Machine learning analyses of low salinity effect due to surface-charge alteration in carbonate porous media

Abstract

Accurate prediction of rock–brine zeta potential is critical for understanding wettability alteration and optimizing low-salinity water flooding (LSWF) in carbonate reservoirs. Although numerous experimental studies have investigated this phenomenon, the complex interactions among brine chemistry, rock mineralogy, fluid properties, and operational conditions remain inadequately quantified. This study systematically evaluated these relationships using nine machine learning models: Multiple Linear Regression (MLR), Regression Tree (RT), Random Forest (RF), Support Vector Machine (SVM), Gaussian Process Regression (GPR), Adaptive Neuro-Fuzzy Inference System (ANFIS), Multi-Layer Perceptron (MLP), Radial Basis Function (RBF) network, and Deep Neural Network (DNN). A dataset of approximately 900 experimental zeta potential measurements was used for model training and validation. Hyperparameters were optimized to maximize predictive accuracy and generalization, with performance assessed using root mean square error (RMSE) and efficiency factor (EF). The RBF network achieved the highest predictive accuracy (overall RMSE = 2.89 mV; EF = 0.96) and demonstrated robust generalization across varying training dataset proportions. DNN and RF also performed strongly, whereas linear (MLR) and fuzzy-based (ANFIS) models showed limited capacity to capture nonlinear electrochemical interactions. Sensitivity analysis indicated that brine ionic composition, particularly Na⁺, Ca²⁺, SO₄²⁻, Sr²⁺, and HCO₃⁻ concentrations, had the greatest influence on zeta potential. The proposed RBF-based predictive framework offers a reliable and efficient tool for estimating surface charge under diverse reservoir conditions, thereby improving mechanistic understanding of wettability alteration during LSWF.