The increasing adoption of groundwater heat pumps due to climate change necessitates optimized citywide placement to minimize negative interactions between heat plumes. So far, these scenarios are simulated on large computational clusters [1]. Our work aims to replace these simulations with flexible, faster machine learning models [2]. We propose a method for rapid prediction of subsurface temperature fields in domains of arbitrary size, accommodating any number of heat pumps. Our approach decomposes the problem into local and global dependencies. Convolutional Neural Networks (CNNs) capture complex local interactions, while simpler global dependencies are handled numerically. To enable scalability and efficient training, we introduce a Local-Global CNN-based (LG-CNN) framework that predicts temperature fields using simulated datasets based on real-world parameters. We apply this method to a simplified two-dimensional scenario with constant operational pump parameters. The model is trained on a single large simulation, partitioned into overlapping sections to accelerate the training. Current experiments include adapting the model to subsurface parameter maps extracted from the Munich region and incorporating realistic variations in operational parameters. In the future, we will extend the model to handle 3D seasonal data. REFERENCES [1] K. Zosseder and GEO.KW-Team, “GEO.KW – Abschlusskonferenz.”, Geothermie, 2022. Available: https://www.geothermie.de/fileadmin/user_upload/Aktuelles/Termine/GeoKW_Abschlusskonferenz.pdf. [Accessed: Oct. 14, 2024]. [2] J. Pelzer and M. Schulte, “Efficient two-stage modeling of heat plume interactions of geothermal heat pumps in shallow aquifers using convolutional neural networks,” Geoenergy Science and Engineering, Jun. 2024, doi: 10.1016/j.geoen.2024.212788.