Abstract: In response to the issues of low accuracy and significant lithological misidentification in traditional well-logging interpretation for gas hydrate exploration in permafrost regions, this paper proposes a fusion lithology identification method based on Bayesian Optimized Convolutional Neural Network–Long Short-Term Memory network (BO-CNN-LSTM). This method adaptively searches for the key hyperparameters of the CNN-LSTM model through Bayesian optimization, effectively integrating the strengths of CNN in extracting well-logging features with the capability of LSTM in modeling sequential information, thereby achieving high-precision lithology identification in gas hydrate-bearing formations in permafrost regions. Experimental results show that the overall identification accuracy of the BO-CNN-LSTM model reaches 96.25%, significantly outperforming traditional CNN-LSTM models and single deep learning models. This demonstrates the promising potential of the proposed method for lithology classification and reservoir evaluation in gas hydrate-bearing formations within permafrost regions.