Abstract
Underground hydrogen storage (UHS) is a critical component of future sustainable energy infrastructure, offering reliable solutions for energy storage and supply security. The relative permeability of hydrogen (H(2)) significantly impacts UHS performance by governing gas mobility during injection and withdrawal cycles. Traditional empirical models often fail to capture the complex interactions in hydrogen-water systems, necessitating advanced predictive approaches. In this study, machine learning (ML) techniques-including Polynomial Regression, Multi-Layer Perceptron (MLP), Gaussian Process Regression (GPR), Kernel Ridge Regression (KRR), Random Forest Regression (RFR), and Gradient Boosting Regression (GBR)-were employed to predict H(2) relative permeability under diverse experimental conditions. A dataset of 130 data points, encompassing variables such as gas saturation, porosity, salinity, and differential pressure, was used for model training and evaluation. Among the tested models, GPR demonstrated superior performance, achieving an R(2) of 0.9356, RMSE of 0.0280, and MAE of 0.0178. These results underscore the potential of ML to provide accurate and efficient predictions, outperforming traditional empirical methods. The study highlights the importance of data-driven approaches for optimizing UHS systems and contributes to advancing sustainable energy solutions. Future research should focus on expanding datasets and incorporating key physical processes, such as hysteresis and gas mixing effects, to further enhance model accuracy and applicability.