ResGAT: A Residual Graph Attention Network for Lithology Identification

Abstract

Lithology identification is crucial for oil and gas exploration and reservoir evaluation, involving the analysis of physical and chemical characteristics of geological samples through well-logging data. This process requires understanding the complex nonlinear relationships between logging parameters and lithology. Recently, graph neural networks have gained prominence for their ability to uncover hidden relationships among samples, enhancing lithology identification. However, the imbalanced distribution of logging data often leads to incorrect interclass connections in logging graphs, which can skew feature aggregation and reduce prediction accuracy. To address this issue, this letter introduces the residual graph attention network (ResGAT), which integrates the residual information of well-logging data into the graph network based on graph relationships, adds residual connections to mitigate the impact of interclass edges, and enhances the weight of original information. To authentically assess the model’s practical effectiveness, we, respectively, conducted cross-well predictions in completely isolated well sets in oil fields in Daqing, China, and Kansas, USA. Compared to conventional GAT and GCN models, our proposed method achieves higher identification accuracy and significantly improves prediction accuracy for minority classes.