Recent advances in machine learning and deep learning for total organic carbon (TOC) estimation: a comprehensive review

Accurate estimation of total organic carbon (TOC) is central to source rock evaluation, reservoir characterization, and unconventional resource assessment. While laboratory measurements provide reliable TOC values, their high cost and sparse availability motivate the use of well logs and data-driven predictive models for continuous subsurface characterization. This review presents a comprehensive and critical synthesis of conventional TOC estimation methods and recent advances in artificial intelligence (AI), encompassing machine learning (ML), deep learning (DL), and hybrid optimization frameworks. A wide range of algorithms, including ensemble learners, kernel-based methods, neural networks, graph-based models, and evolutionary optimization techniques, are systematically compared in terms of predictive performance, data requirements, computational characteristics, interpretability, and geological applicability. The review further examines the influence of input log combinations, feature engineering strategies, lithofacies, mineralogy, kerogen type, and thermal maturity on model behavior and transferability. Although ensemble models such as Random Forest and XGBoost often demonstrate strong performance across many case studies, optimized Support Vector Regression and Gaussian Process Regression provide advantages in error minimization and uncertainty quantification, respectively. Deep learning architectures, including Convolutional Neural Networks, Long Short-Term Memory networks, and Graph Neural Networks, enable advanced feature extraction and spatial dependency modeling but remain constrained by data scarcity and interpretability challenges. Beyond performance comparison, the review critically evaluates data quality, validation practices, and generalization risks, highlighting the impact of small sample sizes, inconsistent train–test splitting, potential data leakage, laboratory measurement uncertainty, depth mis-tie between logs and cores, and limited cross-formation benchmarking. The analysis confirms that no single model is universally optimal and that reliable deployment requires geology-aware validation, uncertainty-aware evaluation, and standardized benchmarking. Future directions emphasize hybrid and physics-informed learning, expanded open-access datasets, explainable AI, transferable model architectures, and tighter integration with reservoir modeling and automated interpretation workflows.