Aquifer Thermal Energy Storage (ATES) is a promising technology for reducing the energy use and carbon emissions associated with heating and cooling of buildings. However, ATES performance depends on complex interactions between groundwater flow, subsurface heat transport, building energy demand, heat-pump operation, and long-term operational strategies. Accurate prediction and optimization therefore require computational models capable of representing both subsurface physics and energy-system dynamics. This project develops an AI-enhanced modeling and optimization framework for ATES systems by integrating physics-based numerical simulation, Physics-Informed Neural Networks (PINNs), long-term monitoring data, and energy-system co-simulation. The research is based on a monitored ATES installation in Stockholm serving approximately 30,000 m² of commercial building area. The available dataset contains long-term measurements of groundwater temperatures, well flow rates, groundwater levels, distributed temperature sensing (DTS), building heating and cooling demand, heat-pump operation, and environmental conditions. High-fidelity groundwater-flow and heat-transport models will be developed and calibrated using MODFLOW 6 and associated Python tools. These models will provide physical reference simulations and training and validation data for PINNs that incorporate governing equations and observational data into machine-learning models of subsurface thermal dynamics. The resulting models will be coupled with building and heat-pump system models using Modelica/FMU-based co-simulation. The project will investigate model calibration, thermal-plume evolution, long-term ATES performance, uncertainty, operational scenarios, and data-driven prediction and control. The ultimate objective is to develop computationally efficient models that can support improved design, forecasting, and operation of ATES systems while retaining physical consistency. The requested NAISS resources will enable large numerical simulation ensembles and GPU-accelerated PINN training that are computationally impractical on local workstation hardware.