Two and Three dimensional predictions of permeability using machine learning
Abstract
Traditional approaches to the prediction of effective properties from porous media focus on coupling the 3D structural imaging of porous media with the full physical simulation of the partial differential equations governing the property of interest. The past 10 years have seen an explosion in the availability of open source machine learning and computer vision tools, allowing for an both in-depth multivariant analysis of porous media structure and powerful new regression techniques. These allow for new approaches for machine learning based prediction of flow and transport properties.
A series of over 200,000 coupled synthetic pore networks and permeabilities were generated using object based network generation and digital rock flow techniques. The networks were then analyzed for the statistics of their structural properties, which were then regressed against permeability using open-source machine learning tools, creating a predictive model. This predictive model was validated using a composite set of real digital rock images ranging in voxel size from 32nm to 10μm, with predicted permeabilities varying over 6 orders of magnitude (10-17m2 to 10-11D). These predictions were then compared with full-physics simulations on the same geometries with an average fractional error of <25%. The most important advantage of such a technique is that it greatly expands the forms of data available for quantitative flow prediction, as the machine learning algorithm can be run in both 2D and 3D. Effective property prediction can be performed from 2D data (e.g. light or electron microscopy), which can be acquired over much larger areas, allowing a much better characterization of structural heterogeneity. This approach was validated using a correlative light, electron and nano-XRM workflow, allowing for the extraction of samples (for nm scale imaging) from large area data, showing consistent results from flow predictions made using the machine learning model trained in 2D (taking either light or electron microscopy data) and the same regions when imaged in 3D (using either micro or nanoXRM) (figure 1). Large area predictions of flow properties can then be used to create bespoke (lithology specific) upscaling functions, allowing for the direct integration of pore and core scales of rock description.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H23M2079A
- Keywords:
-
- 1822 Geomechanics;
- HYDROLOGY;
- 1835 Hydrogeophysics;
- HYDROLOGY;
- 1847 Modeling;
- HYDROLOGY;
- 4558 Sediment transport;
- OCEANOGRAPHY: PHYSICAL