Abstract
Permeability is a crucial analytical variable in petrophysical parameters of reservoir rocks, which is highly related to geo-energy exploration and evaluation. Conventional physics-based models and data-driven permeability estimation methods using pore-structure parameters and image parameters as input both highly depend on core-experimental parameters and have potential limitations in uncored area. Using logging data as input, seven machine learning methods in total, including Linear Regression, Back Propagation Neural Network Regression, K-Neighbors Regression, Random Forest Regression, Support Vector Machine Regression, Gradient Boosting Decision Tree Regression, and Extreme Gradient Boosting (XGBoost) Decision Tree Regression were trained to predict the permeability of low-permeable sandstones of Zhuhai Formation in Wenchang A Sag, Pearl River Mouth Basin. In the seven well-trained models, XGBoost showed the optimal performance in permeability prediction with the coefficient of determination (R~2) of 0.91088 and mean squared error (MSE) of 0.135. Moreover, the continuous permeability profile of one well was predicted successfully using the optimal model of XGBoost. Furthermore, in order to interpret the black box of machine learning of XGBoost model, SHAP (Shapley additive explanations) was introduced to interpret the well-trained model of XGBoost. The results indicated that acoustic transit time and corrected compensated neutron log in the logging series contributed the most to the permeability prediction in the low-permeability reservoir rocks. And each output result could be expressed as the sum of contributions of five input logging parameters in the XGBoost model. This work gains insight into the rapid prediction of permeability of underground aquifers and hydrocarbon-bearing layers using easily available logging data without coring and testing, and provides pattern visualization experience for permeability prediction in low-permeable sandstones.
NSTL