Predictions of flow and transport in fractured media in the subsurface using graph-based machine learning
Los Alamos National Laboratory | LANL
Flow through fractured media in the sub-surface is heavily dependent on microstructural information. Continuum models often eliminate features critical to accurately predicting macroscale behavior but are commonly used since resolving thousands of fractures individually is computationally intractable. We overcome this hurdle by developing compact graph representations of the fracture networks, and using machine learning algorithms to mimic the detailed physics at the microscale. The resulting workflow has been shown to achieve up to 4 orders of magnitude computational speedup while maintaining accuracy.