Data-driven lithofacies classification of marine shale based on deep learning approaches

Shale lithofacies control hydrocarbon generation potential, reservoir properties, and anisotropy, thereby determining the distribution of shale “sweet spots” and guiding the selection of horizontal well target windows. Consequently, predicting shale lithofacies is crucial for the detailed evaluation, efficient exploration, and development of shale reservoirs. However, predicting lithofacies within the Wujiaping Formation of the Permian in the Hongxing area, eastern Sichuan Basin, presents significant challenges due to complex mineral compositions and strong heterogeneity. In this study, we classified the shale and its interlayers within the reservoir into five lithofacies based on total organic carbon content and mineral assemblages, with MF4 identified as the favorable lithofacies. Additionally, six logging curves (GR, AC, DEN, CNL, CAL, and LLD) were selected as input features to train and test deep learning models for automatic lithofacies prediction. To investigate the applicability of deep learning models for lithofacies identifications using well logs, six models were employed, including CNN, CNN-LSTM, CNN-BiLSTM, CNN-GRU, CNN-BiGRU, and Transformer. Notably, the Transformer model outperformed the others, achieving an $\mathrm{Accuracy}$ of 0.8776 and $\mathrm{Precision}$ of 0.9152 in shale lithofacies identification. Specifically, for the favorable lithofacies MF4, the Transformer model yielded the highest prediction performance, with $F1-\mathrm{score}$ of 0.89. This study demonstrates that deep learning models can effectively identify the shale lithofacies from conventional well logs, providing valuable technical insights for developing practical approaches to identify lithofacies.