Attention-Based Residual Convolutional Neural Network-Bidirectional Long Short-Term Memory for Volcanic Lithofacies Prediction

Summary Volcanic rock formation identification is a critical step in oil and gas reservoir exploration and development. However, its diverse composition, complex structure, and strong temporal dependency pose significant challenges to traditional methods. To enhance the accuracy and reliability of lithological identification, we focus this study on volcanic rocks in the Xiquan area of the Junggar Basin and propose a hybrid model based on an attention mechanism (AM), residual convolutional neural network (CNN) (ResCNN), and bidirectional long short-term memory (LSTM) (BiLSTM) network. This architecture combines ResCNN’s multiscale spatial feature extraction capability with BiLSTM’s bidirectional temporal information processing. The AM dynamically assigns weights to features of varying importance, while residual connections effectively mitigate degradation in deep networks. By comparing recognition results between CNN, LSTM, and BiLSTM models with a single algorithmic structure and the model proposed in this paper, the results indicate that ResCNN-BiLSTM-AM demonstrates higher and more balanced positive classification rates across all categories in the identification of individual rock types. For andesite recognition, ResCNN-BiLSTM-AM achieved particularly significant improvement, reaching a positive classification rate of 97%, compared with 93%, 84%, and 92% for CNN, LSTM, and BiLSTM models, respectively. Even in the face of a relatively small amount of tuff, the model’s performance reached its optimum. In terms of the overall recognition situation, the test accuracy of ResCNN-BiLSTM-AM also reached the maximum value of 96.09%. Furthermore, during testing on actual blind and continuous well sections, it demonstrated superior lithological identification results, further validating its practical engineering value in complex and uneven lithological identification tasks.