Due to increasing urbanization and scarcity of land, the need to include the underground in urban planning is growing, especially in cities. The human-made underground structures generate an anthropogenic heat flow which has a significant impact on the ground temperature and leads to the creation of urban underground heat islands (UUHI). In parallel with the urbanization, the consumption of energy increases and the interest in sustainable energy production from thermal energy sources is growing. The UUHI have a high geothermal potential and are therefore of great interest for energy production. A sustainable evaluation of the geothermal potential in cities depends however on many factors and is therefore a complex procedure in which different scenarios must be assessed. This master thesis proposes a machine learning (ML) based approach for the evaluation of the underground waste heat flow of buildings. ML techniques are suitable to very quick and accurate modeling and can therefore be of great interest in the field of complex geothermal potential evaluation. In this project we will focus on the ground temperature change due to heat losses of basements. To this effect, 2D heat maps are created for a given depth and a given simulation time period. In this study, a comprehensive validation of the proposed ML model (random forest algorithm) is presented for different small scale scenarios, using various heat source geometries, locations, numbers and thermal loads. Furthermore, we study the heat maps for diverse depths and simulation time periods. To do so, we assume only a conductive heat flow mode and constant boundary conditions. The proposed methodology involves having access to fully solved numerical simulation to create a data set which is used to train and test the ML algorithm. For this purpose, the finite element software package COMSOL is used. We evaluate and discuss the accuracy of the proposed ML approach on each studied scenario. The results promise a big potential all while requiring only few solved FE models. We obtain a root mean squared error of always less than 5 to 10% depending on the scenario. At the end of this paper a district scale application of the proposed ML technique is presented on the Loop district in Downtown Chicago, US. It will be shown that 25% of the total Loop data from the FE model is sufficient to learn the ML model and to predict the whole district with a root mean squared error of less than 5%.