Enterprise AI Analysis

Data driven prediction of reservoir rock wettability in shale using deep learning and gene expression programming for CCUS

Accurately predicting reservoir rock wettability is important for CO2 retention, migration, and carbon storage. It directly affects the long-term stability of CO2 in shale reservoirs for carbon sequestration and enhanced oil and gas recovery. However, traditional laboratory methods for assessing wettability are slow and complex due to the heterogeneity of shale. That's why to address these challenges, this study uses Transformer, AdaBoost, and Stacking ensemble model to predict the contact angle (CA) using key parameters such as NaCl (M), Quartz (wt%), Carbonate (wt%), Clay (wt%), TOC (%), Pressure (MPa), Temperature (K), Permeability (md), and Porosity (%). The Transformer model showed the highest predictive accuracy (R2 =0.9959) with the lowest RMSE (1.7866) and MAE (1.1760). Moreover, the Taylor diagram, residual plot analysis, and computational time also confirmed its superiority over the other two models. Additionally, a Gene Expression Programming (GEP) model was developed to provide an explicit symbolic correlation for predicting CA, thereby avoiding the need for repeated model training. The resulting GEP equation demonstrated strong predictive quality (R2=0.9949 and 0.9862; RMSE =2.26 and 3.80; MAE = 1.70 and 2.31 for training and validation), validating both reproducibility and interpretability. The Transformer technique, along with the GEP approach, is a novel contribution to shale wettability modeling. Furthermore, a sensitivity analysis identified temperature as the most influential factor, while Quartz, carbonate, clay, TOC, and pressure also affected CA, confirmed by SHAP and trend analyses.