Machine Learning-based History Matching of Discrete Fracture Network Fields at Utah FORGE Enhanced Geothermal System

Abstract

Characterizing the hydrogeological and mechanical properties of an enhanced geothermal system (EGS) is very important for optimal stimulations, efficient and sustainable heat recovery, and safe operations. However, EGS sites are highly heterogeneous with complex geological structures such as discrete fracture networks (DFNs). Moreover, traditional subsurface characterization approaches require a number of thermal-hydraulic-mechanical (THM) simulations, which poses a significant limit on extensive EGS characterization. In this work, we use a deep generative diffusion model and DFN models to characterize the permeability field of the EGS sites and connectivity between natural and induced DFNs with available observation data sets. The diffusion model, a deep generative model, is used to learn the probability distribution of the permeability field based on conceptual geological models of DFNs. The latent diffusion model is chosen in this work to represent the permeability fields in a significantly smaller dimension than the actual numerical grid dimensions. Ensemble smoother-multiple data assimilation through the latent space is then employed to characterize EGS permeability fields by matching multiple measured data such as temperature, pressure, and production rates. For accurate and fast forward simulations, a surrogate model is constructed using the deep learning framework for THM simulations. Examples with discrete fractured permeability fields and with corresponding continuum-based permeability fields based on the Utah FORGE site, where a recirculation test was performed in the presence of DFNs developed by previous multiple stimulations, are presented to show the performance of our proposed data assimilation framework. After model calibration and validation against recirculation tests at the Utah FORGE site, we will evaluate the effect of a few conceptual models such as near wellbore permeability with DFNs and interaction between natural and induced DFNs for long-term behaviors of heat recovery. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.