Unbalanced graph isomorphism network for fracture identification by well logs

Fracture identification and prediction are of great significance for the production of tight oil and gas reservoirs. The high angles of fractures limit their traceability and reduce drilling intersection, leading to significant data imbalance and making fracture identification an imbalanced classification problem. Lithology and fluid properties can create similar features in fracture samples, often resulting in nonlinear relationships and a non-Euclidean structure. This complexity makes fracture identification a nonlinear process. To address this issue, the unbalanced graph isomorphism network (UGIN) algorithm is introduced. This approach leverages the GIN and incorporates a binary cross-entropy loss function specifically designed for unbalanced samples during fracture identification, aiming to adjust the model’s focus toward minority classes by assigning higher penalties to misclassified fracture samples, thereby improving detection accuracy in imbalanced datasets. The identification process is divided into three stages: First, the sample logging similarity information is integrated into the graph structure using the sequence edge method. Second, node-level information is embedded via the GIN algorithm, and nodes are clustered using K-means to derive the local graph’s embedding representation. Finally, nodes are classified using the model. To test the validation of the UGIN algorithm, a dataset of fractured carbonate reservoirs in A Oilfield, of the Zagros Mountain fold belt is used. The results demonstrate robust generalization on both training and test datasets through the use of cross-validation, achieving an AUC score of 0.938, higher than the baseline model. The classification accuracy on test data reaches 96.7%, with particularly strong performance in identifying fracture samples. To evaluate the impact of different graph construction methods on UGIN’s performance, we compare the K-means clustering method, hierarchical clustering method, the comprehensive connectivity method, the enhanced linkage strategy and the sequence edge method. Results indicate that the sequence edge method performs best, maximizing the retention of depth-related information in logging features and enhancing sample embedding.