betaFeedback

WorksPhysics and Astronomy

POREMAPS: A finite difference based Porous Media Anisotropic Permeability Solver for Stokes flow

Abstract

Porous materials are ubiquitous in various engineering and geological applications, where their permeability plays a critical role in viscous fluid flow and transport phenomena. Understanding and characterizing the microscale properties, the effective hydraulic parameters, and also the anisotropy of porous materials are essential for accurate modeling and predicting fluid flow behavior. The study pursues the Digital Rock Physics approach to retrieve intrinsic permeability and its evolution in anisotropic configurations of porous media, which are subjected to pore space alterations. Therefore, we discuss the development and implementation of a computational framework based on the finite difference method to solve the pseudo-unsteady Stokes equations for fluid flow on the pore scale. We present an efficient and highly parallelized implementation of this numerical method for large voxel-based data sets originating from different image-based experimental setups. A comprehensive variety of benchmarks has been conducted to assess and evaluate the performance of the proposed solver. The solver's compatibility with huge domain sizes generated by state-of-the-art imaging techniques is demonstrated. We investigate an open-cell foam undergoing deformation, observing that contrary to initial expectations, no anisotropy emerges. Further, we examine a microfluidic cell experiencing precipitation within its pore space, resulting in clear anisotropic development during the clogging process.

Introduction

Experimental investigations and numerical simulations of fluid flow through porous materials with a large variety of hydro-mechanical properties are of interest in many fields such as hydrosystems modeling, groundwater flow and contaminant transport [8, 25], oil reservoir exploitation [4, 15] and industrial applications like membranes for water desalination [55]. The majority of all investigations operate at the macro, i.e., Darcy scale, and use a coarse-grained continuum theory, which means that individual pore geometries are not physically resolved. Further, the correlation between viscous fluid flow and corresponding driving force is assumed to be linear. Porosity ϕ\phi and the intrinsic permeability tensor k=kij ei⊗ej\mathbf{k} = k_{ij}\,\mathbf{e}_i \otimes\mathbf{e}_j serve as input parameters for these models. Regimes where the components of the permeability tensor kijk_{ij} do not linearly depend on the driving force are denoted as non-Darcian and are beyond the scope of this contribution. In general, ϕ\phi and effective hydraulic properties like k\mathbf{k} of the porous material are highly sensitive to the distinct porous microstructure. This work primarily centers on the assessment and determination of the intrinsic permeability tensor k\mathbf{k}. There are numerous experimental, semi-analytical and numerical methodologies for permeability determination documented in the literature. Given the expense and complexity of flow experiments and their limitations, and the limited applicability of semi-analytical models it is preferable to assess these parameters with computer simulations. Advances in three-dimensional (3D) imaging techniques like micro X-Ray Computed Tomography (µXRCT) and the meanwhile widespread availability of appropriate devices [65, 64, 63, 46] allow to fully resolve 3D pore geometries at high-resolution and allow for the flow problem to be solved directly at the pore scale for many porous materials [12]. Moreover, two-dimensional (2D2\mathrm{D}) data, e.g., generated from images of microfluidic experiments acquired by optical microscopy, extended in the third spatial direction can provide further input data [28] (Weinhardt, Deng, Hommel, et al. 2022).

Deformation of the pore space or precipitation and a corresponding decrease in porosity have a direct influence on hydro-mechanical properties of a material. Certainly, the process of experimentally determining anisotropic permeability in the course of an experiment is inherently intricate. Acquiring the complete permeability tensor k\mathbf{k} is mostly impossible. Nonetheless, this investigation is particularly significant due to the non-trivial implications of deformation or precipitation on the principal directions of hydraulic properties or the material’s anisotropic characteristics. In the literature, numerical solvers are applied to porous materials undergoing changes in pore space, cf. for geochemical alteration like calcite precipitation [41], carbonate precipitation for carbon capture and storage applications [31] or external strain causing deformation [5, 26]. Another field of interest is the benchmark of specially designed materials whereby the permeability can be controlled via the manufacturing parameters on Al-Cu alloys [10]. Due to the limitations of experimentally viable boundary conditions, e.g. undrained boundary conditions along the cylinder of the sample in conventional triaxial cells, a priori assumption has to be made for the orientation of principal directions of k\mathbf{k}. The utilization of the aforementioned imaging techniques combined with computer simulations provides a way for achieving this in a systematic way while circumventing any a priori assumption. However, in order to determine changes in anisotropy or computing porosity-permeability relationships, for example, many data points and correspondingly many individual simulations are required. While various numerical methods like Finite Element Method (FEM) [13, 17], Finite Volume Method (FVM) [17, 36] and Lattice-Boltzmann Methods (LBM) [13, 37], Smoothed Particle Hydrodynamics (SPH) [42, 54, 27] or Pore Network Modeling (PNM) [21, 44, 11] have been employed to simulate pore-scale fluid flow, each method comes with its own set of limitations and advantages. In this work, we present an efficient and reliable solver for the Stokes equations by implementing a Finite Difference Method (FDM) based algorithm. In terms of numerical efficiency, the solver is tuned for voxel-based cartesian grids as directly obtained from image-based characterization methods like XRCT or microfluidics. It stands out as a versatile and well-established choice due to its straightforward implementation and suitability for complex geometries [43, 9, 22, 35, 1]. In addition, FDM is a simple approach in terms of the algorithm, has advantages when using regular grids, and can be effectively parallelized due to its local nature.

The objective is to develop a resilient and modular open-source tool that advances our comprehension of fluid flow at the pore-scale and addresses the issue of evolving anisotropies in porous media under varying conditions. The software tool should be able to handle state-of-the-art data sets from XRCT scans (up to 200032000^3 voxels) on various computing architectures and should allow for further extensions. Besides research codes which do not allow for further extensions (only executable of the software available [22]) or commercial solvers such as GeoDict [39], there are common open-source packages such as OpenFoam, e.g. used in Icardi, Boccardo, Marchisio, Tosco, and Sethi (2014); Guibert, Horgue, Debenest, and Quintard (2016), which are extremely flexible but therefore not tuned for the mentioned demands. Others, like the tool from the National Institute for Standards and Technology [9] which is written in Fortran, are not multi-node parallelized. The rationale behind developing a new solver, despite the existence of current solutions, is multilayered. We present a FDM solver, exclusively dependent on the Message Passing Interface (MPI), that is fully open source and platform independent and works directly on binarized data sets which essentially is an important tool in porous media research. Moreover, its full integration into existing experimental setups streamlines the research process, promoting a more cohesive and comprehensive approach to investigate complex phenomena. Since large domains (state-of-the-art XRCT data-sets comprise up to 10910^9–101010^{10} voxels) have to be simulated, in combination with several snapshots for time-resolved investigations, the efficiency and performance as well as the simplicity of the presented FDM solver is the main focus. The code is written in procedural, functional C++ and is parallelized with the MPI to run on multi-node, multi-core CPU systems.

For these reasons, in addition to the permeability determination derivation (Sec. 2), the mathematical basics (Sec. 3), and the numerical principles (Sec. 4), it is particularly important for us to address the implementation and technical aspects (Sec. 5) as well as to provide a detailed validation against various benchmark cases (Sec. 6). In § 7, we demonstrate the developed solver to investigate two distinct materials characterized by alterations in their pore space. These alterations are anticipated to result in changes not only in the magnitude of the permeability but also in the permeability anisotropy ratio. Specifically, our study focuses on an open-cell foam, which defies expectations by not exhibiting anisotropy as it undergoes deformation. Additionally, we investigate a porous micro-structure exposed to mineral precipitation and subsequently clogging, revealing a development of anisotropy that is notably influenced by the boundary conditions of the underlying experiment.

Permeability tensor and principal permeabilities

Permeability is defined as a proportionality factor between pressure gradient across the examined sample and fluid fluxes (grad⁡p∝q\operatorname{grad} p \propto\mathbf{q}). For homogeneous materials, that are known to have isotropic material behavior a scalar value is used for the hydraulic permeability. Intrinsic permeability quantifies viscous losses in continuum-based Darcy-type models. In the generic case, the effective permeability k\mathbf{k} is a second order tensor

k=kij ei⊗ej=[k11k12k13k21k22k23k31k32k33]ei⊗ej(1)\mathbf{k}=k_{ij}\,\mathbf{e}_i\otimes\mathbf{e}_j= \begin{bmatrix} k_{11} & k_{12} & k_{13}\\ k_{21} & k_{22} & k_{23}\\ k_{31} & k_{32} & k_{33} \end{bmatrix}\mathbf{e}_i\otimes\mathbf{e}_j \tag*{(1)}

where kijk_{ij} are the components and the cartesian basis vectors ei\mathbf{e}_i build the tensorial basis through its dyadic product. Bold notation is employed to represent tensors and vectors, with the implicit assumption that equations in index notation abide by Einstein’s summation convention. The coefficient matrix is symmetric and positive definite [7, 51]. In contrast to the experimental determination of permeability, the numerical approach is capable of computing secondary diagonal elements of kijk_{ij}, which is required for a complete hydraulic characterization. Solving the characteristic polynomial det⁡(k−λI)=0\det(\mathbf{k}-\lambda\mathbf{I})=0 for the eigenvalues λi\lambda_i (also referred to principal permeabilities kIk_I, kIIk_{II}, kIIIk_{III}) we obtain the permeability tensor in its spectral form

k=∑i=13λi e~i⊗e~i(2)\mathbf{k}=\sum_{i=1}^{3}\lambda_i\,\tilde{\mathbf{e}}_i\otimes\tilde{\mathbf{e}}_i \tag*{(2)}

expressed in a basis system with the principle axes e~i\tilde{\mathbf{e}}_i, that are computed by solving (k−λiI)⋅e~i=0(\mathbf{k}-\lambda_i\mathbf{I})\cdot\tilde{\mathbf{e}}_i=0. I\mathbf{I} is the second order identity tensor. The orientation of the principal axes with respect to the basis system ei\mathbf{e}_i is computed via the rotation tensor R=Rik(ei⊗ek)\mathbf{R}=R_{ik}(\mathbf{e}_i\otimes\mathbf{e}_k) whereby e~i=R⋅ei\tilde{\mathbf{e}}_i=\mathbf{R}\cdot\mathbf{e}_i and Rki=cos⁡∠(ek,e~i)R_{ki}=\cos\angle(\mathbf{e}_k,\tilde{\mathbf{e}}_i) hold. Determining the principal permeabilities enables categorization of the material into one of the following classes:

  1. Isotropy: kI=kII=kIIIk_I=k_{II}=k_{III} Same hydraulic properties in all three principal directions.

  1. Orthotropy: kI≠kII ∧ kI≠kIII ∧ kII≠kIIIk_I\ne k_{II}\ \land\ k_I\ne k_{III}\ \land\ k_{II}\ne k_{III} Unique and independent hydraulic properties in three mutually perpendicular directions.

  1. Transverse anisotropy: kI≠kII=kIII ∨ kI=kII≠kIII ∨ kI=kIII≠kIIk_I\ne k_{II}=k_{III}\ \lor\ k_I=k_{II}\ne k_{III}\ \lor\ k_I=k_{III}\ne k_{II} Same hydraulic properties in one plane. Thus, there are two independent constants in the permeability tensor. Typical examples of this material class are wood, unidirectional fiber composites or sedimentary sandstones.

Governing equations on the pore scale

Pore scale refers to the length scale at which individual pores and their geometrical features, such as pore size and shape, are significant. At this scale, the fluid flow pattern is strongly influenced by the morphology of the pores, as well as the viscous momentum interactions between the fluid phase and the solid skeleton. Pore scale simulations are used to study fundamental aspects of porous media flow as well as the determination of properties, used in coarse-grained Darcy scale methods. We are interested on solving for intrinsic effective permeability and therefore consider solving fluid flow at stationary creeping flow conditions. Accordingly, the following applies for the Reynolds number

Re⁡=ρ0 V Lμ<1[−],(3)\operatorname{Re}=\frac{\rho_0\,\mathcal{V}\,\mathcal{L}}{\mu}<1\qquad[-], \tag*{(3)}

and the fundamental equations to be solved are the Stokes equations consisting of the balance of linear momentum for an incompressible Newtonian fluid [43]

0=μ div⁡(grad⁡v)−grad⁡p(4)\mathbf{0}=\mu\,\operatorname{div}(\operatorname{grad}\mathbf{v})-\operatorname{grad}p \tag*{(4)}

and the balance of mass

div⁡v=0(5)\operatorname{div}\mathbf{v}=0 \tag*{(5)}

where ρ\rho, ρ0\rho_0, v\mathbf{v}, μ\mu, pp, V\mathcal{V}, L\mathcal{L} are the fluid density, the rest density of the fluid, fluid velocity vector, (constant) dynamic viscosity of the fluid, pore fluid pressure and the characteristic velocity and length, respectively. No body force is used. In porous media L\mathcal{L} is assumed as the characteristic (mean) pore diameter and V\mathcal{V} is calculated from the applied/measured flux q\mathbf{q}. In addition to characteristic properties, indicated by notation in script typestyle, to scale physical quantities we denote dimensionless variables and operators by (⋅)∗(\cdot)^*. It is well known that solving Eqs. (4) and (5) causes problems due to the special role of pressure. Therefore, this set of governing equations is extended with an artificial time derivative and a so-called pseudo-unsteady method is used [43]. In this manner, for the numerical implementation, we do not fulfill the incompressibility condition as an algebraic constraint. Instead an artificial compressibility formulation will considered with an equation of state for the fluid pressure p(ρ)p(\rho). Hence, we obtain for the nondimensional equations to be solved

ρ∗v˙∗=1Rediv⁡grad⁡(v∗)−grad⁡(p∗)andp˙∗=−c∗2div⁡(v∗)(6)\rho^* \dot{\mathbf{v}}^* = \frac{1}{\mathrm{Re}}\operatorname{div}\operatorname{grad}(\mathbf{v}^*)-\operatorname{grad}(p^*) \quad\text{and} \quad\dot{p}^*=-c^{*2}\operatorname{div}(\mathbf{v}^*) \tag*{(6)}

where c∗c^* is the dimensionless speed of sound also referred to artificial compressibility parameter. This numerical approach solves the steady-state Stokes equations by transforming them into a pseudo-time dependent problem which is feasible when dealing with low Reynolds number fluid flows in porous media. It provides a computationally efficient alternative to directly solving the steady-state equations and it can be shown that the solution converges for t→∞t \to\infty to the steady-state solution of the original problem [18, 22, 43]. The detailed nondimesionalization with introduction of all quantities, dimensionless differential operators and constitutive equations can be found in Appendix § 9.

Finite difference scheme for Stokes equations

Simulation parametrization: We introduce the reference length L=1\mathcal{L}=1 vx (voxel) and the reference velocity V=1Re\mathcal{V}=\frac{1}{\mathrm{Re}}. As driving force a constant pressure gradient across the domain is employed and as initial conditions we set grad⁡(p∗)=1vx\operatorname{grad}(p^*)=\frac{1}{\mathrm{vx}}. Furthermore, no body force is present (b=0\mathbf{b}=0). We require small Reynolds number and fix Re=0.01\mathrm{Re}=0.01 and c∗2=1.5×106c^{*2}=1.5\times10^6 as in Bentz and Martys [9]. Note that the dimension c∗2≈106c^{*2}\approx10^6 fits well to the physical ratio of the speed of sound in water to the assumed characteristic velocity of creeping flow conditions.

Grid: Space is discretized with a regular staggered Marker-and-Cell (MAC) grid [24], where fluid velocities are stored on the faces of the cells, while the pressure values are stored at the cell centers. Storing different quantities on different locations within each cell, allowing for an efficient and accurate computation of the pressure gradients and additionally for the exact modeling of no-penetration conditions at the fluid-solid interfaces.

Boundary in the numerically considered domain and second order derivatives: For the outer boundaries we apply periodic conditions to simulate a periodic or repeating behavior of the physical system in a unit cell being modeled or use no-slip no-penetration boundary conditions initially assumed to be not suitable for anisotropy investigations since the transverse flow would be zero. The outer boundary in direction of the pressure gradient is always periodic for the local flux. In main flow direction we employ fixed pressure conditions. Fluid-solid boundaries correspond to the pixel boundaries of the binary 3D image and we use no-slip no-penetration boundary conditions on the interfaces, cf. Figure 1. Second-order derivatives perpendicular to the local flow direction are analytically determined based on Taylor series approximations. This ensures that the no-slip condition is fulfilled on the fluid-solid voxel surfaces. We distinguish 8 basis cases, whereas the last two ones are special cases of cases 4 and 5, and are only required for the second voxel layers used for the boundary conditions. Therefore, we have theoretically the same number of 6 cases as in Bentz and Martys [9]; Gerke et al. [22]. In Figure 1 the eight distinguished cases are exemplary illustrated for the flow direction eke_k and the perpendicular direction eie_i.

Figure 1. Different cases for the identification of the voxel neighborhood to calculate the second-order derivatives (∂2vk∂xi2)\left(\frac{\partial^2 v_k}{\partial x_i^2}\right) in the perpendicular directions to the considered flow velocity direction for i,j,k=(1,2,3),(2,3,1),(3,1,2)i,j,k=(1,2,3),(2,3,1),(3,1,2).

Based on the velocity definitions in Figure 1, the second-order derivatives for the different cases can be computed by solving a resulting system of equations. For instance, for case 2 we get a set of three equations to compute ∂2vk∂xi2\frac{\partial^2 v_k}{\partial x_i^2} as follows:

\begin{aligned} v_{k1} &\approx v_{k2}-\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0\\ v_{k3} &\approx v_{k2}+\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k4} &\approx v_{k2}+\frac{3}{2}\frac{\partial v_k}{\partial x_i}+\frac{9}{2}\frac{\partial^2v_k}{\partial x_i^2}+\frac{9}{16}\frac{\partial^3v_k}{\partial x_i^3}=0 \end{aligned} \quad\} \quad\sim\quad \frac{\partial^2v_k}{\partial x_i^2}=\frac{8}{3}v_{k3}-\frac{16}{3}v_{k2} \tag*{(7)}\end{aligned}

An overview of all cases of neighborhood consideration can be found in the Appendix B.

Permeability computation: In the so-called creeping flow regime, characterized by low Reynolds numbers, Re⁡<1.0\operatorname{Re}<1.0, it is valid to employ Darcy’s law [19]

q=1μk⋅h  →  qiei=1μkij(ei⊗ej)⋅hkek=1μkikhkei(8)\mathbf{q}=\frac{1}{\mu}\mathbf{k}\cdot\mathbf{h} \;\to\; q_i\mathbf{e}_i=\frac{1}{\mu}k_{ij}(\mathbf{e}_i\otimes\mathbf{e}_j)\cdot h_k\mathbf{e}_k=\frac{1}{\mu}k_{ik}h_k\mathbf{e}_i \tag*{(8)}

to compute the entries of the second order permeability tensor where hi=−∂p∂xi=−p,ih_i=-\frac{\partial p}{\partial x_i}=-p_{,i} is the pressure gradient. In order to determine the nine entries in the coefficient matrix of the permeability tensor, three numerical simulations must be performed. The following pressure gradients hih_i are specified for the different simulations

a) h1≠0∧h2=h3=0;b) h2≠0∧h1=h3=0;c) h3≠0∧h1=h2=0;(9)\text{a) }h_1\ne0\land h_2=h_3=0;\qquad \text{b) }h_2\ne0\land h_1=h_3=0;\qquad \text{c) }h_3\ne0\land h_1=h_2=0; \tag*{(9)}

and we measure three fluxes qiq_i for each case. For case c), all properties are illustrated in Figure 2. By comparing the coefficients (Eq. (8), right), nine equations are obtained for nine entries of the coefficient matrix of the permeability tensor. A detailed list of the equations for determining the complete permeability tensor k\mathbf{k} can be found in Appendix C.

Illustration of the components of Darcy’s law

Figure 2. Illustration of the components of Darcy’s law with the given boundary values for the pressure pinp_{\mathrm{in}} and poutp_{\mathrm{out}} used to compute the gradient h3h_3. Volumetric fluxes q2q_2, q3q_3 are evaluated over the respective boundaries A2A_2, A3A_3. In e1e_1-direction, the volumetric flux is not shown here, but is handled analogously. With given pressure gradients h3≠0h_3\ne0 and h1=h2=0h_1=h_2=0, this represents case c) in Eq. (9).

Computational aspects

Processing input data

Binarized 3D image data (8-bit file format) distinguishing between solid phase and pore space are employed as input for the solver. Typically performed image pre-processing steps depending on the exact imaging technique (e.g., denoising and segmentation) are well-documented in the literature, e.g., [3, 2, 14, 29, 49, 56] (Schlüter, Sheppard, Brown, and Wildenschild 2014) and will not be further elaborated upon here. Depending on the boundary conditions and material, a few pre-processing steps need to be conducted, such as mirroring for symmetric periodicity, domain cropping, or the elimination of disconnected pore spaces. This depends upon the specific problem at hand, a topic that will be delved into within the dedicated application sections. The solver operates under the assumption that the percolation condition is met.

Implementation

The solver is completely implemented in C++ and parallelized with OpenMPI (version 4.1.5) [40] to employ it on distributed memory architectures. Special emphasis is put on keeping the code as simple as possible. For the domain decomposition we use a communicator on a Cartesian topology (MPI_Cart_create) which is particularly well suited for regular meshes and 3D geometries. The file IO is fully parallelized and communication between the ranks is executed by blocking send-receive operations. We work directly with binary data which is RAM-efficient allows for a concise. The source code of POREMAPS is published in Krach, Ruf, and Steeb [33].

Algorithm 1: Program Flow Stokes Solver

Data: Binarized 3D image data Result: Velocity and pressure fields Initialize MPI; Domain decomposition, read partial domains, add halos; Impose initial pressure gradient; Evaluate neighborhood; while div⁡v>ε\operatorname{div} v > \varepsilon do &nbsp;&nbsp;Compute velocity field vi,j,kv_{i,j,k}; &nbsp;&nbsp;Compute pressure field pi,j,kp_{i,j,k}; &nbsp;&nbsp;if i mod CommFrequency≡0i \bmod CommFrequency \equiv0 then &nbsp;&nbsp;&nbsp;&nbsp;Communicate halos &nbsp;&nbsp;end &nbsp;&nbsp;Evaluate permeability and convergence criterion ε\varepsilon; &nbsp;&nbsp;Write log file; end Write pressure and velocity fields;

The inherent domain decomposition of MPI optimizes communication by embedding the virtual topology onto the physical machine as efficient as possible but is not necessarily suitable for domains with high aspect ratios (see Sec. 7.3). Therefore, it is possible to include the desired number of ranks in each direction directly in the input file.

Scaling

The scalability of FDM codes has been studied extensively in the literature. Several factors contribute to weak scalability, including the communication overhead, memory requirements, and load balancing. For the scaling test we use a 50350^3 voxel regular sphere packing per core and the hardware of the experimental compute cluster ”ehlers” of the EXC 2075 ”SimTech” Cluster of Excellence (University of Stuttgart). The CPU partition is comprised of 8 nodes with 128 cores (2 ×\times 64 cores, AMD EPYC 7702) each and 200 Gb/s Infiniband interconnect and 2 TB of RAM.

The algorithm is very communication-heavy. The influence of increasing core-to-core (Figure 3, left) or node-to-node communication (Figure 3, right) causes a decrease in performance. However, as soon as com munication in all three spatial directions is required (>8> 8 cores), the computed time steps per second (TPS) are almost constant. Accordingly, the results are satisfactory for the problems at hand. Considering targeted domain sizes, tests beyond 1024 cores are not significant for our applications and therefore not considered here.

Figure 3. Weak scaling on one node using 1, 2, 4, 8, 16, 32, 64, 128 cores (left) and 1–8 nodes with 128 cores each (right).

Benchmarks

The code has been developed to be able to study complex, heterogeneous porous materials. To justify the application of the solver to diverse domains, a multi-layered benchmark and validation program is conducted. We validate the code against analytical solutions such as Poiseuille flow and channel flow as well as empirical relations such as the Kozeny-Carman equation for sphere packings with different porosities. We further evaluate our workflow for the determination of the anisotropy and the secondary diagonal elements of the permeability tensor. Finally, we compare our code with other solvers like LBM, FEM or mathematical homogenization for different types of porous materials.

Hagen-Poiseuille equation

The laminar creeping flow of a Newtonian fluid through a pipe is a standard benchmark for CFD codes. The tube is per definition periodic in direction of the pressure gradient. We compare results for different resolutions and the analytical Hagen-Poiseuille equation presented in Batchelor [6]

v3(r)=Δp4Lμ(R2−r2)(10)v_3(r) = \frac{\Delta p}{4L\mu}\left(R^2-r^2\right) \tag*{(10)}

with the radial coordinate rr. Length and radius of the tube are L=0.01L = 0.01 m and R=0.001R = 0.001 m. The no-slip condition on the fluid-solid interface results in v3(r=R)=0v_3(r = R) = 0.

We compare simulation results obtained for different resolutions (L/dx={5,10,20,50,100}L/dx = \{5, 10, 20, 50, 100\}) and present the velocity profiles through the center of the tube, see Figure 4, left. All resolutions deliver very similar solutions and are in line with the analytical solution Eq. (10). Figure 4, right, gives the cross-section through the tube visualizing the radially symmetric velocity pattern.

Figure 4. Resolution dependent velocity profiles through a tube compared to the analytical solution (left) and velocity pattern in one cross-section (right).

Channel flow – rectangular cross section

To benchmark slightly more complex structures, we analyze the flow through a rectangular channel. It has an advantage over the tube in § 6.1, since it can be discretized on a cubic lattice without introducing a discretization error. The benchmark provides information about the resolution at which we can expect valid results from the solver. That is of particular interest for the study of microfluidic experiments in order to resolve the channels sufficiently.

Presented are simulation results for different resolutions (h/dx={3,5,10,25,50,100}h/dx = \{3, 5, 10, 25, 50, 100\}) for the channel height hh. The width of the channel equals b=2hb = 2h. Already with a resolution of the channel cross-section using 5×105 \times10 voxels, the volumetric flow rate Q=q⋅AQ = q \cdot A corresponds with results for higher resolutions. Although the solution depends on the discretization, it converges already for low resolutions.

Figure 5. Deviation of the volumetric flux QQ from its analytical solution QanQ_{an} according to White and Majdalani [62] through rectangular cross section channel for different resolutions (left) and corresponding velocity pattern (right).

Regular sphere packings

Due to the periodicity of regular sphere packings, the simulation of flow can be reduced to a cubic unit cell of side-length LL. It is therefore possible to simulate representative porous structures without pushing the domain size too far. In addition, the permeability in direction of the cubic unit cell edges is identical and semi-analytical estimates exist. Therefore, we investigate differently arranged sphere packings (Simple Cubic (SC), Body-Centered Cubic (BCC), Face-Centered Cubic (FCC)) for a sweep over a wide range for the porosity ϕ={0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9}\phi= \{0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9\}. In addition, all simulations are run at three different resolutions L/dx={60,100,200}L/\mathrm{d}x = \{60, 100, 200\}. The results are compared with permeability estimates k1KCk_1^{KC} by the semi-analytical, semi-empirical Kozeny-Carman equation [16, 32]

k1KC=D2cKCϕ3(1−ϕ)2,(11)k_1^{KC} = \frac{D^2}{c_{KC}}\frac{\phi^3}{(1-\phi)^2}, \tag*{(11)}

where DD, cKCc_{KC} are the sphere diameter and the Kozeny-Carman constant for which the value cKC=180c_{KC} = 180 is set. In addition, the Stokes solution is further compared with results from a SPH solver that performs simulations based on weakly compressible Navier-Stokes equations [42].

Figure 6. Comparison of Stokes solver results of normalized permeabilities with Karman-Cozeny equation and SPH solver.

The results of the two solvers and also in comparison with the semi-analytical solution match almost perfectly. It can be concluded that the solver is capable for a wide range of porosities. A single simulation for a resolution of L/dx=100L/\mathrm{d}x = 100 takes in average 40 minutes on a desktop PC using 4 cores (11th Gen Intel(R) Core(TM) i7-11700KF @ 3.60GHz).

Permeability tensor and principal permabilities

To compute the anisotropic permeability behavior of various materials, we must not only consider standard benchmarks and those for simple porous materials but also include benchmarks with known or adjustable principal directions. For this purpose, we create a cube with 1003100^3 voxels corresponding to 1 mm31\,\mathrm{mm}^3 and place an ellipsoid in the center. It has the following semi-axes in Cartesian coordinates ei\mathbf{e}_i each given in absolute voxel numbers: e1:a=35\mathbf{e}_1: a = 35; e2:b=10\mathbf{e}_2: b = 10; e3:c=35\mathbf{e}_3: c = 35. By using this per se periodic structure, one does not have to consider the difficulties that would arise from symmetrical or translational periodization [23]. We rotate the single ellipsoid around e1\mathbf{e}_1-axis in 5∘5^\circ steps for 0∘≤α≤90∘0^\circ\leq\alpha\leq90^\circ and compute fluid flow and permeability, whereby the consideration of the permeability in e1\mathbf{e}_1-direction does not play a role in our evaluation for the time being. Thus the computation of the permeability tensor in Eq. (1) is reduced to a two-dimensional problem and eigenvalues as well as principal directions can be represented by an ellipse.

Each simulation of a domain described above allows us to determine one column of the coefficient matrix kijk_{ij}, or in our simplified case two entries. The other two entries are obtained from the simulation for which the direction of the imposed pressure gradient relative to the ellipsoid is rotated by 90°. Alternatively, instead of rotating the imposed pressure gradient the ellipsoid is rotated. For example, considering a rotation of α1=20∘\alpha^1 = 20^\circ, α2=−70∘\alpha^2 = -70^\circ follows. Due to symmetry, also α2=70∘\alpha^2 = 70^\circ can be set and we get the velocity patterns, illustrated in Figure 7, and compute the following coefficient matrix

Figure 7. Normalized magnitude of velocity around an ellipsoid with 20° (left) and 70° inclination (right). The same pressure gradient in e3e_3-direction is applied, which explains differences in the maximum velocities. Both figures show a cross-section (e2−e3e_2-e_3 plane) in the center of the domain.

kij=[4.541.761.968.71]⋅10−9 m2≈[4.541.861.868.71]⋅10−9 m2(12)k_{ij} = \begin{bmatrix} 4.54 & 1.76 \\ 1.96 & 8.71 \end{bmatrix} \cdot10^{-9}\,\mathrm{m}^2 \approx \begin{bmatrix} 4.54 & 1.86 \\ 1.86 & 8.71 \end{bmatrix} \cdot10^{-9}\,\mathrm{m}^2 \tag*{(12)}

and finally receive the coefficient matrix for the principle axes system

k~ij=[3.82009.42]⋅10−9 m2.(13)\widetilde{k}_{ij} = \begin{bmatrix} 3.82 & 0 \\ 0 & 9.42 \end{bmatrix} \cdot10^{-9}\,\mathrm{m}^2. \tag*{(13)}

Numerically determined, one obtains slightly different values for two corresponding secondary diagonal elements. These are therefore averaged k12=k21=12(k12+k21)k_{12}=k_{21}=\frac{1}{2}(k_{12}+k_{21}). The same procedure would adopted for k13k_{13}, k31k_{31} and k23k_{23}, k32k_{32}. In this reduced case, the angle φ\varphi between the axes in the original basis system and the principal directions is computed by

φ=12tan⁡−1(2k23k22+k33)=12tan⁡−1(2⋅1.86⋅10−9 m28.71⋅10−9 m2+4.54⋅10−9 m2)=0.364≈20.8∘(14)\varphi= \frac{1}{2}\tan^{-1}\left(\frac{2k_{23}}{k_{22}+k_{33}}\right)=\frac{1}{2}\tan^{-1}\left(\frac{2\cdot1.86\cdot10^{-9}\,\mathrm{m}^2}{8.71\cdot10^{-9}\,\mathrm{m}^2+4.54\cdot10^{-9}\,\mathrm{m}^2}\right)=0.364\approx20.8^\circ \tag*{(14)}

which matches very well to what was being specified. The listing of all values can be found in Table 1 and a selection is visualized in Figure 8.

α\alpha$1$051015202530354045
k33k_{33}[⋅10−9 m2][\cdot10^{-9}\,\mathrm{m}^2]9.439.339.198.998.718.367.977.517.126.63
k22k_{22}[⋅10−9 m2][\cdot10^{-9}\,\mathrm{m}^2]3.853.924.044.254.544.915.295.706.176.63
k23k_{23}[⋅10−9 m2][\cdot10^{-9}\,\mathrm{m}^2]4.570.831.141.501.862.212.472.592.802.83
kIIIk_{III}[⋅10−9 m2][\cdot10^{-9}\,\mathrm{m}^2]9.469.459.439.439.429.449.449.369.499.46
kIIk_{II}[⋅10−9 m2][\cdot10^{-9}\,\mathrm{m}^2]3.813.793.803.813.823.833.823.863.803.80
φ\varphi$1$4.68.511.916.120.826.030.835.440.145.0

Table 1. Permeabilities, principal permeabilities and principal directions for all simulations.

Figure 8. Plots of the coefficient matrices of the permeability tensors for rotations α={0∘,10∘,40∘}\alpha= \{0^\circ, 10^\circ, 40^\circ\}. Absolute values of two semi-axes of the ellipse represent the eigenvalues of the tensor. The angle of inclination (geometric information) and the computed angle of the eigenvector highlighted with a green and red line, respectively.

We observe a clearly recognizable trend for increase of the permeability k22k_{22} and decrease of k33k_{33} with increasing rotation α\alpha. Performing the transformation, we obtain the same eigenvalues for the permeability, except for minimal differences, which can be explained by the numerical aspects and issues related to discretization. The calculated angles for small α\alpha still exhibit noticeable deviations from the geometrically prescribed angles, with a maximum deviation of up to 4.6∘4.6^\circ. An error arises since the analysis for α1=0∘\alpha^1=0^\circ includes a simulation α2=90∘\alpha^2=90^\circ where the flow approaches the ellipsoid from its wide side. In Table 1 we can also see two outliers for the secondary diagonal elements for α=0∘\alpha=0^\circ and α=5∘\alpha=5^\circ. The larger the inclination, the more precisely the simulation result reproduces the angle presupposed in the real geometry. For α=40∘\alpha= 40^\circ the error corresponds to only 0.1∘0.1^\circ. For this comparably straightforward benchmark geometry, we can demonstrate the overall effectiveness of the workflow. It is worth highlighting that when the flow approaches the ellipsoid at α=90∘\alpha= 90^\circ, it leads to deviations in the secondary diagonal elements. This intriguingly results in enhanced solution accuracy when the system is maximally distant from one based on the principal directions.

Testing against other codes

Regular thin porous media (2D microfluidic devices): In Sec. 7.3 we investigate structures that have a small thickness compared to their lateral dimensions. The dimensions of such thin (”2D”) porous materials, which are often used in microfluidic characterization of porous media flow, pose problems for numerical methods in terms of efficiency and convergence. To ensure the applicability of the solver for such requirements, we compare it with benchmarks for regular 2D porous materials described in [57]. The dimension of the simulated unit cell is 1 mm×0.091 mm×1 mm1\ \mathrm{mm} \times0.091\ \mathrm{mm} \times1\ \mathrm{mm} and is discretized by 250×27×250250 \times27 \times250 voxels, with a solid frame of two voxels thick each, included in the e2\mathbf{e}_2 direction. All results align with the solutions provided in the benchmark paper.

Radiusϕ\phik11k_{11} range from Wagner et al. [57]k11k_{11} (POREMAPS)
0.35 mm0.6222.9−26.7⋅10−10 m222.9 - 26.7 \cdot10^{-10}\ \mathrm{m}^226.1⋅10−10 m226.1 \cdot10^{-10}\ \mathrm{m}^2
0.40 mm0.5016.4−17.7⋅10−10 m216.4 - 17.7 \cdot10^{-10}\ \mathrm{m}^217.5⋅10−10 m217.5 \cdot10^{-10}\ \mathrm{m}^2
0.45 mm0.367.54−8.59⋅10−10 m27.54 - 8.59 \cdot10^{-10}\ \mathrm{m}^28.08⋅10−10 m28.08 \cdot10^{-10}\ \mathrm{m}^2
0.47 mm0.313.62−5.35⋅10−10 m23.62 - 5.35 \cdot10^{-10}\ \mathrm{m}^23.83⋅10−10 m23.83 \cdot10^{-10}\ \mathrm{m}^2
0.49 mm0.250.46−0.54⋅10−10 m20.46 - 0.54 \cdot10^{-10}\ \mathrm{m}^20.39⋅10−10 m20.39 \cdot10^{-10}\ \mathrm{m}^2

Table 2. Comparison with different types of pore scale solvers (SPH, FEM, LBM from Wagner et al. [57]) for thin porous media samples and comparison of computed permeabilities.

To illustrate the performance of the solver, the SPH solver used in Wagner et al. [57] needs approximately 17 hours on a 4-core CPU, the presented Stokes solver, at the same resolution and on same hardware takes about 10 minutes.

Irregular sphere packings and porous rock: There are several benchmark papers providing suitable geometries ([3, 2, 50]). The solver is compared with different 3D benchmarks where permeabilities are computed with, among others, different LBM solvers ([50]). We compute the sphere packing, one Berea sandstone (Rock1), and one Fontainebleau sandstone (Rock3) sample. The characteristics of the geometries and a comparison of the computed permeabilities k11k_{11} are summarized in Table 3. The results for k11k_{11} are in accordance with the solutions determined by LBM and thus we consider the benchmarking to be successful and completed.

Sampleϕ\phiVoxel sizeDomain sizek11k_{11} [50]k11k_{11} (POREMAPS)
Sphere packing0.347.0 μm7.0\ \mathrm{\mu m}788×791×793788 \times791 \times7932.438−2.903⋅10−10 m22.438 - 2.903 \cdot10^{-10}\ \mathrm{m}^22.512⋅10−10 m22.512 \cdot10^{-10}\ \mathrm{m}^2
Berea0.182.114 μm2.114\ \mathrm{\mu m}1024×1024×10241024 \times1024 \times10244.569−6.889⋅10−13 m24.569 - 6.889 \cdot10^{-13}\ \mathrm{m}^25.772⋅10−13 m25.772 \cdot10^{-13}\ \mathrm{m}^2
Fontainebleau0.092.072 μm2.072\ \mathrm{\mu m}1024×1024×10241024 \times1024 \times10240.642−1.411⋅10−13 m20.642 - 1.411 \cdot10^{-13}\ \mathrm{m}^29.200⋅10−14 m29.200 \cdot10^{-14}\ \mathrm{m}^2

Table 3. Properties for different types of 3D benchmarks (range of LBM results from Saxena et al. [50]) and comparison of computed permeabilities.

Applications

In this section, three different application scenarios are shown to demonstrate the capabilities of POREMAPS. The corresponding domain sizes and computational times are summarized in Table 4.

Domain sizeHardware and resourcesComputation time
Example 1 “Berea” and “sphere packing” (§ 6.5 and § 7.1)
2048×2048×20482048 \times2048 \times2048SimTech Cluster, 2 Nodes36 hrs.
1576×1582×15861576 \times1582 \times1586SimTech Cluster, 2 Nodes24 hrs.
3152×3164×31723152 \times3164 \times3172SimTech Cluster, 4 Nodes≈80\approx80 hrs.
Example 2 “Open-cell foam (ε33=0.0\varepsilon_{33}=0.0 and ε33=−0.5\varepsilon_{33}=-0.5)” (§ 7.2)
800×800×1186800 \times800 \times1186SimTech Cluster, 4 Nodes8 hrs.
800×800×526800 \times800 \times526SimTech Cluster, 4 Nodes3 hrs.
Example 3 “Calcite precipitation” (Sec. 7.3)
12×1200×120012 \times1200 \times1200SimTech Cluster, 1 Node0.5–2 hrs.

Table 4. Overview of computation times of considered application examples.

Characterization of large 3D images

The solver is primarily designed to analyze large, high-resolved 3D images of porous materials typically acquired by 3D imaging methods like µXRCT (Stock, [53]; Ruf & Steeb, [46]; Withers et al., [65]). This allows the digital characterization of porous materials with regard to hydraulic permeability and can be integrated into the imaging workflow as a subsequent standard procedure. In µXRCT imaging, resolution of REVs is often given by 3D images having more than 100031000^{3} voxel. High-resolution X-ray flat panel detectors having around 3000×30003000 \times3000 pixel and more are not uncommon anymore. This means in terms of the reconstructed 3D images, that they can have ≈30003\approx3000^{3} voxel. With appropriate hardware (large memory), the presented code is capable to handle such big 3D images.

To demonstrate this, we use the 3D image of the Berea sandstone (1024×1024×10241024 \times1024 \times1024 voxel), and the sphere packing (788×791×793788 \times791 \times793 voxel) from Saxena et al. [50], already considered for the benchmark tests, cf. Table 3. Both data sets are mirrored in all three spatial directions resulting in 2048×2048×20482048 \times2048 \times2048 voxel and 1576×1582×15861576 \times1582 \times1586 voxel. The corresponding computation times are given in Table 4 for a convergence criteria of ε=∣Δq∣∣q∣<10−6\varepsilon= \frac{|\Delta q|}{|q|} < 10^{-6}. Latter data set is mirrored a second time in each direction resulting in 3152×3164×31723152 \times3164 \times3172 voxel (equals 64 times the original geometry), cf. Figure 9. Due to parallelized IO routines, this can be implemented using 4 nodes, with a total of 512 cores, requiring a total of ≈1.5\approx1.5 TB RAM.

Illustration of a double-mirrored 3D image of a sphere packing

Figure 9. Illustration of the large sized computable domains using the example of a double-mirrored 3D image of a sphere packing (blue) with originally 788×791×793788 \times791 \times793 voxel (yellow). The original 3D image (yellow) is taken from Saxena et al. [50].

Deformation-dependent permeability and permeability anisotropy – uniaxial compression of an open-cell foam

The deformation-dependent permeability of open-cell foams has been studied experimentally by several researches. Different theoretical models were proposed to predict the effect of strain on permeability, cf. Dawson, Germaine, and Gibson [20]; Markert [38] and therein cited literature. In experimental studies, often uniaxial compression loading is imposed on a foam sample and the deformation-depending permeability is measured in one direction, often in the same direction as the imposed load. Measuring the permeability in different directions is technically challenging. Using non-destructive μXRCT imaging is a possible approach to overcome this problem when the fluid-solid interaction is negligible and the foam can be considered as rigid at a given deformation state. In this case, the foam structure is imaged under different loading conditions in 3D. Based on the time-series of 3D images followed by a subsequent post-processing provides the possibility to perform virtual experiments and to characterize the structure and the deformation-dependency in more detail.

This is exemplary shown for an open-cell Polyurethane (PUR) cylindrical foam sample with 10 PPI subjected to uniaxial compressive loading and imaged at discrete loading states (engineering compressive strain ε33=−{0.0,0.1,0.2,0.3,0.4,0.5}\varepsilon_{33} = -\{0.0, 0.1, 0.2, 0.3, 0.4, 0.5\}), see Figure 10(a). For the imaging, the system presented in Ruf and Steeb [46]; Ruf, Lee, and Steeb (2023) was employed. For the series of 3D images, the permeability tensor is determined using cubic subvolumes of size 800×800×526800 \times800 \times526–11861186 voxel with a uniform voxel size of 74.874.8 μm. The permeability tensor (three simulations) for each loading condition is determined in ≈3\approx3-8 hrs.. The permeability in axial direction (e₃) and radial directions (e₁, e₂) and the principle values are shown in Figure 10(b) over the applied strain (left) and the foam porosity (right). At all deformation states, the radial permeabilities k11k_{11} and k22k_{22} are quite similar and differ slightly from the axial permeability k33k_{33} of the cylindrical sample. In general, it can be said that the permeability anisotropy does not correlate in a systematic way, neither with the strain nor with the porosity, which contradicts the expectation.

Deformation-dependent permeability and anisotropy of an open-cell PUR foam sample

Figure 10. Deformation-dependent permeability and anisotropy of an open-cell PUR foam sample (10 PPI) based on μXRCT images (Ruf, Steeb, & Karadimitriou, 2023; Ruf & Steeb, [47]). (a) Segmented μXRCT images for different uniaxial compression loading states. (b) Determined permeabilities in e₁-, e₂-, e₃-direction and principle permeabilities.

Anisotropy development during calcite precipitation

Predicting pore-scale clogging phenomena in heterogeneous porous materials presents a significant challenge. These processes can occur inadvertently, and if they cannot be prevented, there is a need to manage them, such as in the case of clogged filters. Conversely, these processes might also be intentional, such as in the case of blocking subsurface cracks. We focus our investigation on a specific scenario where pore spaces gradually become obstructed due to a chemically induced precipitation process. The experimental data set, Weinhardt, Deng, Steeb, and Class (2022), serves as the basis for our study, and the details of data acquisition and the experimental arrangement are elaborated upon in Weinhardt et al. [59]; Weinhardt (2022); Weinhardt, Deng, Hommel, et al. [60]. We have already described this procedure of investigation on a smaller extent as a proof of concept in Krach and Steeb [34].

Unlike the experimental procedure, we can determine the permeability k33k_{33} in the e3e_3-direction (aligned with the pressure gradient in the experiment) and the permeability k22k_{22} in the e2e_2-direction (perpendicular to the pressure gradient) at various time steps during the experiment. This investigation includes 137 individual simulations, where for 57 time steps the domain is percolating in both directions, resulting in 2 simulations per time step. In e3e_3-direction a flow path remains unblocked for a longer time and 80 time steps are investigated. A simulation (12×1200×120012 \times1200 \times1200 voxel) typically runs for an average of 2 hours on a node using 121 CPU-cores. We explicitly fix the domain decomposition to 1×11×111 \times11 \times11 ranks.

2D image data is used as basis and replicated in the third spatial direction to simulate the original dimensions of the microfluidic experiment in 3D. We are particularly interested in a preferential flow path that forms in the upper region of the domain and stays almost free during the experiment. Accordingly a subdomain is extracted. In the course of the experiment, more and more precipitate is accumulated in the domain, thus increasing the solid fraction and decreasing porosity and permeability. Figure 11 shows the geometries (black - solid body, gray - precipitate) and the velocity patterns. Note that we simulate the same domain twice with pressure gradient from left to right (e3e_3-direction: Figure 11, center column) and pressure gradient from top to bottom (e2e_2-direction: Figure 11, right column). The simulations include all time steps of the experiment up to the point where there is no flow path through the porous material at all. The material clogs first in the e2e_2-direction, at a porosity of ϕ=27.7 %\phi= 27.7\ \%, while at this time a flow path remains open in the e3e_3-direction.

Figure 11. Geometry and flow patterns at different times (increasing from top to bottom) during the experiment. The left column represents the input domain (black - solid columns, gray - precipitate). For the simulations, the precipitates are attributed to the solid. Column 2 and 3 show the normalized absolute velocities based on pressure gradients in e3e_3- and e2e_2-direction.

The focus of the analysis lies on the development of the degree of anisotropy in the course of the experiment. For this purpose, the eigenvalues of the 2×22 \times2 permeability tensor are determined. Figure 12 shows the eigenvalues kIIk_{II} and kIIIk_{III} plotted against the porosity ϕ\phi. The principal permeability kIIk_{II} decreases significantly faster than kIIIk_{III}. The degree of anisotropy, the ratio of the eigenvalues kIIIkII\frac{k_{III}}{k_{II}} (Figure 12, right) increases stepwise linear in the course of the experiment up to a ration of 6. Thus, this analysis provides an additional benefit that experiments cannot provide.

Figure 12. Normalized eigenvalues of the permeability tensor (kIIk_{II}, kIIIk_{III}) over the course of a decreasing porosity (left) and the relation of both (kIII/kIIk_{III}/k_{II}) illustrating a clear development of anisotropic hydraulic properties (right).

Conclusion

A FDM solver, POREMAPS, for the incompressible Stokes equations using the artificially compressible method was presented to compute the permeability tensor k\mathbf{k} of various porous materials. Different benchmarks have been performed, which demonstrate that the solver can handle a variety of problems for a range of porosities swiftly and robustly. We present two compelling applications that perfectly align with the capabilities of the solver, demonstrating its versatility in tackling complex scenarios that demand extensive simulations. The first application focuses on open-cell foams, where we establish that any anisotropy in permeability is unlikely to develop during deformation up to ε33=−0.5\varepsilon_{33} = -0.5. Our findings underscore the absence of permeability directionality, what we would have expected otherwise. The second application addresses plugging porous materials, characterized by a distinct preferred orientation. In this context, we not only quantify the resulting anisotropy but also offer insights that highlight the solvers key advantages and experimental constraints.

While acknowledging the computational time and hardware prerequisites, it’s worth noting that they remain modest when contrasted with alternative methodologies. The solver’s user-friendly nature, coupled with its direct compatibility with experimental imaging data, opens up promising avenues in various domains.

A Nondimensional balance equations:

Acknowledgement

D.K. and H.S. acknowledge funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2075 – 390740016. M.R. and H.S. acknowledge funding from the DFG through the project STE 969/13-1 (Project No. 357361983) and supporting this work within the SFB 1313 (Project No. 327154368). We acknowledge the support by the Stuttgart Center for Simulation Science (SC SimTech).

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Author Contributions

D. Krach: Conceptualization, Methodology, Investigation, Software, Validation, Data Curation, Writing - Original Draft, Visualization, Project administration. M. Ruf: Investigation, Software, Data Curation, Writing - Original Draft, Writing - Review & Editing, Visualization. H. Steeb: Conceptualization, Writing - Review & Editing, Supervision, Funding acquisition.

Data, Code & Protocol Availability

The complete source code including input data of the shown application examples that support the findings of this study is openly available in the Data Repository of the University of Stuttgart (DaRUS) at https://doi.org/10.18419/darus-3676 ([33]).

Nondimensional quantities and operators

We introduce the dimensionless length l∗l^*, velocity v∗\mathbf{v}^*, and time t∗t^* by the ratio of physical property divided by characteristic property (L,V,T=L/V)(\mathcal{L}, \mathcal{V}, T=\mathcal{L}/\mathcal{V})

l∗=lL → l=l∗L,v∗=vV → v=v∗Vandt∗=tT → t=t∗LV.(15)l^*=\frac{l}{\mathcal{L}}\ \to\ l=l^*\mathcal{L},\qquad\mathbf{v}^*=\frac{\mathbf{v}}{\mathcal{V}}\ \to\ \mathbf{v}=\mathbf{v}^*\mathcal{V}\qquad\text{and}\qquad t^*=\frac{t}{T}\ \to\ t=\frac{t^*\mathcal{L}}{\mathcal{V}}. \tag*{(15)}

The nondimensional differential operators grad⁡(∙)\operatorname{grad}(\bullet) and div⁡(∙)\operatorname{div}(\bullet) and time derivative (∙)⋅(\bullet)^{\boldsymbol{\cdot}} are given as follows:

grad⁡∗(∙)=Lgrad⁡(∙) → grad⁡(∙)=1Lgrad⁡∗(∙),div⁡∗(∙)=Ldiv⁡(∙) → div⁡(∙)=1Ldiv⁡∗(∙)(16)\operatorname{grad}^*(\bullet)=\mathcal{L}\operatorname{grad}(\bullet)\ \to\ \operatorname{grad}(\bullet)=\frac{1}{\mathcal{L}}\operatorname{grad}^*(\bullet),\qquad\operatorname{div}^*(\bullet)=\mathcal{L}\operatorname{div}(\bullet)\ \to\ \operatorname{div}(\bullet)=\frac{1}{\mathcal{L}}\operatorname{div}^*(\bullet) \tag*{(16)}
(∙)⋅∗=T(∙)⋅=LV(∙)⋅ → (∙)⋅=VL(∙)⋅∗.(\bullet)^{\boldsymbol{\cdot}*}=T(\bullet)^{\boldsymbol{\cdot}}=\frac{\mathcal{L}}{\mathcal{V}}(\bullet)^{\boldsymbol{\cdot}}\ \to\ (\bullet)^{\boldsymbol{\cdot}}=\frac{\mathcal{V}}{\mathcal{L}}(\bullet)^{\boldsymbol{\cdot}*}.

This allows for the selection of the pressure and density scale

ρ∗=ρρ0 → ρ=ρ0ρ∗andp∗=pρ0V2 → p=p∗ρ0V2,(17)\rho^*=\frac{\rho}{\rho_0}\ \to\ \rho=\rho_0\rho^*\qquad\text{and}\qquad p^*=\frac{p}{\rho_0\mathcal{V}^2}\ \to\ p=p^*\rho_0\mathcal{V}^2, \tag*{(17)}

where ρ0\rho_0 is the rest density of the fluid.

Nondimensional governing equations

For the mutual dependence of pressure pp and density ρ\rho we employ the constitutive equation and its derivative

ρ=ρ0[pKF+1] → ρ˙=ρ0p˙KF.(18)\rho=\rho_0\left[\frac{p}{KF}+1\right]\ \to\ \dot{\rho}=\rho_0\frac{\dot{p}}{KF}. \tag*{(18)}

Artificial compressibility for pseudo-unsteady Stokes equations is introduced by considering the linearized balance of mass (Eq. (5)) for a compressible fluid

ρ˙+ρ0div⁡v=0,(19)\dot{\rho}+\rho_0\operatorname{div}\mathbf{v}=0, \tag*{(19)}

where for steady state ρ˙→0\dot{\rho}\to0 holds. The Stokes equations (Eq. (4)) are analogously extended

ρv˙=μdiv⁡(grad⁡v)−grad⁡p(20)\rho\dot{\mathbf{v}}=\mu\operatorname{div}(\operatorname{grad}\mathbf{v})-\operatorname{grad}p \tag*{(20)}

where ρv˙→0\rho\dot{\mathbf{v}}\to0 holds for steady state. We apply the dimensionless divergence operator, dimensionless velocity and density to (19) and receive

VL(ρ0ρ∗)⋅∗+ρ0Ldiv⁡(v∗V)=0 → (ρ∗)⋅∗+div⁡(v∗)=0.(21)\frac{\mathcal{V}}{\mathcal{L}}(\rho_0\rho^*)^{\boldsymbol{\cdot}*}+\frac{\rho_0}{\mathcal{L}}\operatorname{div}(\mathbf{v}^*\mathcal{V})=0\ \to\ (\rho^*)^{\boldsymbol{\cdot}*}+\operatorname{div}(\mathbf{v}^*)=0. \tag*{(21)}

Using the speed of sound for a compression-type wave c=KFρ0c=\sqrt{\frac{K F}{\rho_0}} with its dimensionless form c=c∗Vc=c^{*}V the constitutive equation for the pressure (18) is transformed in the same way, yielding

ρ˙=ρ0p˙KF→ρ˙=p˙c2→VL(ρ0∗ρ∗)⋅=V(p∗ρ0V2)⋅Lc∗2V2→(ρ∗)⋅=(p∗)⋅c∗2(22)\dot{\rho}=\rho_0\frac{\dot{p}}{K F}\quad\to\quad\dot{\rho}=\frac{\dot{p}}{c^2}\quad\to\quad\frac{V}{\mathcal{L}}(\rho_0^{*}\rho^{*})^{\cdot}=\frac{V(p^{*}\rho_0V^2)^{\cdot}}{\mathcal{L}c^{*2}V^2}\quad\to\quad(\rho^{*})^{\cdot}=\frac{(p^{*})^{\cdot}}{c^{*2}} \tag*{(22)}

Plugging (22) in (21) the nondimensional balance of mass is obtained

(ρ∗)⋅=−c∗2div⁡(v∗).(23)(\rho^{*})^{\cdot}=-c^{*2}\operatorname{div}(\mathbf{v}^{*}). \tag*{(23)}

The nondimensional form of the pseudo-unsteady Stokes problem ((20)) yields

ρ0ρ∗V2L(v∗⋅∇∗)v∗=μL2div⁡∗grad⁡(v∗V)−1Lgrad⁡∗(p∗ρ0V2)→ρ∗(v∗⋅∇∗)v∗=1Rediv⁡∗grad⁡(v∗)−grad⁡∗(p∗).(24)\frac{\rho_0\rho^{*}V^2}{\mathcal{L}}(\mathbf{v}^{*}\cdot\nabla^{*})\mathbf{v}^{*}=\frac{\mu}{\mathcal{L}^2}\operatorname{div}^{*}\operatorname{grad}(\mathbf{v}^{*}V)-\frac{1}{\mathcal{L}}\operatorname{grad}^{*}(p^{*}\rho_0V^2)\quad\to\quad\rho^{*}(\mathbf{v}^{*}\cdot\nabla^{*})\mathbf{v}^{*}=\frac{1}{\mathrm{Re}}\operatorname{div}^{*}\operatorname{grad}(\mathbf{v}^{*})-\operatorname{grad}^{*}(p^{*}). \tag*{(24)}

Implementation of fluid-solid boundary conditions

The Taylor expansion around the point x0x_0 is generally given by the following series:

f(x)=∑n=0∞f(n)(x0)n!(x−x0)n=f(0)+f(1)(x−x0)1+12f(2)(x−x0)2+⋯ .(25)f(x)=\sum_{n=0}^{\infty}\frac{f^{(n)}(x_0)}{n!}(x-x_0)^n=f^{(0)}+f^{(1)}(x-x_0)^1+\frac{1}{2}f^{(2)}(x-x_0)^2+\cdots. \tag*{(25)}

As briefly described in Sec. 4 for case 2 depicted in Figure 1, for the other given cases following system of equations can be derived. By solving the given sets of equation systems, the corresponding second-order derivatives (∂2vk∂xi2)\left(\frac{\partial^2v_k}{\partial x_i^2}\right) in the perpendicular directions to the considered flow velocity direction for i,j,k=(1,2,3),(2,3,1),(3,1,2)i,j,k=(1,2,3),(2,3,1),(3,1,2) can be determined.

Case 1:

\begin{aligned} v_{k1}&\approx v_{k2}-\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0\\ v_{k3}&\approx v_{k2}+\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0 \end{aligned} \quad\longrightarrow\quad \frac{\partial^2v_k}{\partial x_i^2}=-8v_{k2} \tag*{(26)}\end{aligned}

Case 2:

\begin{aligned} v_{k1}&\approx v_{k2}-\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0\\ v_{k3}&\approx v_{k2}+\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k4}&\approx v_{k2}+\frac{3}{2}\frac{\partial v_k}{\partial x_i}+\frac{9}{8}\frac{\partial^2v_k}{\partial x_i^2}+\frac{9}{16}\frac{\partial^3v_k}{\partial x_i^3}=0 \end{aligned} \quad\longrightarrow\quad \frac{\partial^2v_k}{\partial x_i^2}=\frac{8}{3}v_{k3}-\frac{16}{3}v_{k2} \tag*{(27)}\end{aligned}

Case 3:

\begin{aligned} v_{k1}&\approx v_{k3}-\frac{3}{2}\frac{\partial v_k}{\partial x_i}+\frac{9}{8}\frac{\partial^2v_k}{\partial x_i^2}-\frac{9}{16}\frac{\partial^3v_k}{\partial x_i^3}=0\\ v_{k2}&\approx v_{k3}-\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k4}&\approx v_{k3}+\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0 \end{aligned} \quad\longrightarrow\quad \frac{\partial^2v_k}{\partial x_i^2}=\frac{8}{3}v_{k2}-\frac{16}{3}v_{k3} \tag*{(28)}\end{aligned}

Case 4:

\begin{aligned} v_{k1}&\approx v_{k2}-\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0\\ v_{k3}&\approx v_{k2}+\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k4}&\approx v_{k2}+2\frac{\partial v_k}{\partial x_i}+2\frac{\partial^2v_k}{\partial x_i^2}+\frac{8}{6}\frac{\partial^3v_k}{\partial x_i^3} \end{aligned} \quad\longrightarrow\quad \frac{\partial^2v_k}{\partial x_i^2}=2v_{k3}-5v_{k2}-\frac{1}{5}v_{k4} \tag*{(29)}\end{aligned}

Case 5:

\begin{aligned} v_{k1}&\approx v_{k3}-2\frac{\partial v_k}{\partial x_i}+2\frac{\partial^2v_k}{\partial x_i^2}-\frac{8}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k2}&\approx v_{k3}-\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k4}&\approx v_{k3}+\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{48}\frac{\partial^3v_k}{\partial x_i^3}=0 \end{aligned} \quad\longrightarrow\quad \frac{\partial^2v_k}{\partial x_i^2}=2v_{k2}-\frac{1}{5}v_{k1}-5v_{k3} \tag*{(30)}\end{aligned}

Case 6:

\begin{aligned} v_{k1} &\approx v_{k2}-\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}-\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3}\\ v_{k3} &\approx v_{k2}+\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}+\frac{1}{6}\frac{\partial^3v_k}{\partial x_i^3} \end{aligned} \quad\} \sim\frac{\partial^2v_k}{\partial x_i^2}=v_{k1}-2v_{k2}+v_{k3} \tag*{(31)}\end{aligned}

Case 7 (4b):

\begin{aligned} v_{k1} &\approx v_{k2}-\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}=0\\ v_{k3} &\approx v_{k2}+\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2} \end{aligned} \quad\} \sim\frac{\partial^2v_k}{\partial x_i^2}=\frac{4}{3}v_{k3}-4v_{k2} \tag*{(32)}\end{aligned}

Case 8 (5b):

\begin{aligned} v_{k1} &\approx v_{k2}-\frac{\partial v_k}{\partial x_i}+\frac{1}{2}\frac{\partial^2v_k}{\partial x_i^2}\\ v_{k3} &\approx v_{k2}+\frac{1}{2}\frac{\partial v_k}{\partial x_i}+\frac{1}{8}\frac{\partial^2v_k}{\partial x_i^2}=0 \end{aligned} \quad\} \sim\frac{\partial^2v_k}{\partial x_i^2}=\frac{4}{3}v_{k1}-4v_{k2} \tag*{(33)}\end{aligned}

Permeability computation

For Re⁡<1.0\operatorname{Re}<1.0 the entries of the permeability tensor can be computed using Darcy’s law [19]

q=1μk⋅h→qiei=1μkij(ei⊗ej)⋅hkek=1μkikhkei(34)\mathbf{q}=\frac{1}{\mu}\mathbf{k}\cdot\mathbf{h}\quad\to\quad q_i e_i=\frac{1}{\mu}k_{ij}(e_i\otimes e_j)\cdot h_k e_k=\frac{1}{\mu}k_{ik}h_k e_i \tag*{(34)}

where hi=−∂p∂xi=−p,ih_i=-\frac{\partial p}{\partial x_i}=-p_{,i} is the hydraulic gradient. Three numerical experiments must be performed to determine the nine entries in the coefficient matrix of the second order permeability tensor. Thereby, the following pressure gradients are applied for the different cases

a) h1≠0∧h2=h3=0;b) h2≠0∧h1=h3=0;c) h3≠0∧h1=h2=0;(35)a)\ h_1\ne0\land h_2=h_3=0;\qquad b)\ h_2\ne0\land h_1=h_3=0;\qquad c)\ h_3\ne0\land h_1=h_2=0; \tag*{(35)}

and we measure three fluxes qiq_i for each case. By comparing the coefficients for the three equations for the flow qiq_i (Eq. (34), right) we receive

qi=−1μkikhk→i)q1=−1μ[k11h1+k12h2+k13h3]→ii)q2=−1μ[k21h1+k22h2+k23h3]→iii)q3=−1μ[k31h1+k32h2+k33h3](36)\begin{aligned} q_i&=-\frac{1}{\mu}k_{ik}h_k\quad\to\quad i)\quad q_1=-\frac{1}{\mu}[k_{11}h_1+k_{12}h_2+k_{13}h_3]\\ &\to\quad ii)\quad q_2=-\frac{1}{\mu}[k_{21}h_1+k_{22}h_2+k_{23}h_3]\\ &\to\quad iii)\quad q_3=-\frac{1}{\mu}[k_{31}h_1+k_{32}h_2+k_{33}h_3] \tag*{(36)} \end{aligned}

wherefrom arise the following nine equations by means of the constraints a)a) - c)c) given in Eq. (35):

a)q1=−1μk11h1,q2=−1μk21h1,q3=−1μk31h1,b)q1=−1μk12h2,q2=−1μk22h2,q3=−1μk32h2,c)q1=−1μk13h3,q2=−1μk23h3,q3=−1μk33h3.(37)\begin{aligned} a)\quad q_1&=-\frac{1}{\mu}k_{11}h_1,\qquad q_2=-\frac{1}{\mu}k_{21}h_1,\qquad q_3=-\frac{1}{\mu}k_{31}h_1,\\ b)\quad q_1&=-\frac{1}{\mu}k_{12}h_2,\qquad q_2=-\frac{1}{\mu}k_{22}h_2,\qquad q_3=-\frac{1}{\mu}k_{32}h_2,\\ c)\quad q_1&=-\frac{1}{\mu}k_{13}h_3,\qquad q_2=-\frac{1}{\mu}k_{23}h_3,\qquad q_3=-\frac{1}{\mu}k_{33}h_3. \tag*{(37)} \end{aligned}

It results in nine equations for nine unknown entries of the coefficient matrix of the second order permeability tensor. Numerically determined, one obtains slightly different values for two corresponding secondary diagonal elements. These are therefore averaged k12=k21=12(k12+k21)k_{12}=k_{21}=\frac{1}{2}(k_{12}+k_{21}). The same procedure is adopted for k13k_{13}, k31k_{31} and k23k_{23}, k32k_{32}.

References

  1. [1]Adler, P., Jacquin, C., & Quiblier, J. (1990). Flow in simulated porous media. *International Journal of Multiphase Flow, 16*(4), 691-712. doi: 10.1016/0301-9322(90)90025-EDOI
  2. [2]Andrä, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., . . . Zhan, X. (2013a). Digital rock physics benchmarks—Part II: Computing effective properties. *Computers & Geosciences, 50*, 33-43. doi: 10.1016/j.cageo.2012.09.008DOI
  3. [3]Andrä, H., Combaret, N., Dvorkin, J., Glatt, E., Han, J., Kabel, M., . . . Zhan, X. (2013b). Digital rock physics benchmarks—Part I: Imaging and segmentation. *Computers & Geosciences, 50*, 25-32. doi: 10.1016/j.cageo.2012.09.005DOI
  4. [4]Aziz, K., & Wong, T. (1989). Considerations in the development of multipurpose reservoir simulation models. In First and Second Forum on Reservoir Simulation (pp. 77–208).
  5. [5]Bakhshian, S., & Sahimi, M. (2016). Computer simulation of the effect of deformation on the morphology and flow properties of porous media. Physical Review E, 94, 042903. doi: 10.1103/PhysRevE.94.042903DOI
  6. [6]Batchelor, G. K. (1967). An Introduction to Fluid Dynamics. Cambridge University Press.
  7. [7]Bear, J. (1988). Dynamics of Fluids in Porous Media. Dover.
  8. [8]Bear, J., & Verruijt, A. (1987). Modeling Groundwater Flow and Pollution. Springer Science & Business Media.DOI
  9. [9]Bentz, D., & Martys, N. (2007). A Stokes Permeability Solver for Three-Dimensional Porous Media (Techreport No. NISTIR 7416).DOI
  10. [10]Bernard, D., Övgrind Nielsen, S., Sylvain, L., & Cloetens, P. (2005). Permeability assessment by 3D interdendritic flow simulations on microtomography mappings of Al–Cu alloys. Materials Science and Engineering: A, 392(1), 112-120. doi: 10.1016/j.msea.2004.09.004
  11. [11]Blunt, M. J. (2017). Multiphase Flow in Permeable Media: A Pore-Scale Perspective. Cambridge University Press.
  12. [12]Blunt, M. J., Bijeljic, B., Dong, H., Gharbi, O., Iglauer, S., Mostaghimi, P., … Pentland, C. (2013). Pore-scale imaging and modelling. Advances in Water Resources, 51, 197-216. doi: 10.1016/j.advwatres.2012.03.003DOI
  13. [13]Borujeni, A. T., Lane, N., Thompson, K., & Tyagi, M. (2013). Effects of image resolution and numerical resolution on computed permeability of consolidated packing using LB and FEM pore-scale simulations. Computers & Fluids, 88, 753-763. doi: 10.1016/j.compfluid.2013.05.019
  14. [14]Burger, W., & Burge, M. J. (2016). Digital Image Processing. London: Springer. doi: 10.1007/978-1-4471-6684-9
  15. [15]Cantisano, M. T., Restrepo, D. P., Cespedes, S., Toelke, J., Grader, A., Suhrer, M., & Walls, J. (2013). Relative Permeability in a Shale Formation in Colombia Using Digital Rock Physics (Vol. All Days). doi: 10.1190/urtec2013-092
  16. [16]Carman, P. C. (1997). Fluid flow through granular beds. Chemical Engineering Research and Design, 75, S32–S48. doi: 10.1016/S0263-8762(97)80003-2DOI
  17. [17]Chareyre, B., Cortis, A., Catalano, E., & Barthélémy, E. (2012). Pore-scale modeling of viscous flow and induced forces in dense sphere packings. Transport in Porous Media, 94, 595–615. doi: 10.1007/s11242-012-0057-2
  18. [18]Chorin, A. J. (1967). A numerical method for solving incompressible viscous flow problems. Journal of Computational Physics, 2(1), 12–26. doi: 10.1016/0021-9991(67)90037-xDOI
  19. [19]Darcy, H. (1856). Les fontaines publiques de la ville de Dijon: Exposition et application des principes à suivre et des formules à employer dans les questions de distribution d’eau. Paris: Victor Dalmont.
  20. [20]Dawson, M. A., Germaine, J. T., & Gibson, L. J. (2007). Permeability of open-cell foams under compressive strain. International Journal of Solids and Structures, 44(16), 5133–5145. doi: 10.1016/j.ijsolstr.2006.12.025DOI
  21. [21]Dong, H., & Blunt, M. J. (2009). Pore-network extraction from micro-computerized-tomography images. Physical Review E, 80, 036307. doi: 10.1103/PhysRevE.80.036307
  22. [22]Gerke, K. M., Vasilyev, R. V., Khriveich, S., Collins, D., Karsanina, M. V., Sazonenko, T. O., … Mallants, D. (2018). Finite-difference method Stokes solver (FDMSS) for 3D pore geometries: Software development, validation and case studies. Computers & Geosciences, 114, 41–58. doi: 10.1016/j.cageo.2018.01.005
  23. [23]Guibert, R., Horgue, P., Debenest, G., & Quintard, M. (2016). A comparison of various methods for the numerical evaluation of porous media permeability tensors from pore-scale geometry. Mathematical Geosciences, 48, 329–347. doi: 10.1007/s11004-015-9587-9
  24. [24]Harlow, F. H., & Welch, J. E. (1965). Numerical calculation of time-dependent viscous incompressible flow of fluid with free surface. The Physics of Fluids, 8(12), 2182–2189.DOI
  25. [25]Helmig, R. (1997). Multiphase Flow and Transport Processes in the Subsurface: A Contribution to the Modeling of Hydrosystems. Springer.DOI
  26. [26]Hilliard, Z., Evans, T. M., & Peszynska, M. (2024). Modeling flow and deformation in porous media from pore-scale to the Darcy-scale. Results in Applied Mathematics, 22, 100448. doi: 10.1016/j.rinam.2024.100448
  27. [27]Holmes, D. W., Williams, J. R., Tilke, P., & Leonardi, C. R. (2016). Characterizing flow in oil reservoir rock using SPH: absolute permeability. Computational Particle Mechanics, 3, 141–154. doi: 10.1007/s40571-015-0038-7DOI
  28. [28]Hommel, J., Gehring, L., Weinhardt, F., Ruf, M., & Steeb, H. (2022). Effects of enzymatically induced carbonate precipitation on capillary pressure-saturation relations. Minerals, 12(10). doi: 10.3390/min12101186DOI
  29. [29]Iassonov, P., Gebrenegus, T., & Tuller, M. (2009). Segmentation of X-ray computed tomography images of porous materials: A crucial step for characterization and quantitative analysis of pore structures. Water Resources Research, 45(9), W09415. doi: 10.1029/2009wr008087DOI
  30. [31]Jiang, F., & Tsuji, T. (2014). Changes in pore geometry and relative permeability caused by carbonate precipitation in porous media. Physical Review E, 90, 053306. doi: 10.1103/PhysRevE.90.053306DOI
  31. [32]Kozeny, J. (1927). Über die kapillare Leitung des Wassers im Boden. Sitz. b. Sitzungberichte, Mathematik, Astronomie, Physik und Meteorologie, 136, 271–306.
  32. [33]Krach, D., Ruf, M., & Steeb, H. (2024). POREMAPS 1.0.0: Code, Benchmarks, Applications. DaRUS. doi: 10.18419/darus-3676DOI
  33. [34]Krach, D., & Steeb, H. (2023). Comparing methods for permeability computation of porous materials and their limitations. PAMM, 23(1), e202200225. doi: 10.1002/pamm.202200225DOI
  34. [35]Lemaitre, R., & Adler, P. (1990). Fractal porous media IV: Three-dimensional stokes flow through random media and regular fractals. Transport in Porous Media, 5, 325–340. doi: 10.1007/BF01141990DOI
  35. [36]Liao, Q., Wei, Z., Yan, Z., You, S., Cui, M., Liu, X., … Guo, X. (2024). Parallelizable permeability estimation of digital porous media for sandstone using subvolume properties for flow in porous media. Journal of Energy Engineering, 150(5), 04024023. doi: 10.1061/JLEED9.EYENG-5390DOI
  36. [37]Liu, J., Pereira, G. G., & Regenauer-Lieb, K. (2014). From characterisation of pore-structures to simulations of pore-scale fluid flow and the upscaling of permeability using microtomography: A case study of heterogeneous carbonates. Journal of Geochemical Exploration, 144, 84-96. doi: 10.1016/j.gexplo.2014.01.021
  37. [38]Markert, B. (2007). A constitutive approach to 3-d nonlinear fluid flow through finite deformable porous continua. Transport in Porous Media, 70(3), 427–450. doi: 10.1007/s11242-007-9107-6DOI
  38. [39]Math2Market GmbH. (2023). GeoDict Simulation Software Release 2023. 10.30423/release.geodict2023.
  39. [40]Message Passing Interface Forum. (2021). MPI: A message-passing interface standard version 4.0. Retrieved from https://www.mpi-forum.org/docs/mpi-4.0/mpi40-report.pdf
  40. [41]Noiriel, C., Steel, C. I., Yang, L., & Bernard, D. (2016). Effects of pore-scale precipitation on permeability and flow. Advances in Water Resources, 95, 125-137. doi: 10.1016/j.advwatres.2015.11.013DOI
  41. [42]Osorno, M., Schirwon, M., Kijanski, N., Sivanesapillai, R., Steeb, H., & Göddeke, D. (2021). A cross-platform, high-performance sph toolkit for image-based flow simulations on the pore scale of porous media. Computer Physics Communications, 267, 108059. doi: 10.1016/j.cpc.2021.108059DOI
  42. [43]Peyret, R., & Taylor, T. D. (1983). Computational Methods for Fluid Flow. Springer Berlin Heidelberg. doi: 10.1007/978-3-642-85952-6DOI
  43. [44]Piovesan, A., Achille, C., Ameloot, R., Nicolai, B., & Verboven, P. (2019). Pore network model for permeability characterization of three-dimensionally-printed porous materials for passive microfluidics. Physical Review E, 99, 033107. doi: 10.1103/PhysRevE.99.033107DOI
  44. [46]Ruf, M., & Steeb, H. (2020). An open, modular, and flexible micro X-ray computed tomography system for research. Review of Scientific Instruments, 91(11), 113102. doi: 10.1063/5.0019541DOI
  45. [47]Ruf, M., & Steeb, H. (2024). In situ micro-XRCT data set of an open-cell polyurethane foam sample under uniaxial compression load. DaRUS. doi: 10.18419/darus-3010
  46. [49]Russ, J. C., & Neal, F. B. (2016). The Image Processing Handbook (7th ed.). Boca Raton, FL: CRC Press.
  47. [50]Saxena, N., Hofmann, R., Alpak, F. O., Berg, S., Dietderich, J., Agarwal, U., . . . Wilson, O. B. (2017). References and benchmarks for pore-scale flow simulated using micro-ct images of porous media and digital rocks. Advances in Water Resources, 109, 211-235. doi: 10.1016/j.advwatres.2017.09.007DOI
  48. [51]Scheidegger, A. E. (1957). The Physics of Flow through Porous Media. University of Toronto press.DOI
  49. [53]Stock, S. R. (2019). MicroComputed Tomography: Methodology and Applications (2nd ed.). Boca Raton, FL: CRC Press.
  50. [54]Tartakovsky, A. M., Trask, N., Pan, K., Jones, B., Pan, W., & Williams, J. R. (2016). Smoothed particle hydrodynamics and its applications for multiphase flow and reactive transport in porous media. Computational Geosciences, 20, 807–834. doi: 10.1007/s10596-015-9468-9DOI
  51. [55]Thomas, M., & Corry, B. (2016). A computational assessment of the permeability and salt rejection of carbon nanotube membranes and their application to water desalination. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374(2060), 20150020. doi: 10.1098/rsta.2015.0020DOI
  52. [56]Tuller, M., Kulkarni, R., & Fink, W. (2013). Segmentation of X-ray CT data of porous materials: A review of global and locally adaptive algorithms. In S. H. Anderson & J. W. Hopmans (Eds.), Soil–water–root processes: Advances in tomography and imaging (pp. 157–182). American Society of Agronomy and Soil Science Society of America. doi: 10.2136/sssaspecpub61.c8DOI
  53. [57]Wagner, A., Eggenweiler, E., Weinhardt, F., Trivedi, Z., Krach, D., Lohrmann, C., . . . Rybak, I. (2021). Permeability estimation of regular porous structures: A benchmark for comparison of methods. Transport in Porous Media, 138, 1–23. doi: 10.1007/s11242-021-01586-2DOI
  54. [59]Weinhardt, F., Class, H., Vahid Dastjerdi, S., Karadimitriou, N., Lee, D., & Steeb, H. (2021). Experimental methods and imaging for enzymatically induced calcite precipitation in a microfluidic cell. Water Resources Research, 57(3), e2020WR029361. doi: 10.1029/2020WR029361DOI
  55. [60]Weinhardt, F., Deng, J., Hommel, J., Vahid Dastjerdi, S., Gerlach, R., Steeb, H., & Class, H. (2022). Spatiotemporal distribution of precipitates and mineral phase transition during biomineralization affect porosity–permeability relationships: Microfluidic investigations. Transport in Porous Media, 143(2), 527–549. doi: 10.1007/s11242-022-01782-8DOI
  56. [62]White, F. M., & Majdalani, J. (2006). Viscous Fluid Flow. McGraw-Hill New York.
  57. [63]Wildenschild, D., & Sheppard, A. P. (2013). X-ray imaging and analysis techniques for quantifying pore-scale structure and processes in subsurface porous medium systems. Advances in Water Resources, 51, 217-246. (35th Year Anniversary Issue) doi: 10.1016/j.advwatres.2012.07.018DOI
  58. [64]Wildenschild, D., Vaz, C., Rivers, M., Rikard, D., & Christensen, B. (2002). Using X-ray computed tomography in hydrology: systems, resolutions, and limitations. Journal of Hydrology, 267(3), 285-297. doi: 10.1016/S0022-1694(02)00157-9DOI
  59. [65]Withers, P. J., Bouman, C., Carmignato, S., Cnudde, V., Grimaldi, D., Hagen, C. K., . . . Stock, S. R. (2021). X-ray computed tomography. Nature Reviews Methods Primers, 1(18). doi: 10.1038/s43586-021-00015-4DOI

Paper details

vlnc:W_2yb6bs63
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents