Predicting total organic carbon from well logs based on deep spatial-sequential graph convolutional network

The total organic carbon (TOC) is a key geologic parameter for unconventional reservoirs. Conventional empirical [Formula: see text] methods cannot handle the nonlinear relationships between the characteristics of TOC and its well-log responses. Increased data availability has the potential to speed up deep learning applications, which can reasonably propagate the integrated information from well logs to indirectly observable geologic properties, such as TOC. Although the existing convolutional neural network (CNN) has found superior performance to [Formula: see text] for predicting TOC, CNNs feature-learning capability is still constrained by the fact that it can only extract log-specific sequential features of the input logs. However, the cross-log topological association features are potentially essential for the nonlinear mapping between well logs and TOC. Thus, we introduce a novel deep spatial-sequential graph convolutional network (SSGCN) for predicting the TOC by jointly leveraging the cross-log topological association features and log-specific sequential features. Through further use of the previously unaccounted topological interactions, our SSGCN dramatically outperforms the sequence-based CNN. In the southeast Sichuan Basin, SSGCN exhibits beneficial mapping not demonstrated previously: its models achieve a better cross-validation performance within the same gas field wells and a greater generalizability in another gas field well. Our SSGCN method can predict TOC of shale gas field well with the best [Formula: see text] being 0.87 within 1 s on the CPU of a desktop computer, which increases the efficiency of obtaining the TOC parameter. From this study, we recommend graph and sequential convolutions for designing deep learning architectures in the well-log analysis.