Shale oil and shale gas are important unconventional resources. As the total organic carbon (TOC) of shales is related to the generation potentiality of hydrocarbons, a model of TOC prediction can improve development efficiency and reduce cost. However, the intelligent prediction of TOC is not widely studied. This study works on hydrocarbon-rich shales from four basins in China and Canada. ΔLogR, support vector regression (SVR), single-layer artificial neural network (ANN), and fully connected deep network (FCDN) were built to predict the TOC of the studied shales using resistivity, sonic, density, and gamma ray logs. The predicted and geochemical-measured TOC were compared. The correlation of determination (R²) and normalized-root-mean-square-error (NRMSE) were used to evaluate the models. The results indicate that ΔLogR, SVR, and ANN have low R² and high NRMSE values and are unsatisfactory to predict TOC. These models greatly underestimate the TOC, which provide evident deviations of predicted TOC from their true values.

Conversely, FCDN provides accurate TOC predictions. The optimum results are obtained from an eight-layer network that consists of one input layer, six hidden layers, and one output layer. Values of R² and NRMSE are 0.89 and 0.044, suggesting that the TOC predictions using the FCDN are close to their true values. The FCDN greatly outperforms other models by extracting the complicate relationships between well logs and TOC values. The results of this study may suggest the great potentiality of deep learning techniques in the evaluation of unconventional resources.