC predictors. • Explainable AI enhances trust and interpretability in the petroleum industry.