Machine Learning-Based Prediction of the Migration Range of Dissolved CO2 in Deep Saline Aquifers: SHAP Interpretation and Engineering Insights

Abstract

Prediction of the dissolved CO₂ migration range is a key step in evaluating the efficiency and safety of geological storage in deep saline aquifers. To improve model generalization and elucidate the underlying roles of input features, this study proposes a predictive framework applicable to diverse reservoir types by integrating TOUGH-based numerical simulations with a fully connected neural network (FCNN). Furthermore, SHapley Additive exPlanations (SHAP) are employed to quantitatively assess the importance of six input features. Results show that the FCNN model exhibits strong predictive performance on the test set, with an average R² exceeding 0.9 and a normalized root-mean-square error (NRMSE) of approximately 0.04. Permeability and porosity are identified as the dominant factors controlling the maximum migration distance of dissolved CO₂, with feature contributions of 0.77 and 0.34, respectively. Notably, high porosity exerts a suppressive effect: under specific geological and operational conditions, as porosity increases from 0.0129 to 0.621, the maximum migration distance decreases from 7229 to 818 m, and the total storage capacity declines from 24.46 to approximately 20.56 Mt. These findings provide technical support for site screening and injection design in CO₂ geological storage.