The growing demand for eco-friendly power solutions, driven by rising greenhouse gas emissions, highlights hydrogen (H2) as a promising sustainable energy source. H2 reduces emissions and decreases dependence on imported fossil fuels, serving as a versatile energy carrier for the energy transition era. Effective storage solutions are crucial for establishing H2 as a viable low-carbon option, utilizing depleted hydrocarbon reservoirs for Underground Hydrogen Storage (UHS). Accurate well placement within these reservoirs is essential for optimizing the H2 recovery, and traditional reservoir simulations in this context are costly and complex due to geological uncertainties. Surrogate models such as LightGBM and AdaBoost offer a promising alternative by improving accuracy in identifying optimal well locations through efficient handling of non-linear relationships and complex datasets. This study compares these machine learning models with traditional simulations to optimize well placement in a Middle Eastern gas field, aiming to improve the H2 withdraw efficiency and reduce computational costs. Using 10,000 realizations from a commercial compositional simulator, varying parameters such as well locations, perforation layers, and hydrogen injection/production rates, the study found that high-rate hydrogen injection into shorter reservoir intervals resulted in lower recoveries, whereas lower rates into longer intervals achieved higher recoveries. LightGBM achieves strong predictive accuracy (R2 = 0.98) and computes H2 recovery 1.7 times faster than AdaBoost, with a significantly lower RMSE (approximately 53% less) on unseen data. These results highlight the efficiency of machine learning in optimizing H2 storage strategies, surpassing traditional simulations that required 350 s per realization to predict H2 recovery.