The computational effort of a simulation can be reduced by running simulations on a coarse grid and interpolating to a fine one. This interpolation can be done using data-driven neural networks, called super-resolution. To minimize the need to perform expensive simulations to create the datasets required for training, physics-informed neural networks (PINNs) add a physical error term to the learning process. In this work, we extend and apply a super-resolution PINN approach to a groundwater simulation with heat pumps. We extend an existing network model for flow velocity and pressure to include temperature and permeability of the soil and derive a corresponding error term. The model is trained on a data set from a simulation with different pressures and permeabilities. The results are compared with a data-driven network and bicubic interpolation. We find that both neural networks significantly outperform bicubic interpolation, whereas the PINN approach achieves slightly better results than the data-driven network.