Lithology identification of shale gas reservoirs using multi-source intelligent learning algorithms

Lithology classification of shale gas reservoirs is pivotal for oil and gas exploration and development, as it directly dictates the reliability of reservoir evaluation, fracturing design, and productivity prediction. To address the prevalent challenges of inadequate recognition accuracy and weak generalization in traditional machine learning-based lithology identification, this study proposes a novel multi-scale feature fusion model (MSFR-Net) with integrated core innovations. The model is applied to the shale gas reservoir in the F well area of the eastern Ordos Basin, using approximately 3,600 sampling points extracted from raw well logs of seven wells. Eight lithologically sensitive log response parameters were selected via the Pearson correlation coefficient method, including GR, AC, SP, CAL, DEN, PE, RLLD, and RLLS. MSFR-Net integrates an innovative feature extraction module, a random convolution module with dynamic kernel optimization, and a multi-classifier collaborative decision module. The new method synergistically overcomes the limitations of traditional models in feature mining and prediction robustness to enable accurate intelligent lithology recognition of the Triassic Yanchang Formation continental shale reservoir. Experimental results based on real well-log data confirm that MSFR-Net outperforms five widely used models (including BPNN, SVM, and GRU) in accuracy, precision, recall, and F1-score. Specifically, it achieves 93.25% prediction precision for major lithologies in Well X-22, representing 11.33% and 7.68% improvements over SVM and BiLSTM, respectively. The integrated innovative design of MSFR-Net significantly enhances key geological feature extraction and model generalization, providing a novel technical pathway for shale gas reservoir lithology identification.