2021/02/10 by Javier Santos, Santos, Javier, Ying Yin +15
Engineering · Computer Science · #Enhanced Oil Recovery Techniques #Advanced Mathematical Modeling in Engineering #Lattice Boltzmann Simulation Studies
paper · pdf · doi:10.48550/arxiv.2102.07625
The permeability of complex porous materials can be obtained via direct flow\nsimulation, which provides the most accurate results, but is very\ncomputationally expensive. In particular, the simulation convergence time\nscales poorly as simulation domains become tighter or more heterogeneous.\nSemi-analytical models that rely on averaged structural properties (i.e.\nporosity and tortuosity) have been proposed, but these features only summarize\nthe domain, resulting in limited applicability. On the other hand, data-driven\nmachine learning approaches have shown great promise for building more general\nmodels by virtue of accounting for the spatial arrangement of the domains solid\nboundaries. However, prior approaches building on the Convolutional Neural\nNetwork (ConvNet) literature concerning 2D image recognition problems do not\nscale well to the large 3D domains required to obtain a Representative\nElementary Volume (REV). As such, most prior work focused on homogeneous\nsamples, where a small REV entails that that the global nature of fluid flow\ncould be mostly neglected, and accordingly, the memory bottleneck of addressing\n3D domains with ConvNets was side-stepped. Therefore, important geometries such\nas fractures and vuggy domains could not be well-modeled. In this work, we\naddress this limitation with a general multiscale deep learning model that is\nable to learn from porous media simulation data. By using a coupled set of\nneural networks that view the domain on different scales, we enable the\nevaluation of large images in approximately one second on a single Graphics\nProcessing Unit. This model architecture opens up the possibility of modeling\ndomain sizes that would not be feasible using traditional direct simulation\ntools on a desktop computer.\n