Robust machine learning models of carbon dioxide trapping indexes at geological storage sites

Abstract

Geological carbon dioxide (CO2) storage (GCS) in saline aquifers has been recognized as a promising way to slow down global CO2 emissions. The residual and solubility trapping efficiency of CO2 in saline aquifers play a crucial role in monitoring CO2sequestration performance. Due to this fact, the goal of this paper is to determine the effectiveness of three robust machine learning (ML) models — a general regression neural network (GRNN) model and two multilayer perceptron (MLP) models respectively optimized with the Levenberg-Marquardt algorithm (LMA) and Bayesian Regularization (BR) — in predicting the residual trapping index (RTI) and solubility trapping index (STI) of CO2 in saline aquifers. A comprehensive and wide-ranging dataset was compiled from the literature, including over 1,910 simulation samples from numerous CO2 field models. The predicted results revealed that all the proposed ML techniques have an excellent agreement with simulation data. In addition, the error analyses and a comparison of statistical indicators indicated that the GRNN model was more accurate than the MLP-LMA and MLP-BR models as well as two ML models developed in previous studies. The GRNN model exhibited overall coefficient of determination (R2) values of 0.9995 and 0.9998 and average absolute relative error (ARE) percentages of 0.7413% and 0.2950% for RTI and STI, respectively.Furthermore, a trend analysis confirmed the robustness of the GRNN model, as predicted and simulated RTI and STI exhibited strong overlap under four different sets of input parameters. Moreover, the Williams plot reveals that the validity of GRNN model was affirmed, and only a small suspected data was detected from the collected database. Therefore, the GRNN model proposed in this study could serve as a template for evaluating the feasibility of future GCS projects in saline aquifers. Lastly, the findings of this study can help better understanding the promising of machine learning techniques for predicting CO2 trapping efficiency in geological storage sites.