Exploring the power of machine learning to predict carbon dioxide trapping efficiency in saline aquifers for carbon geological storage project

Carbon geological sequestration (CGS) in saline aquifers is an effective carbon utilization approach to decrease the effect of greenhouse gases on the atmosphere. However, the GCS project's fundamental problem remains accurate prediction of carbon trapping efficiency in storage formations. As a result, four machine learning (ML) models were explored and compared in this study to improve the prediction of carbon storage performance in saline aquifers. The adaptive neuro-fuzzy inference system (ANFIS), extra tree (ET), random forest (RF), and radial basis function (RBF) models were used. These were considered for mapping the link between carbon trapping efficiency and influencing factors. A total of 1868 field simulation data points for carbon residual (RTI) and solubility trapping (STI) were extracted from published literature to develop the ML models. The four ML outcomes were assessed and compared using graphical analysis and statistical metrics. The results of the four models, ranked from strongest to weakest, were RF > ANFIS > RBF > ET. The RF scheme achieved the R-2 values of 0.995 and 0.965 and MAE values of 0.0074 and 0.0086 for RTI and STI, respectively. Therefore, an RF robust model is proposed to evaluate the carbon trapping effectiveness in saline aquifers. Simultaneously, the relevancy factor was applied to evaluate the sensitivity of each input parameter. The sensitivity analysis indicated that depth and post injection time were the most influential factors in residual and solubility trapping performance. The findings of this study could help us better understand the potential of ML for predicting carbon trapping efficiencies. The best RF model could serve as a template to quickly predict the performance of carbon storage.

Automated summary

This study compares different machine learning models to improve the prediction of carbon storage performance in saline aquifers. The random forest (RF) model performs the best and can effectively evaluate the carbon trapping effectiveness. Sensitivity analysis identifies depth and post injection time as the most influential factors on residual and solubility trapping performance.