Interpretable Deep Learning-Based Fracture Prediction Method: A Case from the Tarim Basin, China

In tight sandstone gas reservoirs, fractures play a critical role in enhancing reservoir permeability and connectivity. Seismic-based fracture prediction methods offer advantages in characterizing the three-dimensional spatial features between wells. Taking the A wellblock in the Tarim Basin as a case study, this research proposes a fracture prediction method based on interpretable multi-attribute deep learning. Using the number of fractures per meter from image logs as labeled data, initial features were extracted from optimized seismic attributes. These attributes were classified and subjected to principal component analysis to retain principal components containing key fracture genetic information, which served as supplementary feature data. Under the constraints of fracture controlling factors, three models—Backpropagation (BP), Deep Belief Network (DBN), and Long Short-Term Memory (LSTM)—were trained and compared using these feature datasets. The LSTM neural network was selected to construct the final prediction model, which demonstrated high accuracy and strong generalization capability across training, validation, and test datasets. The prediction results align with geological patterns: regions with highly developed fractures are associated with anticlinal axes and fault zones, while low-value zones occur in structurally stable areas such as synclinal cores. To interpret the model and identify the dominant controlling fracture types, the Local Interpretable Model-agnostic Explanations (LIME) method was applied to representative wells to evaluate the contributions of regional structural fractures, structural deformation-related fractures, and fault-derived fractures. The analysis indicates that fault-derived fractures were the dominant factor controlling fracture development in these wells. This method enables precise 3D spatial prediction of fracture density, improves result reliability through genetic attribute analysis, and identifies the dominant factors governing fracture development.