Predicting Fines Migration Threshold in Sandstone Reservoir: A Machine Learning Framework for Predicting Critical Salt Concentration

Abstract

Fines migration is a well-documented phenomenon in sandstones, which impairs permeability and causes productivity and injectivity issues. Critical salt concentration (CSC) defines the lowest brine salinity below which fines detach from sand surface due to increased repulsive forces between fines and sand. While prior research relied mainly on experiments and analytical models, this work presents a novel machine-learning framework that integrates governing physicochemical parameters to rapidly and accurately predict fines migration under diverse reservoir conditions. The DLVO model data for CSC prediction in Berea sandstone, previously validated experimentally, were utilized from our earlier research to train selected ML models. Several new ML algorithms, including AdaBoost, XGBRF, Gradient Boosting, Random Forest, and Extra Trees, have been implemented, and metrics such as Coefficient of Determination (R2), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) were quantified. Following this, hyperparameter optimization was performed for the top-performing ML model, and explainable artificial intelligence (XAI) clarified how the leading model captured the influence of each input parameter within the overall fines-migration prediction framework used in this study. Among all the trained models, the Extra Trees regressor showed the most robust predictive performance, achieving R2 = 1.0000 for training and 0.99996 for testing, with very low RMSE (0.897) and MAE (0.436) values. These minimal errors reflect the model's excellent ability to reproduce DLVO-derived CSC values with negligible bias and variance, confirming its strong generalization to unseen data. The superior performance of Extra Trees comes from its ensemble architecture, which averages multiple decorrelated decision trees, effectively capturing nonlinear interactions between salinity, mineral surface charge, and electrostatic forces that influence fines detachment. Random Forest also achieved high accuracy, with R2 = 0.9999, RMSE = 1.067, and MAE = 0.418 in training, and R2 = 0.9998, RMSE = 1.999, and MAE = 1.034 in testing, demonstrating reliable performance but with slightly higher errors due to increased sensitivity to tree correlation and sample variance. In contrast, AdaBoost was the least effective, shown by much lower R2 values (~0.97) and significantly higher RMSE (~25) and MAE (~19), indicating reduced ability to capture nonlinear electrostatic trends within the DLVO dataset. Explainable AI (XAI) analyses using feature importance highlighted how Extra Trees prioritized key physicochemical descriptors and explained the influence of each parameter on CSC. The current work presents a fast, generalizable ML surrogate that predicts the critical salt concentration (CSC) for fines migration initiation directly from readily available descriptors such as salinity, temperature, mineralogy/fine size metrics, and electrokinetic features. Trained on physics-informed datasets, the best ML model preserves mechanistic consistency while providing rapid, uncertainty-aware CSC thresholds suitable for field-scale screening across varied brine chemistries, mineralogies, and operating conditions.