Underground Hydrogen Storage (UHS) in depleted gas reservoirs, due to their significant storage capacity and established infrastructure, is critical for integrating renewables and enhancing grid stability. However, traditional multiphase compositional simulations for UHS are computationally expensive. To mitigate this, a surrogate model using the XGBoost algorithm was developed to UHS performance, focusing on parameters such as cushion gas molecular weight, injection rate, fracture density, fracture presence, and formation pressure. Our findings reveal that fractures and their density significantly influence H2 withdrawal, while the reservoir’s dip angle minimally affects H2 recovery. Notably, fracture density emerges as the dominant predictive feature, contributing approximately 64% of the model’s predictive power, with formation pressure having minimal impact at around 27%. The achieved Mean Squared Error (MSE) of 10–6 showcases the XGBoost model’s close alignment with actual values, indicating high accuracy. Additionally, achieving an R2 score of 0.97 underscores a strong fit of the model to the actual observations. Lastly, results showed that XGBoost’s rapid training and prediction speed, handling 2400 data points approximately 47 times faster than traditional methods.