Comparative Machine Learning Frameworks for Forecasting CO2/CH4 Competitive Adsorption Ratios in Shale

Abstract

Accurate modeling of CO2/CH4 competitive adsorption behavior is a critical aspect of enhanced gas recovery associated with CO2 sequestration in organic-rich shales (CO2-ESGR). It not only improves the ultimate recovery of shale gas reservoirs that satisfies the increasing energy demand but also provides permanent geologic storage of atmospheric CO2 that contributes to the net-zero energy future. Determining a CO2/CH4 adsorption ratio is essential for the performance prediction of shale gas reservoirs and the evaluation of CO2 storage potential. However, experimental adsorption measurements are expensive and time-consuming that may not always be available for shale reservoirs of interest or at the investigated geologic conditions, and as a result, a sorption ratio cannot be assessed appropriately. Traditional models such as a Langmuir model are highly dependent on extensive experiments and cannot be widely applied. Therefore, a unified adsorption model must be developed to predict the CO2/CH4 competitive adsorption ratios, which is essential for CO2 sequestration and exploitation of natural gas from shale reservoirs. In recent years, the development of machine learning algorithms has significantly improved the accuracy and computational speed of prediction. In this work, we conducted a comparative machine learning algorithm study to effectively forecast the maximum CO2 adsorption capacity and CO2/CH4 competitive adsorption ratios. Four sensitive input parameters (i.e., temperature, total organic carbon, moisture content, and maximum adsorption capacity of CH4) were selected, along with their 50 data points collected from the existing literature. The artificial neural network (ANN), XGBoost, and Random Forest (RF) algorithms were investigated. By comparing the mean absolute errors (MAE) and coefficients of determination (R2), it was found that the ANN models can successfully forecast the required outputs within a 10% accuracy level. Furthermore, the descriptive statistics demonstrated that the CO2/CH4 competitive adsorption ratios were generally from 1.7 to 5.6. The proposed machine learning algorithm framework will provide insights beyond the isothermal conditions of classical adsorption models and the solid support to CO2-ESGR processes into which competitive adsorption can be a driven mechanism.