Azam Khan

Predicting post-acidizing porosity and permeability in sandstone reservoirs remains a fundamental challenge due to complex, nonlinear acid–rock interactions. This study presents a physics-augmented machine learning and deep learning (ML/DL) framework that integrates 24 laboratory core-flooding measurements from the Lower Indus Basin, Pakistan, with 500 synthetic training samples generated by the Panga–Lemerle–Balakotaiah (PLB) two-scale continuum simulator. Eight ML/DL architectures are benchmarked on the augmented test set (n = 75): Support Vector Regression (SVR) achieves the highest porosity R $$^{2}$$ = 0.927 and Random Forest the highest permeability R $$^{2}$$ = 0.907. A systematic domain-gap analysis reveals that naive median imputation of the unknown HCl concentration causes catastrophic external validation failure (porosity R $$^{2}$$ =−2.34). Five targeted remediation strategies are evaluated. Strategy S4—Optuna-tuned LightGBM with Leave-One-Out cross-validated joint augmented+experimental training—raises external R $$^{2}$$ to 0.918 (porosity) and 0.841 (permeability), representing improvements of +3.25 and +2.18 R $$^{2}$$ units over baseline. SHAP interpretability analysis identifies HCl concentration, initial permeability, and the Damköhler number as collectively governing >75% of prediction variance. The framework provides a physically interpretable, calibration-ready forecasting tool for pre-treatment petrophysical prediction aligned with sustainable low-concentration acidizing design in accordance with UN SDG 7 and SDG 12.