The growing impacts of global warming demand urgent climate-change mitigation strategies, with carbon storage in saline aquifers emerging as a promising solution. These aquifers, for their high porosity and permeability, offer significant potential for CO
2 sequestration. Among the trapping mechanisms, solubility trapping—where
[...] Read more.
The growing impacts of global warming demand urgent climate-change mitigation strategies, with carbon storage in saline aquifers emerging as a promising solution. These aquifers, for their high porosity and permeability, offer significant potential for CO
2 sequestration. Among the trapping mechanisms, solubility trapping—where CO
2 dissolves into brine—stands out for its long-term effectiveness. However, CO
2 dissolution alters brine density, initiating density-driven convection that enhances CO
2 migration. Accurate modelling of these density changes is essential for optimising CO
2 storage strategies and improving long-term sequestration outcomes. This study presents a two-step explainable artificial intelligence (XAI) framework for predicting the density of CO
2-dissolved brine in geological formations. A dataset comprising 3393 samples from 14 different studies was utilised, capturing a wide range of brine compositions and salinities. Given the complexity of brine–CO
2 interactions, a two-step modelling approach was adopted. First, a random forest (RF) model predicted the brine volume (as the proxy for the density) without dissolved CO
2, and then, a second RF model predicted the impact of CO
2 dissolution on the brine’s volume. Feature importance analysis and SHapley Additive exPlanations (SHAP) values provided interpretability, revealing the dominant role of temperature and ion mass in the absence of CO
2 and the significant influence of dissolved CO
2 in more complex systems. The model showed excellent predictive performance, with
R2 values of 0.997 and 0.926 for brine-only and CO
2-dissolved solutions, respectively. Future studies are recommended to expand the dataset, explore more complex systems, and investigate alternative modelling techniques to further enhance the predictive capabilities.
Full article