Abstract To address the challenges of reliance on manual interpretation, high non-uniqueness, and limited accuracy in the geological sub-layer division of the C812 layer in the Yanchang Formation, L block, Ordos Basin, this study proposes an intelligent division method based on a CNN-BiLSTM model. This method integrates the strengths of convolutional neural networks (CNNs) and bidirectional long short-term memory (BiLSTM) networks. The CNN component extracts local features from logging curves – such as abrupt changes in logging responses at formation boundaries – while the BiLSTM captures temporal dependencies within the data, enabling accurate geological sub-layer division. Using logging data – including natural gamma, spontaneous potential, acoustic time difference, and resistivity – from 14 cored wells in the study area as training samples, the model achieved an overall recognition accuracy of 89.58 % on the validation set after data preprocessing and training. The recognition accuracies for sub-layers No.1 to No.4 were 90.95 %, 87.75 %, 90.23 %, and 89.24 %, respectively. When the model applied to new wells without training process, the layer interface can be effectively distinguished. The average errors of the three interfaces corresponding to the four groups of geological layers are 0.3125 m, 0.5625 m and 0.1875 m, respectively. Compared with LSTM, BP neural networks, and the XGBoost algorithm, the CNN-BiLSTM model demonstrated superior performance in both division accuracy and continuity, effectively addressing the challenges of fragmented transition zone recognition and dependence on manual interpretation inherent in traditional methods. The research results provide reliable data for the fine-scale characterization of single sand bodies in the C812 layer and present a novel technical approach for geological sub-layer division in continental clastic reservoirs.
Zhao et al. (Thu,) studied this question.