Integration of Rock Fabric Features with a Self-Attention Mechanism for Shale Oil Total Organic Carbon Evaluation

Shale oil has become a global research focus for the unconventional reservoir exploration and development. Total organic carbon (TOC) constitutes a key parameter for evaluating the hydrocarbon potential of the shale reservoirs. Rock fabric is recognized as one of the critical controlling factors for the TOC. This study proposes a novel method for the quantitative characterization of rock fabric using resistivity image log data. It achieves a deep integration of reservoir geological features with artificial intelligence. First, the gradient tensor lamina density (GTLD) method is applied to resistivity image log static images. This process extracts a laminar index to quantitatively characterize the rock fabric features. Second, conventional log data suitable for model input are screened by analyzing their relationship with TOC. Two input data sets, with and without the laminar index, are constructed for the study area to verify the impact of rock fabric features on the TOC evaluation accuracy. Third, three network models are established to evaluate TOC. These models are a bidirectional long short-term memory network, a convolutional neural network-BiLSTM hybrid, and an Atten-BiLSTM model incorporating a self-attention layer. Prediction results of all models show the superior performance of the lamina-index-augmented Atten-BiLSTM model. This model achieves the highest coefficient of determination and the lowest error for the training, validation, and test sets. Application to well data further confirms that the Atten-BiLSTM model provides the highest prediction accuracy. This study elucidates the relationship between rock fabric characteristics and the TOC content of the shale oil reservoir. This proves that incorporating rock fabric information enhances the accuracy of TOC evaluation. The results also show the significant advantage of the self-attention mechanism for handling complex nonlinear regression problems. This research not only improves the accuracy of TOC evaluation in complex shale oil reservoirs but also provides a new idea for the development of shale oil reservoirs. It also provides a novel technical pathway for petrophysical evaluation of unconventional hydrocarbon reservoirs.