Unlocking the full potential of groundwater heat pump deployment in a given region demands a good understanding of the specific groundwater flow dynamics to model the formation of heat plumes and their interactions [1]. This talk presents an approach to model groundwater flow with the goal of optimizing the placement of heat pumps for the example of the metropolitan region of Munich. For practical use in the field, this has to be orders of magnitude more efficient than classical simulations. One of the main difficulties is the correct modelling of the interaction between adjacent plumes in close proximity. Another difficulty are the long-distance influences of locally varying subsurface properties like a varying permeability field. Addressing these challenges, we pursue two approaches. The first approach builds a fully convolutional neural network (CNN) that is trained on small cutouts of a larger simulation with many heat pumps and varying subsurface properties; and applied to arbitrary, large domains as shown in Fig. 3. In the second, bottom-up approach, we place a small box around each heat pump to learn its undisturbed plume. In a second stage, their interaction is learned with a second CNN. To transport all relevant information through the domain, the input feature of the heat pump positions is transformed by variations of a signed distance function.