Stanford Geothermal WorkshopFebruary 9-11, 2026

Accurately predicting flow and thermal performance in fractured reservoirs is essential for the development and optimization of enhanced geothermal systems (EGS). Fracture geometry, connectivity, and heterogeneity strongly control reservoir behavior, but these features are high-dimensional and difficult to observe directly, leading to significant uncertainty in flow and thermal predictions. A key challenge is how to efficiently tune fracture networks using limited observations to reduce the prediction uncertainty. In this work, we propose a latent generative modeling framework to parameterize discrete fracture networks (DFNs) with low-dimensional latent variables while preserving fracture geometry and connectivity. The model is trained on ensembles of 2D DFNs and can generate new realizations that are visually and statistically consistent with the training data. We integrate this generative representation with embedded discrete fracture model (EDFM) simulations and apply an ensemble-based data assimilation method for history matching. By assimilating temperature and tracer observations, this framework updates the latent variables and tune fracture networks that better match observed data. We validate the method using a synthetic 2D EGS test case. The posterior results present significant uncertainty reduction in temperature predictions, along with fracture networks that more closely agree with the ‘true’ synthetic fracture networks. This study demonstrates the potential of deep generative modeling for efficient fracture parameterization and uncertainty reduction in geothermal reservoir characterization.

Topic: Modeling