Physics-Guided Machine Learning for Predicting Critical Salt Concentration and pH in Sandstone Reservoirs

Abstract Fines migration is a critical formation-damage mechanism in sandstone reservoirs, causing permeability reduction and injectivity loss during water flooding and injection operations. The onset of fines mobilization is governed by various key physicochemical thresholds, including critical salt concentration (CSC) and critical pH. These parameters control the balance between attractive and repulsive surface forces at the fine–brine–rock interface through electrical double-layer expansion. Although Derjaguin–Landau–Verwey–Overbeek (DLVO)-based modeling provides mechanistic insight, high-throughput application requires repeated calculations and extensive system-specific electrokinetic characterization. This study presents a physics-informed machine-learning framework that provides rapid and accurate predictions of both CSC and critical pH directly from reservoir, brine, and mineralogical descriptors. The main contribution is the development and systematic evaluation of machine-learning surrogate models capable of reproducing DLVO-derived fines-mobilization thresholds without repeated electrokinetic calculations. Nine ensemble algorithms were evaluated and optimized via Bayesian hyperparameter tuning using Optuna. Model performance was assessed using R2, RMSE, and MAE, with multicollinearity addressed through Pearson correlation analysis prior to training. The Extra Trees regressor demonstrated superior performance for both targets, achieving R2 = 0.9999, RMSE = 0.99, and MAE = 0.42 for CSC prediction, and R2 = 0.9999, RMSE = 0.12, and MAE = 0.03 for critical pH on independent, unseen test sets. Random Forest and CatBoost models were also able to achieve acceptable match between the true and predicted values of total potential; however, AdaBoost and XGBRF failed to capture the nonlinear behavior of electrostatic transitions. The optimized Extra Trees surrogate reproduced the DLVO-derived total interaction potential, Vt, with the highest accuracy among the tested models. CSC and critical pH are therefore obtained through a postprocessing threshold-extraction step applied to the predicted Vt profiles after screening, rather than as independent scalar outputs detached from the DLVO energy criterion. The proposed framework therefore combines mechanistic consistency with computational efficiency and enables rapid screening of fines-migration risks within the investigated ionic-strength, pH, zeta potential, and separation-distance ranges. It can support the selection of injection conditions that minimize permeability impairment and injectivity loss while reducing dependence on laboratory-intensive electrokinetic characterization.