Form Approved

REPORT DOCUMENTATION PAGE

OMB No. 0704-0188
The public reporting burden for this collection of informetion is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources,
gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection
of information, including suggestions for reducing the burden, to Department of Defense, Washington Headquarters Services, Directorate for Information Operations Bnd Reports
10704-0188), 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be
subject to any penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number.
PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.
1. REPORT DATE (DD-MM- YYYYI 2. REPORT TYPE 3. DATES COVERED (From - To)
01062008 Journal Article
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER

Application of a New Grain-Based Reconstruction Algorithm to 5b. GRANT NUMBER

Microtomagraphy Images for Quantitative Characterization and Flow Modeling
EAR-0207788
5c. PROGRAM ELEMENT NUMBER

6. AUTHOR(S) 5d. PROJECT NUMBER

Karsten E. Thompson, Clinton S. Willson, Christopher D. White, Stephanie 5e. TASK NUMBER
Nyman, Janok P. Bhattacharya, Allen Reed

5f. WORK UNIT NUMBER

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

REPORT NUMBER
Naval Research Laboratory
Marine Geoacoustics Division NRL/JA/7430-07-13
Stennis Space Center, MS 39529

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITORS ACRONYM(S)

NSF
National Science Foundation
4201 Wilson Blvd. 11. SPONSOR/MONITOR'S REPORT
NUMBER(S)
Arlington VA 22230

12. DISTRIBUTION/AVAILABILITY STATEMENT

Approved for public release; distribution is unlimited

13. SUPPLEMENTARY NOTES

SPE 95887, June 2008 SPE Journal

20090515193
14. ABSTRACT
X-ray computed microtomography (XMT) is used for high-
resolution, nondestructive imaging and has been applied success-
fully to geologic media. Despite the potential of XMT to aid in
formation evaluation, currently it is used mostly as a research tool.
One factor preventing more widespread application of XMT tech-
nology is limited accessibility to microtomography beamlines. An-
other factor is that computational tools for quantitative image
analysis have not kept pace with the imaging technology itself.
In this paper, we present a new grain-based algorithm used for
network generation. The algorithm differs from other approaches
herause it uses the granular structure of the material as a template
15. SUBJECT TERMS

16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF RESPONSIBLE PERSON
ABSTRACT OF Allen Reed
a. REPORT b. ABSTRACT c. THIS PAGE
PAGES
Unclassified Unclassified Unclassified uu 13
19b. TELEPHONE NUMBER (Include area code!
202-688-5473
Standard Form 298 (Rev. 8/98I
Prescribed by ANSI Std. Z39.18
 Application of a New Grain-Based
Reconstruction Algorithm to
Microtomography Images for
Quantitative Characterization and
Flow Modeling
Karsten E. Thompson, Clinton S. Wlllson, and Christopher D. White, Louisiana State University;
Stephanie Nyman, University of Waikato; Janok P. Bhattacharya, University of Houston; and
Allen H. Reed, Naval Research Laboratory

Summary fluid dynamics community. The boundary clement method is very

X-ray computed microtomography (XMT) is used for high- well suited for low-Reynolds number flow of Newtonian fluids (in-
resolution, nondestructive imaging and has been applied success- cluding multiphase flows). Finally, the lattice-Boltzmann method
fully to geologic media. Despite the potential of XMT to aid in has been favored in the porous-media community because it easily
formation evaluation, currently it is used mostly as a research tool. adapts to the complex geometries found in natural materials.
One factor preventing more widespread application of XMT tech- A less rigorous approach is network modeling, which gives an
nology is limited accessibility to microtomography beamlines. An- approximate solution to the governing equations. It requires dis-
other factor is that computational tools for quantitative image cretization of the pore space into pores and pore throats, and trans-
analysis have not kept pace with the imaging technology itself. port is modeled by imposing conservation equations at the pore
In this paper, we present a new grain-based algorithm used for scale. Network modeling involves two levels of approximation.
network generation. The algorithm differs from other approaches The first is the representation of the complex, continuous void
because it uses the granular structure of the material as a template space as discrete pores and throats. The second is the approxima-
for creating the pore network rather than operating on the voxel set tion to the fluid mechanics when solving the governing equations
directly. With this algorithm, several advantages emerge: the al- within the networks. The positive tradeoff for these significant
gorithm is significantly faster computationally, less dependent on simplifications is the ability to model transport over orders-of-
image resolution, and the network structure is tied to the funda- magnitude larger characteristic scales than is possible with direct
mental granular structure of the material. In this paper, we present solutions of the equations of motion. Consequently, the two ap-
extensive validation of the algorithm using computer-generated proaches (rigorous modeling of the conservation equations vs. net-
packings. These analyses provide guidance on issues such as ac- work modeling) have complementary roles in the overall context
curacy and voxel resolution. The algorithm is applied to two sand- of multiscale modeling. Direct methods will remain essential for
stone samples taken from different facies of the Frontier Formation studying first-principles behavior and subporc-scale processes
in Wyoming, USA, and imaged using synchrotron XMT. Morpho- such as diffusion boundary layers during surface reactions, while
logic and flow-modeling results are presented. network modeling will provide the best avenue for capturing larger
characteristic scales (which is necessary for modeling the pore-to-
Introduction continuum-scalc transition).
Subsurface transport processes such as oil and gas production arc This research addresses one of the significant hurdles for quan-
multiscale processes. The pore scale governs many physical and titative network modeling: the use of high-resolution imaging of
chemical interactions and is the appropriate characteristic scale for real materials for quantitative flow modeling. We focus in particu-
the fundamental governing equations. The continuum scale is used lar on XMT to obtain 3D pore-scale images, and present a new
for most core or laboratory scale measurements (e.g., Darcy ve- technique for direct mapping of the XMT data onto networks for
locity, phase saturation, and bulk capillary pressure). The field quantitative modeling. This direct mapping (in contrast to the gen-
scale is the relevant scale for production and reservoir simulation. eration of statistically equivalent networks) ensures that subtle
Multiscale modeling strategies aim to address these complexi- spatial correlations present in the original material are retained in
ties by integrating the various length scales. While pore-scale mod- the network structure.
eling is an essential component of multiscale modeling, quantita- We refer to the network-generation technique presented here as
tive methods are not as well-developed as their continuum-scale a grain-based algorithm, which refers to the fact that the first step
counterparts. Hence, pore-scale modeling represents a weak link in in the algorithm (prior to network generation) is the characteriza-
current multiscale techniques. tion of the underlying granular structure in the material. The ben-
The most fundamental approach for pore-scale modeling is efits to this approach arc threefold. First, mapping of the granular
direct solution of the equations of motion (along with other rel- structure is less sensitive to image resolution than, for instance, the
evant conservation equations), which can be performed using a direct mapping of a pore skeleton. Second, because the pores in a
number of numerical techniques. The finite-element method is the granular material are necessarily defined by their surrounding
most general approach in terms of the range of fluid and solid grains, a map of the grain structure is an ideal template for gen-
mechanics problems that can be addressed. Finite-difference and erating the pore network; in principle, the map of the grain struc-
finite-volume methods arc more widely used in the computational ture should be essentially independent of image resolution, which
means that the network structure will correlate more strongly to
granular structure than to voxel size. Finally, because the number
of grains is typically orders-of-magnitude smaller than the number
Copyright © 2008 Society of Petroleum Engineers of voxels, the grain-based algorithms are faster than equivalent
This paper (SPE 95887) was first presented at the 2005 SPE Annual Technical Conference voxel-based algorithms.
and Exhibition. Dallas. 9-12 October, and revised for publication. Original manuscript re- In this work, we use a series of computer-generated sphere
ceived for review 14 July 2005. Revised manuscript received 17 September 2007. Paper
peer approved 8 November 2007. packings to validate and quantify errors in the grain-based net-

164 June 2008 SPE Journal

 TABLE 1--PARAMETERS USED TO DEFINE THE PORE NETWORK
Variable Association Variable Name Variable Type Dimension

Network Domain dimensions Vector Length

Pore Location Vector Length
Void volume Scalar Length3
Maximum inscribed radius Scalar Length
Throat lnterconnectivity:periodicity Scalar: vector
Cross-sectional area Scalar Length2
Maximum inscribed radius Scalar Length
Surface area Scalar Length2
Hydraulic conductivity Scalar Length3

work-generation algorithm. The algorithm is then used to perform (1956). For an extended period, network modeling techniques em-
network modeling on two different facies from a well- ployed lattice-based networks, usually decorated with a distribu-
characterized outcrop (Gani and Bhattacharya 2003), which illus- tion of pore-body and/or pore-throat sizes. These lattice-based
trates the potential of network modeling to be used in a multiscale models are valuable for qualitative studies of transport in inter-
framework. Few details of the grain-reconstruction algorithm (the connected, heterogeneous structures. However, they have not
precursor to the network-generation algorithm) are given in this proved to be effective for quantitative modeling of real materials.
paper because they can be found elsewhere (Thompson et al. Beginning in the early 1990s, new techniques were developed
2006). It was originally developed for the analysis of marine sands for quantitative network modeling. Bryant et al. (1993b) created
(Reed et al. 2005), but is being applied to other materials. This physically representative network models from the highly charac-
work extends the grain-based approach to relatively low-porosity, terized Finney packing (Finney 1970). Oren and coworkers created
cemented sandstones, which pose different imaging and analy- synthetic (computer-generated) sandstones and developed a tech-
sis challenges than unconsolldated sands; however, the main focus nique for extracting networks from these structures (Bakke and
of the current paper is the generation of pore network models 0ren 1997; 0ren et al. 1998). Their group has continued to de-
rather than grain characterization. Although grain structure pro- velop this approach and the resulting networks have been used by
vides an initial template, ultimately the parameters in the pore a number of other investigators (Patzek 2001; Lopez et al. 2003;
network are computed directly from the voxel image. Hence (as Valvatne and Blunt 2003). Lindquist et al. (1996) worked with 3D
explained next), even if the level of consolidation in a sandstone microtomography images and computed the medial axes of the
presents problems for the grain-reconstruction algorithm, the re- pore structure; this approach was then extended to allow for direct
sulting pore network will still be a quantitative, one-to-one map of generation of network structures (Sok et al. 2002). Thompson and
the pore structure. Fogler (1997) applied the techniques of Bryant et al. (1993b) to
computer-generated packed beds, and Al-Raoush et al. (2003) ex-
Background tended this work to allow for network structures with arbitrary
XMT. XMT provides nondestructive and noninvasive 3D images connectivity. Ioannidis and coworkers have used simulated anneal-
of the interior of objects by mapping the X-ray absorption through ing and other algorithms to produce networks that conform to key
the sample. The amount of absorption depends on the chemical statistics in real materials (Talukdar et al. 2002; Liang et al. 2000;
composition and structure of the material and the X-ray energy. Ioannidis et al. 1997).
XMT is based on the reconstruction of the cross-section of an For the network modeling used in this work, we borrow the
object from its projection data, which is obtained by passing a terminology physically representative network models (Bryant et al.
series of rays through an object and measuring the attenuation of 1993b) to describe the general class of models, and we note two
these rays using detectors placed on the downbeam side of the important characteristics of these structures. First, this type of
object. Projections are obtained by measuring the X-ray attenua- network is a one-to-one mapping of a specific porous material of
tion coefficients of the sample at different angles as the sample is interest, which ensures that subtle spatial correlations in the pore
rotated about the vertical axis. These attenuation values are rep- structure are retained. Second, the networks are described using
resented in images as discrete elements (pixels in 2D images and rigorous geometric parameters, which ensures that the pore mor-
voxels in 3D images). Synchrotron radiation has several advan- phology is not compromised despite the need to discretize the pore
tages over traditional X-ray sources, including high flux intensity space. This latter point contrasts with techniques in which the pore
(number of photons per second); a high degree of collimation structure is transformed into a network of interconnected capillar-
(source divergence leads to image blur); and the ability to tune the ies from the outset, an approach that has been shown to cause
photon energy to a single energy or frequency over a wide range ambiguity in the subsequent modeling of flow (Balhoff and
using an appropriate monochromator, which can be used to make Thompson 2004).
element-specific measurements. There is no unique or correct discretization of most real pore
Over the last two decades, XMT has played an increasingly structures (exceptions being simple structures such as cubic pack-
important role in the characterization of porous media flow and ings of spheres). Likewise, there is more than one approach that
transport. Because of its nondestructive nature and increasingly can be taken to describe a network model. In this work, we use the
high resolution, synchrotron-based XMT provides the high-quality set of parameters shown in Table 1 to describe the network struc-
datasets necessary to capture the 3D microstructure of the media ture. In addition, the network remains linked to the original data
(Liang et al. 2000a, b; Lindquist et al. 2000; Al-Raoush and Will- (whether it is a voxel image or a computer-generated material).
son 2005b) and the distribution of fluids within the pore space This methodology ensures that additional characterization could be
(Seright et al. 2002; Seright et al. 2003; Al-Raoush and Willson performed if warranted by a particular modeling algorithm. (It
2005a). The ability to characterize and correlate the void space guarantees, in essence, that none of the original morphologic data
microstructure and fluid distributions provides data to improve and are lost.)
validate pore-scale models. The flow modeling itself is performed by imposing conserva-
tion equation (s) at each pore in the network. This results in a set of
Network Modeling. Network modeling has a long history in the linear or nonlinear algebraic equations, depending on the physics
oil and gas industry, beginning with the landmark paper of Fatt of the process being modeled. A description of flow-modeling

June 2008 SPE Journal 165

algorithms is beyond the scope of the current paper, but this in- severely reduced by calcite concretions. Concretions locally re-
formation can be found in many other papers on network modeling duce the permeability by several orders of magnitude in approxi-
(Bryant et al. 1993b; Bakke and 0ren 1997; Patzek 2001; Lopez mately 15 volume % of the tidal sandstone facies (Lee et al. 2007).
et al. 2003; Valvatne and Blunt 2003; Thompson and Fogler 1997; The sediments preserved in these rocks were derived from uplifts
Balhoff and Thompson 2004; Bryant et al. 1993a). to the north and west. Lee et al. (2007) report that the average
composition is approximately 51% quartz, 21% rock fragments,
Network Generation From Voxel Data. Despite the advances and 28% feldspars, with calcite as the dominant diagenetic cement.
described in the previous section, what remains surprisingly dif- Feldspar and volcanic components have been highly altered by
ficult is the generation of physically representative network mod- diagenesis. In uncemented rocks of the tidal sandstone facies, po-
els directly from microtomography images of real materials. The rosity averages approximately 0.22 and permeability is in the range
difficulty stems from the distinct differences in the form and scale of tens to a few hundred millidarcies, with a mean of 110-140 md
of the data structures. The XMT data consist of hundreds of mil- (higher for fluvial-dominated deposits of the delta front).
lions (sometimes billions) of voxels on a Cartesian grid. In con- Sample B is from a channel sandstone facies within the same
trast, the network representation of the same porous medium is delta front sandstone body as sample A. The provenance of this
likely to be described using a much smaller data set (order 103—10s sample is thus similar to Sample A, except for calcite cement
pores), which are interconnected by a complex structure of flow- replacing some of the pore space. This sample is interpreted to be
paths rather than aligned with any specific coordinate system. from a relatively mud-free cross-stratified channel sandstone fa-
Transforming the former to the latter is a nontrivial procedure. cies. Depositional considerations would imply very high reservoir
Most previous work in this area employs the medial axis as a quality. The grain size is on average larger, and sorting is better
basis for characterizing the pore structure. For discretized images than sample A. However, it is from a cemented region within the
such as those from XMT, the medial axis is obtained by thinning sandstone, and its reservoir quality is therefore relatively low. The
the original pore structure until a one-voxel-thick skeleton re- greater abundance of calcite concretions in some channel sand-
mains, or by computing distance transforms in the void space (to stones is inferred to be related to the abundance of mud chips (or
the void/solid interface) and retaining the skeleton of local shale clasts). Although mud chips would decrease initial perme-
maxima. Different medial-axis structures are obtained depending ability, their inhibiting effect on concretion formation actually re-
on the method used, the order for thinning, the type of distance sults in higher, post-diagenetic reservoir quality for channel sand-
transform used, and/or the rules invoked to ensure that topology is stones with mud chips (Lee et al. 2007).
retained (Lohman 1998). Lindquist et al. (1996) used the medial
axes of 3D images to compute statistical parameters associated XMT of the Sandstone. Small samples of the sandstones were
with the void space. They later extended their algorithm, using the impregnated with an epoxy resin under vacuum and then cored to
medial axis as a basis for obtaining pore- and throat-size distribu- a length of 25 mm and a diameter of 5 mm. 5-mm-long sections of
tions, which in turn were mapped onto a network structure the cores were Imaged at 33.07 keV energy at the 13-BMD tomog-
(Lindquist et al. 2000). Specifically, the medial axis was trimmed raphy beamline, operated by GeoSoilEnviroCARS (GSECARS) at
down to the percolating fraction and nodes were then merged (by the Advanced Photon Source. Image reconstruction was performed
comparing the distance separating neighboring nodes to the dis- using algorithms developed by GSECARS to convert CT attenu-
tance to local surfaces) to define pore locations. Pore throats were ation data to 3D volumetric data The resultant 3D gray-scale im-
found by dilating the medial axis until the dilated cylinder con- ages have a voxel resolution of 7.63 microns. Segmentation, the
tacted the bounding surfaces; pore-throat geometries were then process of converting a gray-scale image to a 1 -bit or binary image
obtained by triangulating between the medial axis and voxels by separating the images into two populations based on gray-scale
along the perimeter of the constriction. Sok et el. simulated im- values, was performed using the indicator kriging technique (Oh
miscible displacement on networks generated by this algorithm and Lindquist 1999) that is part of the 3DMA software package. In
and compared them to immiscible displacements on regular lat- this work, the pore voxels are assigned a value of 0 and the grain
tices having identical values for key statistical metrics (Sok et al. voxels are assigned a value of 1. Fig. 1 contains images of the
2002). Mirroring what Bryant et al. (1993a) observed for single binary volume files.
phase flow, they note that multiphase flow behavior is not repro- The combination of mineralogy (i.e., higher absorbing ele-
duced correctly on statistically equivalent networks, emphasizing ments) and X-ray energy created a relatively high signal-to-nolse
the need for algorithms that capture true pore morphology. Delerue ratio. This less-than-ideal image quality made it difficult to com-
et al. (1999) applied a medial-axis technique to a 3D image of a pletely resolve the solid and void phases and to remove the ring
resin-impregnated soil, and then defined the pore-size distribution artifacts. Since the time that these experiments were performed,
in the soil by measuring the maximum inscribed balls for all voxels these issues have been addressed, which means that images from
contained on the skeleton. Mercury intrusion was simulated di- current experiments are generally free from these problems.
rectly on the voxel map (rather than a network representation).
Delerue and Perrier (2002) describe in detail the various compu- Network Construction
tational elements used in the algorithm. Silin et al. (2003) em- Overview. The input for the algorithm is a binary volume file
ployed a similar approach, except that the maximal inscribed ball describing the porous material of interest: voxel labels indicate
was found for each void-phase voxel in the packing. Though com- whether the voxel is contained in the void phase or solid phase of
putationally more intensive, this approach allows the pores to be the material. The algorithm operates by identifying grains, search-
found independently of the skeleton. They tested their algorithm ing for pores based on the grain locations, and then creating the
using computer-generated sphere packings and a CT image of interconnected network using a novel restricted-burn algorithm.
Fontainbleau sandstone. However, pore network models were not The significant differences between this algorithm and other net-
created, and a quantitative assessment of the pore locations, sizes, work generation techniques include the following:
and connectivity was not made. l.The first step in the algorithm is the identification of the
granular structure, a process that is less sensitive to image resolu-
Materials and Methods tion than techniques such as skeletonization. Consequently, net-
Sandstone Samples. The samples used for this application were works created for the same material at two different image reso-
taken from the Wall Creek Member of the Cretaceous Frontier lutions have very similar structure and properties.
Formation, Wyoming, USA as part of an integrated geologic, geo- 2. The algorithm uses the granular structure as a template to
physical, and engineering study (Gani and Bhattacharya 2003). help define pores. This means that the computationally expensive
Ten wellbores were drilled through the Wall Creek Member and search for pores becomes linked to the number of grains in the
sample plugs were selected. image rather than the number of voxels, and the computational
Sample A is from a tidally-reworked sandstone facies within a penalty for increasing image resolution is less severe than with
delta front sandstone body. The permeability of this sample is other algorithms.

166 June 2008 SPE Journal

 Step 2. Location of Extrema. A search is performed to find
local maxima in the burn assignments. These local extrema are
%. • _•— ~ *c",^5d!r** -SL •r•"»«i. *"*L—*-<^» J islands of one or more voxels that are surrounded by voxels with
lower burn numbers. In the simplest form of the algorithm, the
local maxima are taken to be the grain centers. However, refining
the grain locations using optimization leads to better results
(Thompson et al. 2006). Simultaneously, the local minima can be
found (which according to the sign convention are in the void
phase). In practice, this step is skipped to reduce computation time
because die void-phase extrema are not used to define the pore
structure. However, the minima are found and reported for com-
pleteness in this paper.
Step 3. Tessellation of the Grain Centers. A periodic Delaunay
tessellation is performed using the locations of tiie grain centers
(Thompson 2002). The purpose of the tessellation is to identify
likely pore locations based on the granular structure. The role of
the tessellation is somewhat subtle, and the following points are
worth clarifying:
1. In contrast to other techniques where the Delaunay tessella-
tion is used to define pore structure (Bryant et al. 1993b; Thomp-
son and Fogler 1997; Al-Raoush et al. 2003; Bryant et al. 1993a),
it does not influence the structure of the pores in the current al-
gorithm. In fact, the only restriction that it imposes is on the total
number of pores: the algorithm will not find more pores than the
number of tetrahedrons in the tessellation. This limitation should
not be of practical consequence because the tendency of the De-
launay tessellation is to identify too many pores [by splitting single
pores into multiple tetrahedrons; see Al-Raoush et al. (2003)].
Furthermore, this restriction can be relaxed if necessary at the
expense of increased computation time (see Item 3).
2. Use of the Delaunay tessellation to aid in the identification of
pore locations provides an advantage over erosion/dilation tech-
niques because the number of pores that are located for a given
porous medium is relatively insensitive to voxel resolution. This
issue is demonstrated later in the validation section.
3. Finally, the tessellation is a valuable but not essential part of
the algorithm. An alternative approach would be to perform the
optimizations described in the next step beginning with every
void-phase voxel. However, even for the relatively small sand-
stone datasets used here, this approach requires solving -5,000,000
nonlinear optimization problems (in contrast to -25,000 optimiza-
tions when the tessellation is used as a template).
Step 4. Locating Pores. In the Introduction, we note that the
division of void space into pores is somewhat arbitrary. In this
work, we use the distance function d(x,y,z), whose value gives the
minimum distance from the point [x,y,z\ in the void space to a
Fig. 1—Binary volume files for two sandstone samples: (a) tid- grain surface (Luchnikov et al. 1999), and then define a pore as
ally-reworked sandstone fades; (b) channel-sandstone facies. any local maxima in d. In practice, this definition corresponds to
an accepted definition for pores: locations of maximum-diameter
spheres that can be inscribed into the void space, and that are
3. Because the grain locations provide the framework for de- constrained from movement by the surrounding solid phase
termining pore structure, the algorithm operates without ever com- (Scheidegger 1974). In Step 4, these maximum-diameter inscribed
puting a skeleton of the pore space. By eliminating this step, a spheres are located by performing repeated optimization proce-
number of problematic issues such as nonuniqueness, dependence dures (to maximize d) using the tetrahedrons as seed locations to
on image resolution, and the formation of internal loops are avoided. start the optimizations.
4. A key intermediate step between the original binary map and The optimizations themselves are performed using a modified
the final network structure assigns an integer label to every voxel Powell's method (Press et al. 1992), which is a direction set
(both void and solid phase) to mark the grain or the pore to which method effective for situations where gradients in the objec-
that voxel belongs. The reason for emphasizing this intermediate tive function cannot be calculated directly. In essence, the pro-
step is that it makes obtaining statistical information and construct- cedure repeatedly performs ID line minimizations. Various
ing the network straightforward and unambiguous. schemes are available to choose the minimization directions, most
based on estimating new conjugate directions from the history of
Grain Reconstruction Algorithm. Step 1. Dual-Phase Burn. A the optimization.
simultaneous grain-phase and void-phase burn is performed (using As each extremum is found, the voxel containing its x.y.z lo-
the terminology of Lindquist et al. 1996). Voxels in the solid phase cation is marked with the pore number. If a different seed has
are labeled with positive integers denoting the burn level and vox- already converged to this same voxel (though the coordinate lo-
els in the void phase are labeled with the negative of the burn level. cation would rarely be exactly the same), a new pore is not added
This convention is not necessary if the burn map remains coupled to the list. Hence, Step 4 ends having generated a list of N pore
with the binary material array; however, the use of opposing signs locations and inscribed radii (i.e., maximum d values) along
allows the burn numbers to be written over the initial material with N void-phase voxels labeled with the corresponding pore
array without losing the phase information. number. At this point, the pores are no longer tied to their seed

June 2008 SPE Journal 167

tetrahedrons and there is no further need for the Delaunay tessel- of all pore throat surface areas equals the total surface area in the
lation In the algorithm. packing), thus providing a good theoretical foundation for its use
Step 5. Pore Merging. A viable option is to use all pore loca- in modeling.
tions identified in Step 4, and proceed with constructing the net- The last parameter that should be mentioned is the pore-throat
work. However, in real cases, many pores overlap with one another hydraulic conductance of each throat, which is necessary for com-
by a significant amount, the overlaps being caused by one of two puting permeability and performing dynamic flow simulations. In
reasons: (a) the local pore geometry causes two independent ex- general, the conductance is computed using some combination of
trema to be in proximity; (b) two different seeds have led to the the previously mentioned parameters, through generation of an
same extremum, but numerical error or optimization tolerances equivalent capillary. In this work, we use essentially the same
have caused the computed extrema locations to differ by at least approach used in Thompson and Fogler (1997) (without the FEM
one voxel. computations). Further work includes evaluating the best way to
In either of these cases, there is good reason to merge two compute throat conductivities for various types of materials.
largely overlapping pores into one. Various merge criteria can be The network itself is defined by mapping out the connectivity
devised. In the current algorithm, pores are merged only if one of each pore, which is defined by the list of neighboring pores that
inscribed sphere encompasses the center of a neighboring in- share voxel-voxel contacts. There are no limitations on the struc-
scribed sphere. The location and radius of the larger inscribed ture of the networks thus obtained (i.e., coordination numbers,
sphere is used as a seed, and a local optimization is performed once etc.), and the resulting network files have exactly the same format
again to verify the location of the merged pore. With the adjusted as network descriptions from computer-generated media (Balhoff
pore locations (owing to the reoptimization of merged pores), we and Thompson 2004). Once the network is constructed, plots of
have found it advisable to make another pass to check whether pore-size distribution, throat-size distribution, and coordination
additional pores should be merged, and indeed to continue this number can be made directly from the data in the network file,
merge—>re-optimization procedure iteratively until no more pores without having to return to the large voxelized data sets.
are merged. Step 5 ends with the same information as Step 4 (pore
locations, radii, and corresponding voxel assignments), but with Validation. The grain-reconstruction algorithm has been tested
fewer and more spatially independent pores. extensively using computer-generated packings of spheres and cyl-
Step 6. Grain and Pore Assembly. Step 6 is the key interme- inders as well as XMT data from unconsolidated particles and
diate step mentioned in the previous overview: the assignment of sands. Separately, the network-generation algorithm has been
all voxels in the image to one of the grains or pores identified in tested using a series of computer-generated packings of spheres.
Steps 2 or 5, respectively. This step is performed using a novel This approach is valuable for three reasons: First, the pore struc-
restricted burn algorithm, which is described in more detail in the ture in the material is known exactly; hence, the validity of the
context of grain reconstruction elsewhere (Thompson et al. 2006). network generation process can be assessed in a quantitative man-
In short, it assembles a cluster of voxels together that are tied to ner. Second, the computer-generated structures can be discretized
one of the grain or pore locations respectively, in accordance with at arbitrary voxel resolution, which provides good benchmarks for
the local geometry. the accuracy that can be expected with data from real materials.
A summary of the logic for this restricted burn procedure (Step Third, the computer-generated data are free from noise and arti-
6) is the following: facts, which allows validation to be focused solely on the network
1. Set the minimum bum level to the largest (absolute) value construction (rather than imaging and segmentation issues).
found in the pore space. Three sphere packings were used for validation: a cubic pack-
2. Loop through all voxels in the domain. ing, a rhombohedral packing, and a random packing. The two
3. If the current voxel borders a voxel already assigned to a regular packings span the full range of attainable porosities for
particular pore and its absolute burn level is greater than or equal monodisperse spheres and the pore structure is known exactly. The
to the current minimum burn level, assign it to the same pore as the random packing is more representative of real materials. No single
assigned neighbor. network is correct for the random case. However, the modified
4. If any new voxels were assigned during the last pass, go to Delaunay Tessellation (MDT) algorithm (Al-Raoush et al. 2003) is
Step 2. a fairly rigorous method for extracting the pore structure and can
5. If no new voxels were assigned during the last pass, reduce be used for comparison to the networks generated from the vox-
the minimum burn level by 1; go to Step 2. elized images.
Step 7. Pore Morphology and Network Construction. Once Table 2 explains the notation used for reporting the statistical
Step 6 is complete, determining morphologic parameters and con- results. Selected quantitative data are presented in Table 3. (Data
structing the network is a straightforward process. The total vol- for the computer generated packings are in units of sphere diam-
ume of each pore is obtained by summing the volume of all voxels eters; data for the sandstone shown later are in microns.) Note that
assigned to that particular pore. The inscribed pore radii are al- for the three sphere packings, the top row of each section contains
ready known from the Step 4 and Step 5 computations. parameters obtained from the MDT networks. These parameters
In our definition of network structure, pore throats have no quantify the error in the voxel-based networks. For the cubic and
volume but rather are defined by the faces where two pore- rhombohedral packings, the MDT values agree with theoretical
elements come into contact. Hence, the pore-throats, like the pores, values from a unit-cell analysis and are exact within the numerical
have rigorous geometric parameters associated with them. The unit tolerances set in the MDT algorithm. As mentioned previously,
normal for the pore-throat interface is found by averaging the there are no "correct" morphologic values for the random packing.
orientation of all voxel interfaces associated with the given throat. However, the MDT algorithm is an excellent benchmark because
The total cross-sectional area of the throat is then found by sum- it is based on Delaunay analysis (Finney 1970; Mellor 1989) and
ming the areas of voxel faces at this interface projected onto this allows for variable pore connectivities (which are not possible with
local unit normal. Use of the projected area prevents overesti- the unmodified Delaunay tessellation).
mating the area because of the staircase-like surface created by Identification of Pores. Column #£Vis the number of extrema
the voxels. The inscribed area of the throat is found by determin- in the void-phase burn. This parameter is the most logical choice
ing the largest inscribed sphere whose center is located on the for determining pore locations if the burn information were to be
throat interface. used directly. However, notice the very strong dependence of these
Grain surface area is also assigned as a throat parameter be- values on image resolution (at least for the noncubic structures).
cause it affects permeability as well as phenomena such as absorp- For the rhombohedral packing, the number of void-phase extrema
tion and chemical reactions on grain surfaces. The surface area is varies from four to 328 as image resolution is increased over an
assigned by estimating the surface area for each surface voxel, and order-of-magnitude range (the true number of pores is 192). For
then assigning that element of surface to its closest throat. This the random packing, number of burn extrema varies between 63
approach ensures that the surface area is conservative (i.e., the sum and 803 as resolution varies from 10 VPD to 125 VPD (the best

1G8 June 2008 SPE Journal

 resolutions. For the rhombohedral packing, it is exact for resolu-
TABLE 2—EXPLANATION OF PARAMETERS
tions of 16.7 VPD and above. For the random packing, an exact
USED IN TABLES
number of pores can never be defined unequivocally. However, the
Parameter Description values obtained from the current algorithm and the MDT algorithm
agree reasonably well. Results from the random packing also il-
Aa Cross sectional area of a pore throat (ave) lustrate the efficiency of the grain-based approach. Examining
A, Surface area assigned to a pore throat (ave) the number of tetrahedrons (each of which generates a seed point
DG Grain diameter (ave) to search for a pore) shows that -620 nonlinear optimization
DP Pore diameter (inscribed) (ave) procedures were performed to find the initial pore locations,
independent of image resolution. In contrast, -32 million of these
DPT Pore-throat diameter (inscribed) (ave) same computations would have been required for the 100 VPD
UEG Number of extrema in the grain-phase burn image if each void-phase voxel were tested to find its maximum
UEV Number of extrema in the void-phase burn inscribed sphere.
K Permeability (subscripts indicate flow direction) Accuracy of Pore Locations and Sizes. The pore location is
defined as the center of the largest sphere that can be inscribed into
Ur Length of pore throat (ave)
a given void space. For the cubic packing, the results are essen-
MBG Maximum burn number in the grain phase tially exact in cases where an integer VPD value is used. For
MBV Maximum (absolute) burn number in the void noninteger VPD values, the voxels cannot be divided evenly be-
phase tween pores in the packing. Consequently, there remain slight
NG Number of grains distributions in pore parameters, even at high VPD values. How-
Np Number of pores ever, error remains fairly low in key parameters such as pore
volume (see V^ and pore location (Table 3, column PLE). The
PLE Average error (%) in the computed pore locations
rhombohedral packing is a much more rigorous test because the
#P, Number of pores initially found from the pores are not symmetric with respect to the Cartesian voxel grid
tetrahedron seeds and also because they are small. The error in pore locations is
UPM Number of one-to-one pore matches significant at low resolutions (where pore dimensions can be as
(reconstructed packing vs. original) small as a single voxel), but it appears to decrease monotonically
#P« Final number of pores after the iterative with increasing voxel resolution.
merging/reoptimization process The accuracy for the random packing network is somewhat
#Te( Number of tetrahedrons used as seeds for pore harder to assess, mostly because the number of pores found does
locations not agree exactly for the MDT vs. voxel-based networks. Column
vVmin Minimum number of voxel faces required for a #PM in Table 3 provides statistics for the fraction of pores widi a
pore throat to be formed one-to-one match with the MDT network (defined when the center
of one and only one pore from the voxel network lies inside an
#VOK Number of voxels in the data set
inscribed pore of the MDT network). Also shown is the average
V, Volume of a pore (ave) error in pore location for these one-to-one matches.
VPD Voxels per particle diameter Pore parameters are computed with increasing accuracy as the
VR Voxel resolution voxel resolution increases. This trend is easy to confirm by exam-
Z Pore coordination number ining the results for the cubic packing. The same trend occurs for
the other two packings, but is harder to assess from simple statis-
i Porosity tics: a rhombohedral packing contains two distinct types of pores
Other (which is why the standard deviation remains non-zero for both the
ave Arithmetic average MDT results and the high-VPD results) and die random packing
Sd Standard deviation has a distribution of pores.
Coordination Number. Determining the pore coordination
number (the number of throats emanating from a pore and con-
estimates for number of pores in this packing are -400). Clearly, necting to other pores) remains the biggest problem for this algo-
the void-phase burn provides a poor indicator of pore location, rithm (as well as other network generation algorithms). For the
which is related to the problems with skeletonization that were cubic packing, the coordination number is exact for integer VPD
mentioned previously. values, again because of the fact that pore boundaries are coinci-
In the current algorithm, the grain structure is used as a tem- dent with voxel boundaries. For resolutions that produce noninte-
plate for determining pore locations. The initial seeds for pore ger VPD values, the average coordination numbers are slightly
locations are based on a Delaunay tessellation of the grain loca- higher and the maximum coordination numbers are significantly
tions. Table 3 lists the number of Delaunay tetrahedrons (see col- higher than they should be. The problem is illustrated in Fig. 2,
umn #7"er). The number of largest-inscribed-spheres obtained from which shows two pores assembled from a cubic packing with a
these seeds is listed in column #P7. Note that the numbers in both noninteger resolution of 21.3 VPD. Although they are constructed
of these columns are relatively insensitive to image resolution correctly (within the limitations of the voxel resolution), the fact
since they are tied to the grain structure rather than the voxel structure. that the pore boundaries are not coincident with voxel boundaries
The cubic packing has significantly fewer inscribed spheres creates a contact between two voxel faces at a position that should
(i.e., pores) than tetrahedron seeds. The reason for this difference be a corner-corner contact in reality (and therefore should not
is that assembling a cubic pore requires at least five tetrahedrons. register as a throat connection). The numbers from the noninteger
If the optimization routine were exact, all five seeds would con- VPD cases indicate that this problem occurs less often than once
verge on the same central pore location. In reality, more than one per pore on average (though certain pores are assigned many extra
seed will usually converge on the same central voxel, and is listed neighbors, as evidenced by the high values for the maximum co-
only once. The fact that #P7 decreases with increasing resolution is ordination numbers). For the rhombohedral packing, the problem
somewhat counterintuitive (because higher resolutions provide is similar, though no resolution generates the exact coordination
many more voxels in which the optimizations can land). However, pattern because the rhombohedral pore boundaries never lie along
the reason for this effect is the increased accuracy of the optimi- voxel boundaries.
zation procedure as resolution increases (which is a consequence For random pore structures, the problem is exacerbated. The
of more precise distance-to-surface calculations). pore geometries are more irregular (anomalies are not limited to
The more important value is the final number of pores after corner-comer contacts as with the cubic packing) and, like the
merging (#P/J- For the cubic packing, this value is exact for all rhombohedral packing, the Cartesian voxel system does not allow

June 2008 SPE Journal 169

 in TJ- en •* in IO CO CM (O O *- ^>
t 6 (D (i O) So en h- co co en
co -rr o <•- o to r»- o> o i- o
O O CM O O q o o o *- t- T-
O O O O O d d d d d d d o o o

N. f^ n O CO r» CM co in co co r*- Q *- co t- CM •* m
00 f O) CM <3> CM co co co m co m in h- to ^r O) CM m
0 CN m 0> O C> CM tO I*- 01 CO O « ^t in ^ n ^ t
01 CD CO CD O CO CM CM CM CM CM CO C» CM CM CM CM CM CM
d d d d d d d d d d d d d d
10 •»- f-» O) ** <- o
«-
it
*7
CM
CM
O
lO
i-
O
01
CO
*•; 3 S
*-
CM
m co
CO CO ss
a O O O O O O O O O o

^ o ai ID tr o m
N 5 O Ol tf) s SCM *r to CO "* co
CM O m O rO *- to to *n ^r CM T-
m m TJ- in m m CO CM CN (N CM CM CM
o o o o o o

to ^ O 00 m oi i- o *- m
co in to
CO Tfr
p o p
CM
r»
cn
o
O
CM
O
CO
CM
^ g g o
s o* -
o
ooooooooo o o o
<!
Sen CO CD *T to
i- r^. in co en ST" CD CO O CO CO CO CO
to m o co v- -^ CO r-- if CO CM CO tO Ol CO CM O CJ>
•<t -<r ^- co «t -* r-t CM r- T- T- ,- r- «SJ CM CM CM CM T-
o oododddd d d d d d d d
to
CO
in
N
CO
% 5
CD CD
s
CD
CM
CO
o
CO
tototocotocDtbtoto to to to CO to to

tOCOfOtOtOtOCMtOtO CO O O 01 (O o
<N CM CM CO ro

totococotocotocoto CM »- CM *- CO CM T- CM T-

s 00
<D
CO
l£)
T-
-f a
* tp
CN
CO
CN
rO
CN
*-
CM
CN
IN

a en n (71
P
2 CDcOcOtOCOtOCOCO § o o O
8
Q-
tococootocc^oto
CM"
en
CM
en § § I s JI
N o CO
CN
CO
S «rt
CO

(ft ifi
O
O)
LD
in CM
o o o
*- o
o sss ss s g O O O CO c>
ooooooooo d d d d d d
CO CM
O CM O O N S O) tO CO CO to 00 CO O CO
S co CM Tf -^ tO r Q O) 0) Q SN ID in O CO
in m co x m in to O) CO CO CO N
5? 5
t V t f t V o o o o o f O O O o o
oodddddd d d d d d

co co TT CM o en T- *- to o ao CM r- cn
o o o o o •<- o CN 5 « to 5 s m
o to «- CO O i- o O) tt) O) Ol O) O) O)
o o o o o o o
o o o o o o o o o o o o d d d d d d d

O CO <J> CO i- T- m a) t- N a v
h» h- en CM t~- co Q cn r- m to o o co
00 CO O CD ao CO 0) m & in ^t ^ co to
<N CO CM CM CM CM CM co co co co co co co
ooooooooo d d d d d d d
t- to Ol <o T s
cotototoubtotOtoto o -^ co r- o o
co CO co co ^t t

m*-C0cnOr-r-3' Is. •*- t- OJ CM i~ O) *- CM

t-Or-fON-incOCTl N O N CO to CO Ol T- T-
MOJCNr-fNJi-t-i- CM CM CM •d in in m to to

m cn o> o cn CD
<- T- i- CM »- i-
to to to to CO to

E CDCDtOtDtOtOiOCO
co
to
r--
CM
m
m
co
r^-
o
r-
co
o
r r N tO CO

uUj o o o o o
COtOcOtOtOtOtotO to to to to to CO

MnND'tm--'- CM CM 't (0 O) CO

g co ro ^J- co in co m o oi ro CD cn CN CD s
T- CN CO

0) m m BQ
o *- TT o o
in co co
m r- tO i- CM CO

X
r^
X
o h- CO
1- i
<§ 2 o o m CO CO O to CO N. £31 CO
I
CO
CO O
^ sX
CO

i '-
m co co
X X
5 T£ 8 I i i II CM CM Tt
|
•
CO
CO K
X X
CO
cn
•*
CO
s
o r-- o o f^- p o o o o o o
I ocoomocoop
OO^CMO^OQ
^-X-T-^-CMCMIOO
d
T-
d in d to S
*- CM ID CD o
E
«- P.CN ip p o in
CM T

O in o o O o cn
° s o o o o o
Q
5
8 3 CM
O
O
o
o
6
o
T3

a ^
o
o
UO
O
O
5
o
CM
o
o

170 June 2008 SPE Journal

 20 40 60 80 100
Number of voxel faces per throat

Fig. 3—Histogram of the number of voxel-voxel contacts in pore

throats in a random packing of spheres.

Fig. 2—Non-physical pore-throat connection in a cubic packing caused the other parameters to deteriorate, indicating that real
of spheres. throats are being discarded with this higher limit.
Unfortunately, it Is difficult to generalize an effective rule for
how to limit interconnectivity. Creating a histogram (e.g., Fig. 3)
for perfect breaks in the pore geometry. Estimates of average pore
on a case-by-case basis is a sensible and fairly easy approach
coordination number fall into the mid-sixes range. These values with an automated algorithm. However, tomographic data from
are probably too high, again because of identification of voxel-
real materials that we have tested do not show the bimodal distri-
voxel connections that should not register as pore-pore connec- bution found in Fig. 3 and therefore do not provide a strong ra-
tions (a fact also evidenced by the large values of maximum co- tionale for a cutoff value. We suggest the following approach. For
ordination number). the purposes of network generation, small values or no cutoff
The obvious solution to this problem is to set a minimum limit should be used because extra throats will have negligible effect on
for the number of voxel-voxel contacts that constitutes a pore most transport processes (because of their very small size), and
throat. For the cubic packing, requiring pore throats to be com- because one risks eliminating what the tomography has identified
posed of three or more voxel faces led to perfect coordination as "real" connections. For the purposes of statistical analysis, the
numbers at all non-integer resolutions tested. For the 40 VPD extra throats do cause a problem because they affect the calcula-
random packing, Fig. 3 shows a histogram of the number of voxel tion of coordination number and average throat parameters. In
faces found at a throat connection. The histogram has a broad these cases, it is worth the extra effort to assess the distribution of
minimum before the population begins to increase at around 70 voxel contacts, and provide a cutoff if a case can be made to do so.
voxel-faces per throat. Interestingly, the theoretical minimum for Finally, the use of geometric averages should improve the statis-
throat size at this resolution is 64 voxel-faces per throat, which tical characterization even in cases where small throats are mis-
lends credence to die histogram. These two factors suggest that takenly identified.
connections made at less-than -64 voxel-faces/throat are anoma-
lies and should be discarded. Table 4 shows the results for various
minimum-voxel limits applied to the 40 VPD random sphere pack- Results
ing. For a conservative minimum of 50 voxel-faces/throat, the Networks were created using the sample A and sample B sand-
revised throat parameters are in much better agreement with the stone images. For illustration purposes, ball-and-stick depictions
MDT values, with the average pore coordination number reduced of the network structures are shown in Fig. 4. Comparing these
to 4.67. A more aggressive limit of 63 voxel-faces per throat networks to the Fig. 1 tomography images emphasizes the one-to-

TABLE 4—CHANGES IN PORE-THROAT PARAMETERS BY LIMITING CONNECTIVITY

Pore-Throat Parameters
z D °T Acs As
min max ave ave sd ave sd ave sd

MDT 2 12 4.60 0.2627 0.0846 0.3548 0.0963 0.3369 0.1233

0 26 6.60 0.2083 0.1015 0.2484 0.1479 0.2392 0.1448
1 26 6.60 0.2083 0.1015 0.2484 0.1479 0.2392 0.1448
2 23 6.08 0.2215 0.0950 0.2673 0.1386 0.2607 0.1475
3 21 5.87 0.2269 0.0921 0.2757 0.1338 0.2697 0.1477
4 20 5.71 0.2310 0.0898 0.2820 0.1299 0.2770 0.1472
5 20 5.71 0.2310 0.0898 0.2820 0.1299 0.2770 0.1472
10 19 5.24 0.2422 0.0842 0.3019 0.1164 0.3018 0.1414
50 18 4.67 0.2523 0.0814 0.3251 0.1008 0.3386 0.1255
63 0 17 457 0.3113 0.0960 02981 0.1536 0.3108 0.1792

June 2008 SPE Journal I7l

 Fig. 5—Grain reconstruction of the sandstone sample A.

process were identified. The first problem is single grains being

"broken" into more than one piece. This problem is associated with
noise in the data and/or the misidentification of grain centers
(which in turn stems from either unusual grain shapes or limita-
tions in the voxel resolution and resulting burn). The second prob-
lem is the opposite situation: the identification of a single large
grain that should probably be broken into more than one piece
(although this decision is subjective in consolidated materials).
Both problems appear to be more pronounced for the current sand-
stone data compared to unconsolidated sands. Solutions to these
problems are being implemented, but are not yet validated, and
thus were not used in the current work.
Fortunately, the main problem with the current image appears
to be the division of single grains into more than one piece, which
will not affect the network generation process in a detrimental
way; an increased number of grains will simply cause the algo-
rithm to search for more pores. However, the final pore locations
are determined by the optimization procedure and reflect local
geometry regardless of whether the local grains were reconstructed
in a reasonable way. Hence, the grain-based algorithm generates a
Fig. 4—Ball and-stick schematic illustration of network models physically representative network even if problems occur in the
corresponding to the two different sandstone samples. grain reconstruction process.
Table 5 contains selected parameters for the two different net-
one mapping that is achieved using a physically representative work generation techniques used on Sample A. Fig. 6 shows his-
network modeling technique. tograms of the inscribed pore-size, inscribed pore-throat-size, and
For the sample A data set, two different network reconstruc- pore coordination number for the grain-based network. Fig. 7 is a
tions have been performed. The first uses the grain-reconstruc- graphical representation of the two different network structures in
tion algorithm. For comparison, a second run was performed in a small interior section of the sandstone.
which the search for pores was performed directly on the void- The most striking difference between the two networks is the
phase voxels, without use of the grain reconstruction and the De- number of pores, with the grain-based network having many fewer
launay tessellation. pores and thus longer pore throats spanning the gaps between the
All runs were performed on IBM Power5 8-way 1.90GHz p575 pores. This is a consequence of the logic in the grain-based algo-
machines, which are operated by LSU's high-performance com- rithm, which uses the Delaunay tessellation of the grains as seed
puting center. The grain-based algorithm required -1 hr of run locations for potential pores. In contrast, the voxel-based algorithm
time, while the voxel based network required -24 hrs to run. searches for pores from each voxel location, which results in the
identification of many more local extrema (i.e., locations of maxi-
Morphologic Parameters. Fig. 5 contains an image of the grain mum inscribed spheres). Although the difference in pore numbers
reconstruction process for sample A, which is included for general is large, it is still not possible to declare one network more correct
interest. Shading is assigned randomly to grain numbers to give a than the other because discretization of the pore space is an arbi-
visual indication of the discretization into individual grains. (The trary process (within reason of course). The key is that in both
missing piece at the bottom left corner is an artifact caused by the cases the network is a one-to-one mapping of the pore structure,
graphics software.) and in both cases the pores are defined using the same rigorous
In previous tests of the algorithm on tomography images of definition: a local maximum in the distance function d{x,y,z). The
unconsolidated sands, two problems with the grain-reconstruction most effective illustration of these points is to examine the two

\T£ June 2008 SPE Journal

 TABLE 5—QUANTITATIVE RESULTS FROM NETWORK MODELING
OF THE SANDSTONE IMAGES

Sandstone A Sandstone B

Grain-Based Network Voxel-Based Network Grain-Based Network

WG 2,334 2,185
Da (^m) 101 100
NP 10,768 65,574 9,063
Dp (fjm) 34 18 39
Dprium) 31 26 38
LPT (//m) 77 35 92.8
Z 3.21 3.16 3.15
£ 0.186 0.187 0.160
K„ (mD) 410 430 1630
K„ (mD) 530 570 850
KB (mD) 320 360 631

networks side by side, as in Fig. 7, and note that, despite differ- lation, and brighter shades indicate higher flowrates. This graphic
ences in the details, the trends in network structure are the same shows clearly the heterogeneity in the flow pattern that results
(e.g., the appearance of large pores at the same locations; gaps in from the pore morphology, and demonstrates the conventional
the network structure at the same locations, and so on). wisdom that relatively few pores carry a large fraction of the fluid.
The other notable point from the data presented here is the This graphic is also an effective illustration of the rationale for
relatively small difference between the pore-throat diameters and using network modeling, as opposed to simpler effective medium
the pore diameters, which is a consequence of the more tube-like or bundle-of-tubes models. Although these simpler models may
pore structure that can occur in consolidated sands. In fact, in the predict correct permeabilities given the proper pore statistics, they
voxel-based network, the average pore size is smaller than the do not capture the flow heterogeneity that contributes strongly to
average throat size. Although this is counter to the traditional pore processes such as solute dispersion.
and pore-throat model, it is a consequence of the high density of
pores in the second network. The void space comprises strings of Discussion
largely overlapping pores rather than distinct pores connected by A grain-reconstruction algorithm (originally developed for uncon-
long throats; the pore-throat diameter is simply the size of the solidated sands) has been applied to consolidated sandstones. With
channel at the point where two inscribed spheres overlap, and thus additional development and testing, we believe that this algorithm
is not necessarily smaller than the adjacent pores. can be used as a tool for quantitative analysis of the granular
structure in sandstones. This paper uses a grain-reconstruction pro-
Flow Modeling. The networks were used to model single-phase cess as a template for generating a physically representative net-
low-Reynolds-number flow of a Newtonian fluid. The flow mod- work model of the pore structure. The main advantages of this new
eling was performed after cropping 100 /xm from all sides of the approach are that the network is tied to a correct characteristic
networks so as to avoid edge effects that are caused by the bound- scale for the materials (i.e., the grain scale), the algorithm is more
ary of the volume data. than an order of magnitude faster (compared to an equivalent
A pressure gradient was imposed in one of the coordinate di- voxel-based algorithm), and the resulting networks are less depen-
rections and the resulting volumetric flowrate was computed. Per- dent on voxel resolution.
meability was then calculated using Darcy's law (see Table 5). Future research should examine how different network struc-
This process was repeated for all three coordinate directions de- tures (i.e., two different networks that are intended to represent the
fined by the imaging volume. Note that Kyl does not imply orien- same material) affect the modeling of transport phenomena. To
tation with respect to a bedding plane; because the sample was begin studying this issue, we ran the current algorithm in two
taken from a small cutting, no attempt was made to align the ways: as a voxel-based algorithm and as a grain-based algorithm.
tomography image with the geologic strata. These two approaches resulted in dramatically different network
The ability to obtain dimensional permeability values is a con- structures: the most obvious difference is in the number of pores,
sequence of using physically representative networks. That is, the but this in turn affects most other parameters, including coordina-
network is a map of a specific-sized region of the sandstone, which tion number, throat length, and throat conductivity. Interestingly,
allows for the computation of dimensional volumetric flowrates despite the fact that the voxel-based network contained six times
and cross-sectional areas for flow. Additionally, because of the the number of pores, the average permeability difference (for the
one-to-one mapping, the model has no adjustable or scaling pa- three directions) is only 7.5%. This result is encouraging because
rameters. [The pore-throat hydraulic conductivities are determined it means that the physically representative network generation pro-
using the formulas originally developed for sphere packings cess (which is free of adjustable parameters) is working as in-
(Thompson and Fogler 1997) and therefore are not treated as ad- tended with respect to single-phase permeability. Nonetheless, we
justable parameters.] expect that multiphase simulations will be more sensitive to factors
The similarity in permeabilities for the two different networks such as pore coordination number because current multiphase
created from sandstone A may be surprising, considering the dif- models rely more on rule-based algorithms compared to the single-
ferent network structures. However, this fact is again a conse- phase models. This is a topic of ongoing research.
quence of using physically representative networks: in the network We conclude by commenting on the issue of image resolution.
with fewer pores and pore throats, the throats are longer and there- From our experience with grain-based modeling, we suggest a
fore exhibit larger hydraulic resistance. Put simply, everything minimum of 10 VPD in order to extract reasonably good statistics
comes out in the wash. for the solid-phase structure in unconsolidated materials (Thomp-
Fig. 8 is a grayscale mapping of single-phase flow through the son et al. 2006). However, this value (10 VPD) may need to be
same region of sandstone A shown in Fig. 7b (the voxel-based significantly higher for pore networks. Consider that when three
network). The pressure gradient is from left to light for this simu- spheres are placed in contact to form a pore throat (i.e., the ge-

June 2008 SPE Journal 173

 ieoo

1400 •

1200 •
'000

ooo
•
600

400
1
200
•
n
20 40
mlii..-
60 80 100 120 140 160 180
Inscribed pore diameter (urn)
2600

2000

I 1000

500

5 0 5 10 15 20 25 3
Pore coordination number

Fig. 7—Sections of two different networks generated from the

same tomography image: (a) created using the grain-based al-
gorithm; (b) created using the voxel-based algorithm.

sortium funded by BP and Chevron, Geophysical Consortium and

Center for Lithospheric Studies, and selected for analysis using
funding from the US Department of Energy, contracts DOE/SC/
50 100 150
Inscribed throat diameter (urn) 15167-1 to LSU and DE-FG0301ER15166 to UTD. This material
is based in part upon work supported by the National Science
Fig. 6—Morphology of the pore space in sandstone A: (a) pore- Foundation under Grant No. EAR-0207788 (CW) and ACS PRF
size distribution (inscribed); (b) pore-coordination-number dis- grant 41169-AC 9 (CW). Use of the Advanced Photon Source was
tribution; (c) pore-throat-size distribution (inscribed). supported by the US Department of Energy, Office of Science,
Office of Basic Energy Sciences, under Contract No. W-31-109-
ometry In a rhombohedral packing), the inscribed diameter of the Eng-38. This research was conducted using high-performance
passage is equal to 0.155Z?sphere. Hence, a 10 VPD image will computational resources provided by Louisiana State University
provide only 1-2 voxel resolution at a pore throat, which is the (www.lsu.edu/hpc).
crucial area for flow modeling. This issue is exacerbated when a
particle size distribution exists, or with consolidation. Table 3 from References
this paper provides a good starting point for assessing the errors in
Al-Raoush, R.I. and Willson, C.S. 2005a. A Pore-scale Investigation of a
pore parameters that stem from voxel resolution issues.
Multiphase Porous Media System. / Contain. Hydrology 77 (1-2):
Acknowledgments 67-89. DOI: 10.1016/j.jconhyd.2004.12.001.
The core samples were obtained in a study funded by the Univer- Al-Raoush, R.I., and Willson, C.S. 2005b. Extraction of physically realistic
sity of Texas at Dallas Quantitative Sedimentology Research Con- pore network properties from three-dimensional synchrotron X-ray mi-

17-1 June 2008 SPE Journal

 front, Wall Creek Member, Frontier Formation. Wyoming. AAPG Bui
letlnto (2): 191-214.
Liang, Z., Ioannidis, M.A., and Chatzis, I. 2000a. Geometric and topolog-
ical analysis of three-dimensional porous media: Pore space partition-
ing based on morphological skeletonizaUon. J. Colloid Interface Sci.
221 (1): 13-24. DOI: 10.1006/jcis.l999.6559.
Liang, Z., Ioannidis, M.A., and Chatzis. I. 2000b. Permeability and elec-
trical conductivity of porous media from 3D stochastic replicas of the
microstructure. Chem. Eng. Sci. 55 (22): 5247-5262. DOI: 10.1016/
SOO09-2509(00)00142-l.
Llndquist, W.B.. Lee. S.M.. Coker, D.A., Jones. K.W., and Spanne. P.
1996. Medial axis analysis of void structure in three-dimensional to-
mographic images of porous media. J. Geophys. Res. 101 (B4): 8297-
8310. DOI: 10.1029/95JB03039.
Lindquist, W.B.. Venkatarangan, A„ Dunsmuir, J., and Wong, T.-F. 2000.
Pore and throat size distributions measured from synchrotron X-ray
tomographic images of Fontainebleau sandstones. / Geophys. Res. 105
(B9): 21509-21527. DOI: 10.1029/2000JB900208.
Lohman, G. 1998. Volumetric Image Analysis. New York City: John Wiley
& Sons.
Lopez, X., Valvatne, P.H., and Blunt, M.J. 2003. Predictive network mod-
eling of single-phase non-Newtonian flow in porous media. J. Colloid
Interface Sci. 264 (1): 256-265. DOI: 10.1016/S0021-9797(03)00310-2.
Luchnlkov, V.A.. Medvedev, N.N., Oger. L„ and Troadec, J.-P. 1999.
Voronoi-Delaunay analysis of voids In systems of nonspherlcal par-
Fig. 8—Representation of single-phase flow in a small section
ticles. Phys. Rev. £59 (6): 7205-7212. DOI: 10.1103/PhysRevE.59.
of the voxel-based network from the tidally-reworked sandstone
(sample A). 7205.
Mellor, D.W. 1989. Random close packing (RCP) of equal spheres: struc-
ture and implications for use as a model porous medium. PhD thesis.
crotomography images of unconsolidated porous media systems. Jour- Milton Keynes, U.K.: Open University.
nal of Hydrology 300 (1-4): 44-64. DOI: 10.1016/j.jhydrol.2004.05. Oh, W. and Lindquist, B. 1999. Image thresholding by indicator kriging.
005. IEEE Transactions on Pattern Analysis and Machine Intelligence 21
Al-Raoush. R.I.. Thompson, K., and Willson, C.S. 2003. Comparison of (7): 590-602. DOI: 10.1109/34.777370.
network generation techniques for unconsolidated porous media. Soil 0ren, P.-E., Bakke, S., and Arntzen, O.J. 1998. Extending Predictive Ca-
Sci. Soc. Am. J. 67: 1687-1700. pabilities to Network Models. SPEJ 3 (4): 324-336. SPE-52052-PA.
Bakke, S. and 0ren, P.E. 1997. 3-D Pore-Scale Modelling of Sandstones DOI: 10.2118/52052-PA.
and Flow Simulations in the Pore Networks. SPEJ 2 (2): 136-149. Patzek, T.W. 2001. Verification of a Complete Pore Network Simulator of
SPE-35479-PA. DOI: 10.2118/35479-PA. Drainage and Imbibition. SPEJ6 (2): 144-156. SPE-71310-PA. DOI:
Balhoff, M.T. and Thompson. K.E. 2004. Modeling the steady flow of 10.2118/71310-PA.
yield-stress fluids in packed beds. AIChE J. 50 (12): 3034-3048. DOI: Press, W.H.. Teukolsky, S.A.. Vetterling, W.T., and Flannery, B.P. 1992.
10.1002/alc.l0234. Numerical Recipes In Fortran 77. New York City: Cambridge Univer-
Bryant. S.L.. King. PR., and Mellor, D.W. 1993a. Network model evalu- sity Press.
Reed, AH. et al. 2005. Quantification of sediment properties and geo-
ation of permeability and spatial correlation in a real random sphere
acoustic parameters from pore structure and grain contacts: A micro-
packing. Transport in Porous Media 11 (1): 53-70. DOI: 10.1007/
computed tomography analyisis of SAX04 sands. Proc, International
BFO0614635.
Conference on Boundary Influences in High-Frequency, Shallow Wa-
Bryant, S.L.. Mellor, D.W., and Cade, C.A. 1993b. Physically Represen-
ter Acoustics. University of Bath, Bath. U.K., 5-9 September.
tative Network Models of Transport in Porous Media. AIChE J. 39 (3):
Scheidegger, A.E. 1974. 77ie Physics of Flow Through Porous Media.
387-396. DOI: 10.1002/aic,690390303.
Toronto. Ontario, Canada: University of Toronto Press.
Delerue, J.-F. and Perrler, E. 2002. DXSoil. a library for 3D image analysis
Serlght, R.S., Liang, J.-T., Lindquist. W.B.. and Dunsmuir. J.H. 2002.
in soil science. Computers & Geosciences 28 (9): 1041-1050. DOI:
Characterizing Disportation Permeability Reduction Using Synchro-
10.1016/S0098-3004 (02)00020-1.
tron X-Ray Computed Tomography. SPEREE 5 (5): 355-364. SPE-
Delerue, J.F., Perrier, E„ Yu, Z.Y., and Velde, B. 1999. New algorithms in
79717-PA. DOI: 10.2118/79717-PA.
3D image analysis and their application to the measurement of a spa-
Seright, R.S.. Liang, J.-T., Lindquist, W.B.. and Dunsmuir, J.H. 2003. Use
tiallzed pore size distribution in soils. Phys. Chem. Earth A 24 (7):
of X-ray computed microtomgraphy to understand why gels reduce
639-644. DOI: 10.1016/S1464-1895(99)00093-9. relative permeability to water more than that to oil. / Pet. Sci. Eng. 39
Fart, I. 1956. The Network Model of Porous Media: I. Capillary Pressure (3-4): 217-230. DOI: 10.1016/S0920-4105(03)00064-0.
Characteristics. Trans., AIME 207: 144-159. SPE-574-G. Silin, D.B., Jin. G., and Patzek, T.W. 2003. Robust Determination of the
Finney, J.L. 1970. Random packings and the structure of simple liquids I. Pore Space Morphology in Sedimentary Rocks. Paper SPE 84296 pre-
The geometry of random close packing. Proc. Roy. Soc. Lond. A 319: sented at the SPE Annual Technical Conference and Exhibition, Den-
479-493. ver, 5-8 October. DOI: 10.2118/84296-MS.
Ganl. MR. and Bhattacharya, J.P. 2003. Bed-scale facies architecture of an Sok, R.M. et al. 2002. Direct and stochastic generation of network models
ancient delta lobe deposit of the Wall Creek Member, Central Wyo- from tomograpliic images; effect of topology on residual saturations.
ming. U.S.A. AAPG Annual Convention, Salt Lake City. Utah, 11-14 Transport in Porous Media 46 (2-3): 345-371. DOI: 10.1023/
May. A:1015034924371.
Ioannidis, M.A.. Kwiecien, M.J.. Chatzis, I., MacDonald, I.F., and Dullien, Talukdar, M.S.. Torsaeter, 0„ Ioannidis, M.A., and Howard, J.J. 2002.
F.A.L. 1997. Comprehensive Pore Structure Characterization Using 3D Stochastic reconstruction. 3D characterization and network modeling
Computer Reconstruction and Stochastic Modeling. Paper SPE 38713 of chalk. J. Pet. Sci. Tech. 35 (1-2): 1-21. DOI: 10.1016/S0920-
presented at the SPE Annual Technical Conference and Exhibition, San 4105(02)00160-2.
Antonio, Texas, 5-8 October. Thompson, K.E. 2002. Fast and robust Delaunay tessellation in periodic
Lee, K. et al. 2007. Three-dimensional facies architecture and three- domains. Int. J. Numer. Meths. Eng. 55 (11): 1345-1366. DOI:
dimensional calcite concretion distributions in a tide-influenced delta 10.10O2/nme.558.

June 2008 SPE Journal 175

Thompson, K.E. and Fogler, H.S. 1997. Modeling Flow in Disordered gineering. White holds a BS degree from the University of Okla-
Packed Beds from Pore-Scale Fluid Mechanics. AIChEJ. 43 (6): 1377- homa as well as MS and PhD degrees from Stanford University,
1389. DOI: 10.1002/aic.690430602. all in petroleum engineering. Stephanie L. Nyman is a PhD can-
Thompson, K.E., Willson, C.S., and Zhang, W. 2006. Quantitative com- didate and Assistant Lecturer at the University of Waikato,
puter reconstruction of paniculate materials from mlcrotomography Hamilton, New Zealand. Her geologic specialty is clastic dia-
genesis, with a research interest in carbonate cement within
images. Powder Technology 163 (3): 169-182. DOI: 10.1016/ subsurface hydrocarbon seep systems. She holds an MSc de-
J.powtec.2005.12.016. gree from the University of Texas at Dallas. Janok P. Bhatta-
Valvatne, P.H. and Blunt. M.J. 2003. Predictive Pore-Scale Network Mod- charya is the Robert E. Sheriff Professor of Sequence Stratigra-
eling. Paper SPE 84550 presented at the SPE Annual Technical Con- phy in the Department of Earth and Atmospheric Sciences at
ference and Exhibition. Denver, 5-8 October. DOI: 10.2118/84550- the University of Houston. Following a Natural Sciences and
MS. Engineering Research Council post-doc at the Alberta Geo-
logical Survey in Edmonton, Janok worked for the Bureau of
Economic Geology in Austin; ARCO Research in Piano, Texas;
Karsten E. Thompson is an associate professor of chemical en- and the University of Texas at Dallas before joining the Univer-
gineering at Louisiana State University. His research interests sity of Houston. His research interests include deltaic sedimen-
include computational modeling of porous materials. He holds tology and sequence stratigraphy, the local control of struc-
a BS degree from the University of Colorado and MS and PhD ture on stratigraphy, and reservoir architecture of clastic
degrees from the University of Michigan, all in chemical engi- depositional systems. He has authored or co-authored ap-
neering. Clinton S. Willson is an associate professor in the De- proximately 50 technical papers and over 100 abstracts. He
partment of Civil and Environmental Engineering at Louisiana holds a BSc degree from Memorial University of Newfound-
State University. His research interests include high-resolution, land, Canada, and a PhD degree from McMaster University,
3D imaging of multiphase flow in porous media. He holds a BS Hamilton, Ontario, Canada. Allen H. Reed is a geologist with
degree in aerospace engineering from Pennsylvania State Uni- the Naval Research Laboratory. His research interests are in
versity, as well as an MS degree in environmental health engi- marine sedimentology, including microfocus-computed to-
neering and a PhD degree in civil engineering, both from the mography and image analysis of porous media, and the influ-
University of Texas at Austin. Christopher D. White has been on ence of muddy sediments on dissipation of wave energy. He
the faculty of the Department of Petroleum Engineering at holds a BS degree in oceanography from Humboldt State Uni-
Louisiana State University since 1999. He formerly worked at versity, an MS degree in marine sciences from the University of
Shell Development Company and the Bureau of Economic Southern Mississippi, and a PhD degree in oceanography and
Geology. His interests include reservoir geomodeling and en- coastal studies from Louisiana State University.

170 June 2008 SPE Journal