DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication
Otros (Other)
linuswalter/PINN-for-Subsurface-Flow
This software computes the diffusion of fluid pressure p(x,t) in a 1D domain based on the concept of Physics Informed Neural Networks (PINN) for a heterogeneous modeling domain., GNU General Public License 2 or later (GPL-2.0), Field-data assimilation to calibrate rock properties in numerical physics-based reservoir models of hydrogeological applications is challenging. Recently, Artificial Neural Networks (ANN) have emerged as a promising alternative to handle noisy data seamlessly. However, in hydrogeology, even though data are abundant over time, observation wells are sparse over space, which results in insufficient data to train ANN. Here, we propose Physics-Informed Neural Networks (PINN) to bridge the gap of sparse spatial data by imposing physical conditions. We test ANN and PINN on a synthetic dataset for fluid pressure diffusion p(x,t) through a low-permeable porous medium that hosts a high-permeability equivalent fracture material. By comparing the ANN and the PINN as a function of the number of observation wells, we find that the PINN model outperforms the ANN when having less than 14 wells. Adding noise to the training data reveals the advantage of PINN to be more robust for random measurement errors or ambient noise. We finally test the applicability and limitations of PINN to represent fractures as a thin equivalent material while having no observations inside the domain and find that accuracy can be maintained when reducing the thickness of the equivalent material at the expense of increasing the computation time. Given that hydrogeological applications count with a limited number of observation wells, PINN appears as a more suitable machine learning tool than purely data-driven ANN for reservoir modeling. Therefore, we consider this work a starting point for developing more realistic reservoir models with heterogeneous material distribution based on the PINN architecture., LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809., Peer reviewed
Project:
EC / H2020 801809