A Hybrid Machine Learning Method for Dynamic Monitoring of CO2 Sequestration Using Pulsed Neutron Logging

This study proposes a hybrid machine learning model based on full-spectrum pulsed neutron logging data to address the monitoring challenges of Carbon Capture, Utilization, and Storage (CCUS) under complex geological conditions. Traditional interpretation models for sequestered CO2 saturation (e.g., macroscopic capture cross-section model, characteristic peak count model, and ratio model) heavily rely on prior parameters such as porosity, formation water salinity, and lithology. Acquiring these parameters in real time during practical engineering is often costly and difficult. To reduce the rigid dependence of accurate CO2 saturation monitoring on complex prior parameters like porosity and salinity under heterogeneous geological settings, this research focuses on the Pearl River Mouth Basin, a core carbon sequestration target area in the Guangdong-Hong Kong-Macao Greater Bay Area, based on the evaluation results of offshore carbon sequestration macro-regions in China. Taking the primary reservoirs of the Enping and Wenchang Formations as typical geological prototypes, a high-fidelity, full-spectrum neutron–gamma response database was constructed using Monte Carlo simulations. Two machine learning strategies are proposed: a direct regression model (NMF+SVR) and a joint model (NMF+SVC/KMeans+SVR). Based on Monte Carlo simulated data, experimental results demonstrate that, compared with traditional petrophysical baseline models and simple machine learning models, the proposed joint learning method effectively reduces the dependence of CO2 saturation monitoring on lithology and porosity. Furthermore, it is proven that even with a single-detector tool configuration, the method exhibits high prediction accuracy under complex lithological conditions. Notably, the two-step joint model achieves a Root Mean Square Error (RMSE) as low as 4.200%, significantly outperforming traditional physics-based models and single machine learning models such as MLP and RF. This study provides a physically interpretable and accurate technical reference for applying machine learning to pulsed neutron-logging-based CO2 geological sequestration monitoring.