Seeing permeability from images: fast prediction with convolutional neural networks

Abstract

Fast prediction of permeability directly from images enabled by image recognition neural networks is a novel pore-scale modeling method that has a great potential. This article presents a framework that includes (1) generation of porous media samples, (2) computation of permeability via fluid dynamics simulations, (3) training of convolutional neural networks (CNN) with simulated data, and (4) validations against simulations. Comparison of machine learning results and the ground truths suggests excellent predictive performance across a wide range of porosities and pore geometries, especially for those with dilated pores. Owning to such heterogeneity, the permeability cannot be estimated using the conventional Kozeny-Carman approach. Computational time was reduced by several orders of magnitude compared to fluid dynamic simulations. We found that, by including physical parameters that are known to affect permeability into the neural network, the physics-informed CNN generated better results than regular CNN, however improvements vary with implemented heterogeneity.Computation of pore-scale transport properties from pore-scale images is an important aspect of image-based pore-scale studies. Such computations are generally performed in two ways, i.e., direct simulation approach and simplified network approach. In the first approach, the microscopic transport equations are solved directly on the geometry shown by the porescale images to obtain averaged properties such as permeability, relative permeability, or dispersion coefficient. Both single and multiphase flows can be accounted for, and both reactive and non-reactive transport equations can be solved. This direct approach is generally considered to be more accurate, but the computational cost is very high. For processes such as multiphase flows and reactive transport with slow kinetics, it is nearly impossible to solve the governing equations in a medium of even a moderate size. Therefore, the second alternative approach is to first abstract the porous medium as a discrete network. By applying simplified flow and transport laws on the network, the computational cost to obtain averaged properties can be effectively lowered [11].Some transport properties of porous media such as permeability are solely functions of pore geometry. Therefore, it should be possible to predict them using a neural network approach, which is to develop a surrogate model that directly maps a pore geometry to physical properties. Such a task resembles that in image classification [12,13], where a model takes an image as input and give the classification label as output by recognizing the object in the image, e.g., cars, animals, or even subtypes thereof (i.e., car make or animal breed). Once constructed, such surrogate models can potentially enable fast prediction of physical properties of porous media without performing direct simulations or network calculations. The recent studies of chemical imaging of rocks also involve surrogate models. For example, Hao et al.[14] generated a molecula...

Predicting permeability via statistical learning on higher-order microstructural information

“…As a complement to rigorous approaches to estimate effective properties from the microstructure, data-driven methodologies to establish structure-property relationships are increasingly being used [43][44][45][46][47][48][49] . The rapid increase in computational resources facilitates the computation of effective properties for very large data sets (hundreds or thousands) of different microstructures.…”

“…Barman et al 46 studied effective diffusivity prediction in 36 virtual porous polymer films using tortuosity and constrictivity. In a different direction, there are several attempts to use 2D and 3D convolutional neural networks (CNNs) to extract information directly from the binary image data describing the structure [47][48][49][51][52][53] in order to predict effective properties. However, these models are usually difficult to interpret and hard to rescale.…”

A Convolutional Neural Network Architecture for Auto-Detection of Landslide Photographs to Assess Citizen Science and Volunteered Geographic Information Data Quality

“…They found that CNNs demonstrate outstanding performance in predicting the physical parameters of rock if sufficient input data are available. Wu et al [32] predicted the permeability from images when applying CNN and presented a procedure using CNN with simulated data.…”

Towards Bayesian Quantification of Permeability in Micro-scale Porous Structures – The Database of Micro Networks

“…In particular, the quantification of the permeability (PE) of a porous structure has gained attractionowing to its significance in the petroleum industry. In this sake, the stochastic modeling [20] and deep learning strategies have notably been used [21][22][23][24][25][26][27][28] with the latter being applied to the log data as well as the 2D/3D CT-scan images. For stochastic modeling, Eugene et al (2005) [20] applied a combination of lattice Boltzmann method (LBM) and firstorder reliability method (FORM) to construct cumulative distribution functions for randomly-generated porous structures.…”

“…Furthermore, the deep learning strategy was applied to core data, on a microscopic scale, where processing of the micro-CT images has been on focus. For permeability prediction, Jinlong et al (2018) [25] proposed a physics-informed convolutional neural network approach, which involves a series of fluid dynamics simulations in order to build on the training dataset required. On the other hand, geometrical features in binary segmented images were used by some researchers to estimate permeability, applying multilayer neural network (MNN) and convolutional neural network (CNN) methods [26].…”