53 / 2025-11-20 15:47:47

Improved BP Neural-Network-Based Prediction of Shale TOC: Incorporating Morris Sensitivity Analysis

Morris sensitivity analysis;Back Propagation Neural Network;Dropout and L2 regularization; Total organic carbon

Accurate prediction of total organic carbon (TOC) is essential for shale gas reservoir evaluation, yet remains challenging due to strong geological heterogeneity and limited generalization capability of conventional logging-based methods. To address these issues, this study proposes a sensitivity-guided intelligent prediction framework by integrating Morris global sensitivity analysis with an improved Backpropagation Neural Network (BPNN). The Morris algorithm is first applied to identify the seven most TOC-sensitive logging parameters, enabling dimensionality reduction and reducing multicollinearity. Subsequently, a regularized BPNN incorporating both Dropout and L2 weight penalty is developed to enhance robustness and mitigate overfitting.Application to the Qianfoya Formation shale in the Sichuan Basin demonstrates that the proposed model achieves superior predictive performance, with an R² of 0.9316, outperforming traditional Support Vector Regression (SVR) and Particle Swarm Optimization–Least Squares SVM (PSO-LSSVM), which yield R² values of 0.8247 and 0.8325, respectively. The improved network also exhibits significantly reduced MAE and RMSE, and effectively avoids negative TOC predictions that commonly arise in conventional models. Sensitivity interpretation reveals that AC and DEN provide the strongest responses to TOC variations due to the influence of organic matter enrichment on rock elasticity and bulk density.Overall, the integrated workflow enhances both the interpretability and prediction accuracy of logging-based TOC estimation and provides a practical and geologically consistent approach for unconventional shale gas resource characterization and sweet-spot evaluation.