Abstract The growing deployment of carbon capture and storage (CCS) requires robust site screening and risk assessment tools to ensure safe and efficient CO2 injection. Traditional reservoir simulations, while highly accurate, are computationally expensive and impractical for real-time risk evaluation. Existing proxy models are unable to associate real time subsurface conditions with the extent of the fracture created, limiting their ability to inform real time operational decisions. This study aims to develop a machine learning (ML)-assisted proxy model that can predict critical subsurface integrity issues by linking subsurface conditions to fracture likelihood, enabling proactive evaluation of storage risks before injection and improving risk mitigation strategies through real-time monitoring data. A high-fidelity numerical simulator was used to generate a dataset of 150 simulation cases, covering diverse geological and operational conditions derived from actual CCS projects. Latin Hypercube Sampling (LHS) was employed to ensure comprehensive parameter coverage while maintaining data sparsity. The ML proxy model was trained using a neural network algorithm, testing different combinations of hyper-parameters until an optimum combination was found that ensured prediction accuracy while avoiding overfitting. The model was used to predict the occurrence of fractures. Model reliability was evaluated using key performance metrics such as Mean Absolute Error (MAE), R2 scores, Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Additionally, the developed model is tested with real field datasets to validate its applicability in practical storage scenarios. The ML-assisted proxy model demonstrated a high degree of accuracy compared to full-physics reservoir simulations, significantly reducing computational costs while maintaining predictive reliability. Results showed that the model successfully identified scenarios where fractures happen and could lead to CO2 leakage, quantifying fracture occurrence to inform corrective actions, such as reducing injection rates or increasing production from a pump well. Model also shows what well monitoring data (temperature and pressure) could look like in real-time, making it a valuable tool for offering dynamic insights into subsurface behavior. This study presents a scalable framework for site selection and operational decision-making. This work introduces a novel approach to evaluating subsurface CO2 storage integrity risks with high accuracy and real-time adaptability. Unlike existing methodologies, it accurately predicts fracture occurrence allowing for proactive risk assessment before injection and operational adjustments during storage. The opportunity to integrate real-time monitoring data to this model transforms conventional CCS risk management, making large-scale CO2 storage safer and more predictable. The broader impact extends beyond CCS, offering a novel framework applicable to underground hydrogen storage and geothermal reservoir integrity assessments.
Sekar et al. (Mon,) studied this question.