Robust development of data-driven models for methane and hydrogen mixture solubility in brine

Within the domain of hydrogen storage initiatives inside subterranean structures, the accurate estimation of solubility of methane and hydrogen mixtures in brine becomes vital. In this paper, we aim to form robust data-driven intelligent algorithms founded on various machine learning methods of Support Vector Machine, Random Forest, AdaBoost, Decision Tree, K-nearest Neighbors, Multilayer Perceptron Artificial Neural Network and Convolutional Neural Network to model solubility of hydrogen/methane blend in brine under realistic conditions of underground hydrogen storage projects by utilizing an experimental dataset collected from the existing body of published research. An outlier detection method is utilized for checking out the data reliability for the model development. Also, sensitivity study is done to explore relative impacts of input parameters on solubility. The findings show that the Ensemble Learning model (R2 = 0.994842, MSE = 0.012959, AARE% = 3.842907) and AdaBoost model (R2 = 0.996241, MSE = 0.009444, AARE% = 3.607931) provide the highest accuracy in forecasting hydrogen/methane solubility in brine. These models attain the greatest determination coefficient (R2) and the lowest error metrics (MSE and AARE%), underscoring their remarkable capability to identify complex patterns and provide accurate predictions for estimating hydrogen/methane solubility in brine. The results indicate that Ensemble Learning and AdaBoost yield the highest accuracy algorithms in prediction capability as they tend to illustrate the lowest values of mean squared error and mean absolute relative error (%) and highest R-squared values. In addition, it was shown that solubility is mostly affected by hydrogen mole fraction in mixture and pressure. The developed models can be made use of for the estimate task of hydrogen/methane solubility in brine without needing experiments that are extremely laborious and require a lot of time.