H036-0007
Multiscale networks for learning fluid flow through fractured permeable media

Tuesday, 8 December 2020
Poster
Javier E. Santos1, Ying Yin2, Nicolas Lubbers3, Masa Prodanovic4, Michael Pyrcz1, Hari Selvi Viswanathan5 and Qinjun Kang6, (1)University of Texas at Austin, Austin, TX, United States, (2)Xi'an Jiaotong University, Xian, China, (3)Los Alamos National Lab, Los Alamos, United States, (4)The University of Texas at Austin, Hildebrand Department of Petroleum and Geosystems Engineering, Austin, TX, United States, (5)Los Alamos National Laboratory, Los Alamos, NM, United States, (6)Los Alamos National Lab, Los Alamos, NM, United States
Abstract:
Forecasting fluid flow through fractured permeable media is of great importance for subsurface engineers. Modern high-performance computers can compute fluid flow via direct simulation methods (using Lattice Boltzmann or similar schemes) in large 3D volumes, but with significant computational effort -- a machine learning (ML) model that returns results in seconds for usage as a screening or analysis tool would be of great benefit.

Recent work has shown that 3D convolutional neural networks (CNNs) can learn the velocity fields associated with homogeneous porous geometries (i.e. spherepacks), where a local region around 803 voxels, is analyzed, and the results of many local regions are stitched together to construct the flow field of the entire system. This paradigm rests crucially on the stationarity of flow statistics, and thus depends on a statistically homogenous pore geometry. Which means that a local heterogeneity should not have a strong impact on the final solution.

However, stationary fluid flow is a fundamentally global problem, and the presence of fractures and other multi-scale geometric characteristics fundamentally require the ML method to train the entire system at once. Prior network architectures fail at this problem -- most of these were constructed in the context of 2D image processing workflows and learn with many consecutive layers of successively larger feature maps. As a result, their computational pre-factor makes them extremely difficult to scale to large 3D systems.

Here we present our efforts to develop a CNN network architecture geared to capturing geometric information across large 3D volumes with far lower memory and computational cost, allowing the processing of far larger sample volumes by a single model on a single GPU. This is based on a fundamentally hierarchical, multiscale network design. Along the way, we pay special attention to the physics of fluid flow and introduce domain-specific physics informed strategies for enforcing the properties of fluids in porous media.

<< Previous Abstract | Next Abstract >>