Predicting particle-induced permeability impairment during water injection using ensemble machine learning models

• Machine learning predicts particle-induced permeability impairment during water injection. • Ensemble boosting models (LightGBM, XGBoost) achieve R² > 0.95 with low error. • SHAP analysis reveals PVI and particle concentration as dominant predictors. • External validation confirms strong robustness under varying injection conditions. • Models enable data-driven injectivity management and digital twin integration. Water injection is a widely applied secondary recovery method, yet the presence of solid particles in injected water often causes pore throat plugging and deposition, leading to permeability impairment near the wellbore. Accurate prediction of this impairment is critical for optimizing injection strategies and reservoir management. In this study, machine learning was employed to predict permeability reduction using six input parameters: pore volume injection (PVI), initial permeability, porosity, particle size, particle concentration, and injection rate. A dataset of 1003 samples was compiled and analyzed using five supervised learning algorithms, Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LightGBM), and Categorical Boosting (CatBoost), optimized through Optuna and evaluated via 10-fold cross-validation. All ensemble models achieved strong performance, with correlation coefficients (R 2 ) above 0.95 and mean squared error (MSE) ranging below 0.039 across models and datasets. LightGBM provided the most accurate predictions (R² = 0.981, MSE = 0.018), closely followed by XGBoost and Gradient Boosting, while residual and learning curve analyses confirmed strong generalization and the absence of overfitting. External validation using two independent experimental datasets further demonstrated model robustness with MSE < 0.039, where partial pore reopening introduced additional complexity. SHAP-based feature importance highlighted PVI and particle concentration as the dominant predictors, followed by injection rate and initial permeability, with particle size and porosity contributing secondarily. Overall, the study demonstrates that boosting algorithms, particularly LightGBM and XGBoost, provide a fast, cost-effective, and physically consistent approach for predicting permeability impairment, with strong potential for integration into reservoir management and digital twin frameworks.