AI-driven predictive framework for CO2 sequestration and enhanced oil recovery: Insights from a depleted oil reservoir

Abstract

In light of the urgent need to mitigate the atmospheric consequences of carbon dioxide (CO2) emissions while safeguarding the planet and its inhabitants, carbon capture, utilization, and storage (CCUS) emerges as a viable solution. The co-optimization of CO2 sequestration and CO2-enhanced oil recovery (CO2-EOR) has become a decisive strategy to simultaneously enhance oil recovery and long-term CO2 trapping in depleted oil reservoirs. This study presents efficient techniques for predicting oil recovery and the CO2 solubility trapping across over 800 experimental designs using Computer Modelling Optimization and Sensitivity Tool- Artificial Intelligence (CMOST-AI). To improve prediction accuracy and address the complexity, seasonality, outliers, noise, and scatter in the experimental dataset, an efficient preprocessing technique called Improving Distribution Analysis (IDA) and Quantile Transformer (QR) is applied to tackle these challenges. To evaluate their performance, these preprocessing techniques are combined with two advanced deep learning models: the One-Dimensional Convolutional Neural Network (CNN1DF) and the Multilayer Neural Network (MLNNF), both of which have bilateral predictions. It reduces reliance on time-consuming simulations and supports real-time decision-making, offering theoretical insight and field-level applicability. This combination provides robust frameworks that avoid overfitting and consistently outperform, achieving high R2 values of 81.20 % and 81.08 % for the full feature set and 82.36 % and 84.21 % for the reduced feature set in predicting both the oil recovery factor and CO2 solubility trapping. Furthermore, the results surpass the predictions of the CMOST-AI software models by more than 50 % while avoiding the overfitting problem. The findings illustrate that the proposed frameworks significantly improve predictive performance and effectively address overfitting issues, outperforming many of the compared models.