Geothermal energy and underground thermal storage (UTES) are vital to the low-carbon energy transition, yet their optimization is bottlenecked by multi-scale heterogeneity, coupled thermal–hydraulic–mechanical–chemical (THMC) processes, and the high computational cost of full-physics simulations. This review systematically evaluates machine learning (ML) as a foundational paradigm for overcoming these computational and scale-bridging challenges. We categorize current advances into three key functional roles. First, data-driven upscaling directly maps pore-scale features to macro-scale effective properties, replacing traditional empirical homogenization. Second, deep surrogate models mimic high-fidelity THMC simulations at a fraction of the computational cost, enabling real-time prediction and uncertainty quantification. Third, physics-informed digital twins integrate real-time sensor streams with cloud architectures for dynamic reservoir management. Furthermore, we address the generalization limits of purely data-driven approaches, highlighting physics-informed machine learning (PIML) and hybrid architectures that embed conservation laws as strict constraints. Finally, we outline future pathways toward multimodal data fusion and edge-cloud deployment, marking a shift from static offline modeling to dynamic, physics-safeguarded real-time reservoir optimization.