A total organic carbon prediction algorithm for heterogeneous shale based on interpretable neural network: a case study of Qiongzhusi Formation shale in the Sichuan Basin

Total Organic Carbon (TOC) is a fundamental parameter for evaluating source rock quality, yet the strong heterogeneity of the Qiongzhusi Formation shale reservoir in the Sichuan Basin severely limits the applicability of conventional TOC prediction models. To address this challenge, this study proposes a novel TOC prediction algorithm (INN-BIC) that integrates an Interpretable Neural Network (INN) with the Bayesian Information Criterion (BIC). By employing feature decoupling and a dynamic polynomial degree selection mechanism, the method enhances both prediction accuracy and model interpretability in complex geological settings. The model successfully quantifies the contribution of well-log parameters such as uranium content, natural gamma ray, and deep/shallow resistivity to TOC, and accurately captures TOC variations in stratigraphic transition zones. Experimental results demonstrate that the INN-BIC model significantly outperforms traditional methods, improving the R 2 score by 79% and 25% compared to Backpropagation Neural Network (BPNN) and Support Vector Machine (SVM) models, respectively, and achieving a 65% enhancement over the original INN model. This verifies the model's effectiveness and reliability in strongly heterogeneous environments, supporting its practical application in shale gas sweet spot evaluation and efficient development.