Аңдатпа
Abstract Hydrogen diffusion in the presence of cushion gas is an important factor controlling storage efficiency in Underground Hydrogen Storage (UHS). Experimental and molecular simulation approaches for evaluating hydrogen diffusion are time-consuming, underscoring the need for the development of fast, predictive tools. In this study, a machine learning framework was proposed to estimate hydrogen self-diffusion coefficients under the influence of CO2 as a cushion gas. A dataset of self-diffusion coefficients ranging from 10−8 to 10−6 m2/s was compiled from a reliable source of molecular dynamics simulations. The final dataset consisted of 600 data points, with 80% used for training and 20% for testing. The input parameters include pressure, temperature, gas type, and hydrogen mole fraction. In this study, the gas types are CO2 and H2 to facilitate predicting their corresponding diffusion coefficients. Three machine learning tools, decision tree (DT), random forest (RF), and gradient boosting regression (GBR), were trained using 5 k-fold cross-validation coupled with a grid search. The results showed that all models achieved high predictive accuracy. However, GBR outperformed DT and RF, with a testing root mean square error (RMSE) of 0.085 and a testing coefficient of determination (R2) of 0.991. This study proposed a new model to predict hydrogen diffusion in the presence of CO2 as a cushion gas across wide ranges of pressure (5-50 MPa), temperature (323-423 K), gas type (1 for H2 and 2 for CO2), and hydrogen mole fraction (0-1). Sensitivity analysis showed that hydrogen consistently exhibited higher diffusivity than CO2 across all pressures, temperatures, and compositions. This might be due to its lower molecular weight and lower density. Increasing the temperature increases the diffusivity of both gases, whereas increasing the pressure decreases it. The hydrogen diffusivity decreases with increasing CO2 content at all temperatures and pressures. These consistent trends confirm the integrity of the proposed model. The developed machine learning model provides a fast and accurate alternative to laboratory- and simulation-based studies. The proposed model offers valuable insights into predicting hydrogen diffusion in depleted gas reservoirs with CO2 as a cushion gas.