A hybrid deep learning framework for total organic carbon prediction from geophysical well logs using multi-scale feature extraction and reversible normalization

The precise estimation of total organic carbon (TOC) through well logging measurements is a critical objective in petroleum exploration and production. It serves as a fundamental metric for evaluating the hydrocarbon potential of source rocks, assessing reservoir characteristics, and guiding exploration decisions. However, predicting TOC presents challenges due to the complexity of fluid dynamics in geological structures and the diverse nature of subsurface environments. This study proposes a novel hybrid model architecture leveraging advanced computational techniques to address these challenges. The model combines multi-scale feature extraction and global modeling capabilities to enhance prediction accuracy and robustness. It captures local fluid–structure interactions and geological heterogeneities through multi-scale feature extraction, while integrating multi-head self-attention mechanisms and reversible instance normalization to model global dependencies and spatiotemporal fluid dynamics. A residual enhancement mechanism improves training efficiency and generalization performance. Experimental results show the proposed model significantly outperforms traditional methods and mainstream deep learning models in metrics such as root mean square error, mean square error, mean absolute error, and coefficient of determination, showcasing superior adaptability and accuracy in complex fluid systems. This research provides an efficient and accurate solution for TOC prediction in unconventional oil and gas exploration and contributes to the broader application of computational fluid dynamics and machine learning in geophysical and multiphase flow studies. Future work will optimize the model architecture and integrate domain-specific knowledge to enhance interpretability and applicability in fluid dynamics and subsurface flow analysis.