Deep learning methods for the simulation and optimization of shallow geothermal energy systems

Smajil Halilovic 

Shallow geothermal energy (SGE) systems are crucial for decarbonizing the heating and cooling sector. Their planning, design and operation, however, rely on the simulation of heat transport in the subsurface, a task that is computationally demanding and particularly prohibitive for multi-query applications such as sensitivity analysis and optimization. Deep learning (DL) has recently emerged as a powerful means of accelerating or replacing these numerical models. This paper reviews the state of the art of DL methods for the simulation and optimization of SGE systems, with a focus on subsurface thermo-hydraulic processes rather than on above-ground system components. The existing literature is categorized by system type: closed-loop and open-loop systems; and by the applied DL methodology, the predicted quantities and the underlying numerical model. The review reveals that the field is still dominated by purely data-driven, system-level models, whereas physics-informed architectures and operator-learning approaches that explicitly resolve the subsurface remain scarce. Based on these findings, we discuss the central challenges of the field: multi-scale dynamics, well singularities, data scarcity; and outline future research directions, including physics-informed and hybrid models, neural operators for parametric problems, and generative models for uncertainty quantification, ultimately pointing toward real-time digital twins of SGE systems.