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

Pan Zhang¹Zilong Ren^1*Fengjiao Zhang²Wenhao Wang¹Lijie Cui¹Cheng Feng¹Weibiao Xie¹

• ¹China University of Petroleum (Beijing) Karamay Campus, Karamay, China
• ²Sinopec Matrix Corporation, Qingdao, China

Total Organic Carbon (TOC) is one of the key parameters in the evaluation of source rock properties. 11 Previous studies and XRD analysis have consistently confirmed the strong heterogeneity of the 12 Qiongzhusi Formation shale reservoir in the Sichuan Basin, which severely limits the applicability of 13 traditional TOC prediction models. To address this, this study proposes a TOC prediction algorithm 14 (INN-BIC) that integrates an Interpretable Neural Network (INN) with the Bayesian Information 15 Criterion (BIC). By employing feature decoupling and a dynamic polynomial degree selection 16 mechanism, the method successfully enhances both the prediction accuracy and model interpretability 17 for TOC in complex heterogeneous geological settings. The model reveals the contribution of well-log 18 parameters such as uranium content, natural gamma ray, and deep/shallow resistivity to TOC, and 19 accurately captures TOC variations in stratigraphic transition zones within the shale reservoir. 20 Experimental results demonstrate that the INN-BIC model significantly outperforms traditional 21 Backpropagation Neural Network (BPNN) and Support Vector Machine (SVM) methods, with 22 improvements in R² of 79% and 25%, respectively, and a 65% enhancement over the original INN 23 model. The study verifies the effectiveness and reliability of the proposed model in strongly 24 heterogeneous geological environments, supporting its application in shale gas sweet spot evaluation 25 and efficient development.