Keyvan Eghtedari, Parviz Darvishi, Asghar Lashanizadegan, Shahin Kord, Siavash Partonia, Sadraddin Sobhani
This study develops an AI and machine learning (ML) framework for predicting and mitigating asphaltene-induced formation damage during CO₂ injection for enhanced oil recovery (EOR) and carbon sequestration in low-permeability dolomitic reservoirs. Systematic core flooding experiments on 44 plugs under reservoir conditions (80–220°F, 2200–6500 psig, CO₂: 20–100 vol.%) quantified recovery factor (16.8–67.9%) and damage intensity (3.4–48.5%). Analysis of 18 input parameters, including mineralogy (dolomite: 60–80 wt.%, anhydrite: 2–14 wt.%), asphaltene content (3.11–13.56 wt.%), and operational variables, revealed that CO₂ concentration and pressure increased both recovery and damage, whereas temperature enhanced recovery while suppressing deposition. The experimentally constrained dataset was expanded to 1000 samples using physics-guided bootstrapping. Six AI/ML models were trained and evaluated. The 1D Convolutional Neural Network (1D-CNN) achieved the highest predictive accuracy, with an R² of 0.8844 (RMSE=3.073) for recovery and an R² of 0.9682 (RMSE=3.362) for damage intensity, and a top composite score of 0.9605. Its validation loss curve converged smoothly within 60–80 epochs, showing minimal generalization gap, while error distributions were the narrowest (±10%) and most symmetric. Parity plots confirmed near-perfect alignment along the 45° line. In contrast, the ResNet-Tabular model displayed oscillating, non-convergent loss and poor performance (R² < 0.55). The superior 1D-CNN captures localized feature interactions essential for modeling complex precipitation dynamics under gas injection. This validated AI model provides a reliable digital tool for optimizing CO₂-EOR and carbon sequestration projects by enabling proactive formation damage management, thereby supporting sustainable hydrocarbon recovery and secure carbon storage in carbonate formations.