Geothermal energy is a renewable source of heat and electricity which can contribute to Europe's future energy independence. To scale this technology, reservoir characterization in geothermal projects needs to become more reliable. A major challenge here is to identify the spatial distribution of distinct rock strata and their hydraulic properties from well-pumping tests. These experiments are carried out to infer permeability, but offer only a few observation points in space and require large extrapolation through inversion. Physics Informed Neural Networks (PINN) are a recently emerging approach that can seamlessly incorporate pumping test data while enforcing accordance with physical laws in the domain between observation points. We implement this concept via the loss function of an Artificial Neural Network (ANN) that consists of several distinct loss terms for observational data and a mass balance equation. In this study, we represent the heterogeneous permeability k(x) and the spatiotemporal fluid pressure distribution p(x,t) each by an ANN, respectively. We compare our regression output to a virtual experimental dataset from the numerical code "OpenGeoSys". The results indicate that our PINN model can approximate k(x) and interpolate p(x,t) in a 1D domain for a setup of real-world material parameters. In this way, PINN proves as a promising method that deserves further investigation to also provide reliable performance at the reservoir scale.

Trabajo presentado en IDAEA - Young Researchers' Day YRW (2022), celebrado en Barcelona el 10 de noviembre de 2022.