Вы находитесь на странице: 1из 27

J. Non-Newtonian Fluid Mech.

102 (2002) 447473

Models for flow of non-Newtonian and complex


fluids through porous media
J.R.A. Pearson , P.M.J. Tardy
Schlumberger Cambridge Research, Cambridge CB3 0EL, UK

Abstract
This article first provides a brief and simple account of continuum models for transport in porous media, and of
the role of length scales in passing from pore-scale phenomena to Darcy continuum scale representations using
averaged variables. It then examines the influence of non-Newtonian rheology on the single- and multi-phase transport parameters, i.e. Darcy viscosity, dispersion lengths and relative permeabilities. The aim is to deduce functional
forms and values for these parameters given the rheological properties of the fluid or fluids in question, and the
porosity, permeability, dispersion lengths and relative permeabilities (based on Newtonian fluids and equivalent
capillary pressures) of the porous medium. It is concluded that micro-models, typically composed of capillary networks, applied at a sub-Darcy-scale, parameterised using data for flows of a well-characterised set of non-Newtonian
fluids, are likely to provide the most reliable means. 2002 Elsevier Science B.V. All rights reserved.
Keywords: Darcy viscosity; Newtonian fluids; Transport parameter

1. Scope and preface


This article does not seek to prove a new result or to justify a new technique. Nor is it intended as a
comprehensive review. It is restricted to presenting in simple form selected parts of the most important
approaches that have been used to predict flow behaviour in porous media, giving special attention to the
effects of non-Newtonian rheology. All of the seminal work reported here has been done by others and
reference is made in the short bibliography below to a few of the textbooks and papers that are well-known
and have been found helpful in this work. The indulgence is craved of any other authors whose papers or
texts have not been referenced, particularly if they feel that their contributions deserve special mention.
The subject is important in a variety of contexts. This contribution has been prompted by efforts to
design treatment fluids and intervention processes to control the flow of gas, oil and water in hydrocarbon
reservoirs. The difficulty of conducting realistic experiments to model every option, and hence the need to

Corresponding author. Tel.: +44-1223-3509-27; fax: +44-1223-3530-99.


E-mail address: jrap@pearson.co.uk (J.R.A. Pearson).
Dedicated to Professor Acrivos on the occasion of his retirement from the Levich Institute and the CCNY.
0377-0257/02/$ see front matter 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 0 2 5 7 ( 0 1 ) 0 0 1 9 1 - 4

448

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

provide computational models for prediction, is overwhelming. Although even qualitative models based
on a basic understanding of the dominant mechanical, physical and chemical processes involved can be
useful, reliable quantitative models are found necessary for successful applications. This article concerns
criteria for selecting such quantitative models.
2. Introduction
Many of the fluids used in the oil industry to treat and stimulate reservoirs are significantly nonNewtonian: in differing degrees they display shear-dependence of viscosity, thixotropy (or rheopexy)
and elasticity. They are also multi-component and chemically complex, and can interact with the porous
formations (sands, carbonates, clays or hard rock) through which they are pumped and with the other
fluids with which they come into contact within the pore structure. These structures are themselves
complex, displaying inhomogeneities on a wide range of length scales, from 106 to 103 m. For quantitative
purposes, good mathematical models of the transport processes involved are necessary. This has in the
past been based on continuum representations.
2.1. Darcys law and permeability: Newtonian fluids
The starting point for such continuum models is typically Darcys law, relating gradients of pore fluid
pressure to the pore fluid flux within a robust, essentially rigid, porous matrix. Symbolically, this is written
as
 


k
k
u=
(p g)
p
(2.1)

where u, is the Darcy velocity, is a mean velocity of the pore fluid relative to the matrix, p the mean
absolute pressure of the pore fluid, and are the density and viscosity of the fluid, g the gravity (or
other field force) and k is the permeability of the matrix, with dimensions of length squared. Various
assumptions have been made in this much used model:
(i) a suitable averaging procedure has been prescribed to obtain the Darcy-scale u and p from the actual
point values of fluid velocity u and pressure p within the matrix;
(ii) the fluid is Newtonian;
(iii) the matrix is isotropic;
(iv) the flow is slow (in formal terms Rep = |u|k 1/2 / 1) and so the same linear law applies at each
instant.
For extended media, it is natural to suppose that u, p and k can become functions of position x and time t.
If both fluid and matrix are incompressiblean approximation that will be adopted for simplicity
throughout this accountthen the mass conservation law becomes
u=0
and the field equation for pressure is
 
k
ln
p + 2 p = 0

(2.2)

(2.3)

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

449

If the porous medium and the pore fluid are homogeneous, and this was often assumed in early work, then
the reduced pressure p obeys the harmonic equation. Boundary conditions on either pressure or velocity
lead to well-posed problems, which can be solved relatively easily.
It is known that real porous media are not homogeneous, though in practice the spatial variation of
k(x) is at best inferred from a limited series of tests on small samples of the porous medium in question.
It is, therefore, conventional to regard k(x) as a stochastic function of position, describable by a limited
number of numerical parameters defining its statistics (those familiar with classical fluid turbulence will be
familiar with the use of one, two and three point correlations and covariances in this context). Solutions
now become statistical (probabilistic) rather than exact, and the subtleties of predicting consequences
(interpreting results) for given boundary conditions have led to a large and growing literature.
Reservoirs are normally highly stratified and so the scalar k has to be replaced by a tensor permeability
k. Darcys law becomes
u = k

(2.4)

with consequential complication of Eq. (2.3). This issue will not, again for simplicity, be considered
further here.
2.2. Dispersion: Newtonian fluid
If a tracer is present, at some averaged concentration c(x, t), in the pore fluid, then it is well known
experimentally that it is dispersed as well as advected by the mean flow field u. This dispersion is
where u is the mean fluid velocity
attributed to the pore-scale variations in the flow field, i.e. to u u,
within the pore space and differs from the Darcy velocity u by a factor 1/m, where m is the porosity, the
average fraction of total volume occupied by pores. In our case of a homogeneous isotropic medium, the
advectivediffusive equation for c is conventionally written as
c
+ u c = (D c)
t

(2.5)



uu
uu
+ I|u|
D ||
|u|
|u|

(2.6)

m
where

|| (x) and (x) being the dispersion lengths parallel and perpendicular to the Darcy velocity vector and
where molecular diffusion is neglected. It is worth noting that the dispersion term, on the right-hand side
of Eq. (2.5), is linear in u as is the advection term on the left-hand side. This means that the change in
the tracer field depends not on the flow-rate, but only on the displacement. For Newtonian fluid flows,
|| and are regarded as further parameters characterising the porous medium.
2.3. Multi-phase flow: Newtonian fluids
If two or more mutually immiscible fluids are present within the porous matrix, they are treated as
interpenetrating fluids (simultaneously present at x in some average sense), and the phase fractions
within the pore space are called their I phase saturations s(I ) . They are kept separate at the pore-scale by

450

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

internal interfaces across which interfacial tensions act. These lead, in some average sense, to capillary
(I,J )
pressure differences pcap
= p(I ) p (J ) between phases I and J: clearly, these average quantities are
functions not only of the porous medium in question, but also of the interfacial tensions between the phases,
the wetting characteristics (usually described by contact angles for any given pair of fluids meeting along
a contact line on the pore surface) of the internal pore surfaces, the current saturations and, as has often
been observed, the saturation history within the medium.
If one or more of the phases are driven by pressure gradients p (I ) , then corresponding averaged
velocities u(I ) will arise. For slow flows, the relation between them is written formally, by analogy with
Eq. (2.1), as
u

(I )

)
kk(I
rel
= (I ) (p(I ) (I ) g)

(2.7)

(I )
which, in
where all the complexity of the situation is wrapped up in the relative permeabilities krel
the absence of any experimental or theoretical results, have to be regarded as functions of at least the
(I,J )
saturations s(I ) , the capillary pressures pcap
, the viscosities (I ) and the flow-rates u(I ) , as well as of the
pore geometry and nature of the formation.
The mass balance Eq. (2.2) is decomposed into phase relations

s (I )
+ u(I ) = 0
t

(2.8)

It is not surprising that the general situation is so complex that severe approximations are always made
when modelling multi-phase flows, and that these are chosen to suit the particular flow process involved.
Thus, adjustments are made specifically to cover imbibition of water into oil-rich media (typically, cores),
drainage of water from water-rich media, partially miscible phases, and flooding of depleted reservoirs
by water or gas. Almost all calculations are based on locally one-dimensional flow; if capillary pressure
gradients are neglected, as they often are at large-scales, the hyperbolic nature of equations (2.7) and
(2.8) means that shocks in s(I ) can arise: these are typical of water floods and so lead to the tracking of
the fronts so defined; flow lines remain approximately perpendicular to pressure and saturation lines.
2.4. Flow of non-Newtonian fluids in porous media
We naturally wish to extend the basic flow Eq. (2.1) to cover fluids with complex rheology. The method
used for defining u and p in terms of any pore-scale flow will not be altered by the rheology of the
fluid; it is now assumed, and all detailed pore-scale models confirm, that all non-Newtonian effects can
be subsumed in a suitable definition of , which will continue to have the dimensions of Newtonian
viscosity. Much of the rest of this article (Section 4, Appendix A) investigates the way bulk rheological
behaviour of the pore fluid will affect this eff ; put otherwise, how might we hope to derive the value
of eff from a set of bulk rheological measurements, given k and other readily available data? Next, we
consider in Section 5 extending the dispersion relations (2.5) and (2.6) with the hope of redefining ||
and to cover the change in flow field associated with non-Newtonian rheology.
In Section 6, we discuss what changes have to be made to any set of relations found adequate for
)
multi-phase flow of Newtonian fluids. In particular, is it sufficient to use the value of (I
eff found relevant
(I )
for single-phase flow instead of the Newtonian viscosity (I ) in the phase flow (Eq. (2.7)), keeping krel

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

451

(I,J )
unaltered and estimating pcap
in the same way as for Newtonian fluids? In Section 7, we consider other
effects, such as adsorption and desorption of viscosifying agents, boundary slip, adhesive layers and mass
interchange between phases.
Few firm general conclusions are reached in Section 8. The best that can be hoped for at this stage is
to set out some guidelines for those seeking to characterise flow behaviour of non-Newtonian systems in
porous media and to analyse particular flow situations.

3. The role of length scales in characterising porous media and flow through them
3.1. Homogeneity in continuum models
How do we derive continuum models that do not require an exact knowledge of the detailed structure
of any given porous medium? What do we mean when we say that a porous medium is homogeneous?
Can we provide a simple and reasonable criterion for homogeneity?
To answer these questions, we start by noting that any such medium when dry (containing no fluid in
its pores) will consist of one continuous region of solid, the matrix, and regions of void, the pores (for
simplicity, here we take the pore region to be continuously connected); these two regions will be separated
by a surface whose configuration (geometry) defines both matrix and pore spaces. We now take a sample
of characteristic dimension lsample (e.g. the diameter of a spherical sample, the side length of a cube or
more generally the cube root of its volume) and refer points x within the sample to an origin 0 fixed within
the sample. We then associate with each point x a presence function, I1 (x), which has the value 1 if x is
in a pore and 0 if x is in the matrix. The average porosity, a primary characteristic mentioned in Section
2.2, is then given by

sample I1 (x) dx
msample C1 sample =
(3.1)
volumesample
We can also define a joint presence function
I2 (x1 , x2 ) = I1 (x1 )I1 (x2 )
and hence, a mean correlation function

sample I2 (x1 + r, x1 ) dx
C2 sample (r) =
volumesample msample

(3.2)

(3.3)

The medium is isotropic and so we can take C2 to be a function of |r|. The difficulty that arises in defining
I2 when x1 + r lies outside the sample will be addressed further later, and is accommodated formally here
by using the notation sample . It is relatively easy to see that higher-order correlations Cn can be defined
similarly, involving (n 1) separations rj . All the Cn are dimensionless.
There is no unique way of defining a characteristic size for the pore space (pores). The simplest to
envisage is to consider the distances between intersections of straight lines passing through the sample
with the pore surfaces, and to take the average length of those segments corresponding to pores; this
defines one pore size, lpore . A distribution of pore lengths is correspondingly obtained, which would be

452

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

formally related to the Cn . If


0 =

lpore
lsample

1,

we have two widely separated length scales. The variable


x
x pore =
lpore
describes rapid variations of I1 on the pore-scale, while
x
x sample =
lsample

(3.4)

(3.5)

(3.6)

defines variations on a sample scale. By definition C1 is a constant, but C2 is a variable in terms of the
argument rpore . We expect C2 to asymptote rapidly to msample for rpore  1, and so we might expect

 

 C2 (r)

1 dr
(3.7)
lcorr =
m
sample
0
to be of order lpore (this lcorr could be used as an alternative definition of lpore ).
We now consider the sample to have been taken from a larger physical block of an extended porous
medium, whose own length scale is lblock . If
1 =

lsample
1,
lblock

(3.8)

we are able to take many separate samples from the block. If lpore , lcorr (or C2 (r)), and all higher correlations
prove to be independent 1 of sample position in the block, then our block is said to be homogeneous (and
evidently isotropic) and vice versa.
Now clearly, as the ratio 0 tends to unity by taking smaller and smaller samples, the sample will appear
less and less homogeneous, while from experience in large sedimentary rock masses, by taking lblock
progressively larger, the block will appear to become inhomogeneous. The largest amount of information
about the porous medium, treated as a continuum, will be obtained by using the widest range of length
(min)
(max) 2
scales, given by the pair {lsample
, lblock
} , that leads to local homogeneity. On the other hand, for many
problems the extent of the medium involved is so great that little information is available on the scale
(max)
of lblock
and so for calculation purposes the relevant continuum properties are required on a much larger
scale. The issue of how to obtain formal relations between continuum parameters on different length
scales is an important one, generally called the problem of up-scaling, but will not be addressed here.
3.2. Averaging in porous media
We now return to the question of averaging in porous media to obtain continuum parameters and
variables. In a formal sense, any sample can be regarded as a particular, geometrically defined, region of
1

Strictly speaking, we must expect some statistical variation, but we can accept a variance that is at most of the order of the
larger of 0 or 1 , or even of a fractional power of these small quantities.
2 (min)
lsample is sometimes referred to as the representative effective volume (REV).

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

453

the medium, which remains unaffected by any of the fluids flowing through it. A prime requirement of any
continuum model is that the parameters and variables defining it should be continuous variables spatially.
This is easily achieved formally if the volume (sample, REV) over which averages are taken is thought
to move continuously through the medium. Thus, a generalised form of Eq. (3.1) can be written as

I1 (xpore )W (xpore )A(xpore ) dx pore

A(xsample ) =
(3.9)
I1 (xpore )W (xpore ) dx pore
where the origin of the co-ordinate system x pore is at x sample and W (xpore ) is a suitable weighting function.
Here, the same symbol, A in this case, is used for volume-averaged and pore-scale variables, the associated argument determining its meaning. The difficulty mentioned earlier in connection with Eqs. (3.2)
and (3.3), and by extension with Eq. (3.7), is now removed because x2 is now defined outside the averaging volume. The situation is different if experimental evaluations are undertaken. Physical samples
are unique, and cannot be moved continuously, so at best a discretised representation (usually, sparse in
practice) can be obtained for real media; however, this form of discretisation happens to coincide with
most computational representations for specific media, e.g. reservoirs or geological formations generally,
and so the passage from given experimental data to computational model can in principle be direct.
A difficulty arises at the continuum scale when interpreting boundary conditions, which formally and
physically are applied at pore-scale, i.e. to A(xpore ) over a two-dimensional surface. One way out of the
difficulty is to apply them formally at the sample scale, i.e. by writing them in terms of A(xsample ) over
the same surface; it should not be presumed, however, that these boundary conditions would be identical
at the two scales.
An associated difficulty arises when seeking to define the permeability k, used in Eq. (2.1), formally;
we have assumed, rather than proved, that it will be a material constant, at least over length scales large
compared with lsample and small compared with lblock ; formally, we treat k as a slowly-varying function of
x sample . If we have a physical sample, say cubic of side Lsample , we can apply pressure differences across
two of its faces and measure the flux Q. We can replace p by p/L
sample and u by Q/L2sample ; knowing ,
k is obtained. If, however, we seek to derive it independently by solving the Stokes equations of motion for
a geometrically fully-defined pore-scale model, we soon realise that the pore spaces outside the sample
volume have to be considered when solving for the flow within the sample, particularly, close to the sample
boundary; in order to carry out the averaging (Eq. (3.9)) for u and p, an extended solution is required.
Some authors have sought to avoid this difficulty by prescribing spatially periodic media, where the
sample represents one period along each of three principal axes. This has the added advantage that the
medium becomes by definition homogeneous; the constant k (or its more general counterpart k) in Darcys
law can be calculated for a flow field u(xsample ), p(x
sample ) which is uniform with respect to xsample ;
equally all the Cn (r1 pore , . . . , r(n1) pore ) will be independent of position xsample . In moving to the larger
scale xblock , it has to be supposed that the assumption of periodicity is only asymptotically valid as1 0,
though this can only be done formally, since in any real context lpore and lblock are given and so  1 and  0
must range between finite limits.
4. The Darcy viscosity of non-Newtonian fluids
In practice, the permeability of porous media is obtained by direct measurement of pressure-drop and
flow-rate using cylindrical cores from blocks of coherent porous media or lightly compacted cylinders of

454

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

unconsolidated granular media. It has been verified that the permeability obtained by using a Newtonian
fluid with one viscosity is the same as that obtained using another Newtonian fluid of different viscosity,
and that the pressure-drop and flow-rate are linearly related, thus confirming that Stokes law applies at
the pore-scale. This measurement of permeability requires but a single steady-state measurement on a
fully-saturated cylindrical sample of known length and cross-sectional area, sealed along its sides with
entry and exit on plane ends orthogonal to its axis. If the fluid flowing through the medium of known
permeability is now a non-Newtonian fluid (of known bulk rheology), how do we derive an equivalent of
Darcys law?
4.1. Generalised Newtonian fluids
These fluids, often called pseudo-plastic because most are shear-thinning, are such that the (Newtonian)
stress/rate-of-strain law
T = E

(4.1)

with T the deviatoric stress tensor and


E = ( v + v T )

(4.2)

(twice) the rate-of-strain tensor, can be retained, but where the viscosity
GN = (|E|)

(4.3)

becomes a function of strain rate. This representation (with properly regarded as a characteristic
parametric function of the fluid concerned) is suitable for solution of a pore-scale flow, but it is by no
means clear how |E| = E eff should be defined for use of GN in the Darcy-scale relation (2.1). On purely
dimensional grounds |u|/k1/2 is seen to be a rate-of-strain, and many authors have supposed that
Eeff = |u|k 1/2

(4.4)

with a near-universal constant would be a suitable approximation. On general grounds, we can suppose
that any relation (4.3) will yield a pair of characteristic values c , Ec and that Darcys law would become
u=

k
p
c

(4.5)

with a joint matrix/fluid constitutive function of the reduced shear rate


E =

|u|
k 1/2 Ec

(4.6)

Viewed as a material function, would depend upon the detailed nature of the pore geometry, to allow for
the possibility that two separate media having the same permeability would distinguish between different
generalised Newtonian fluids.
An obvious way to evaluate would be to measure its value for a given sample over a suitable range
of Darcy velocities (or pressure gradients), but this is seen to be impractical for the range of media, fluids
and flow-rates involved in many design procedures.
We are led to consider a series of model media that will reproduce relevant features of real porous media
as far as permeability is concerned. As we have restricted ourselves to isotropic and homogeneous media,

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

455

simple one-dimensional tubular models suggest themselves. In Appendix A (Section A.2), we consider
bundles of cylindrical tubes. It immediately becomes clear that it is convenient to replace relation (4.1)
by a flow-rate Q versus pressure-drop P relation for a tube. By considering only circular pipes of length
L and diameter D such a relation is readily obtained by integration, given GN (E) (it is interesting to note
that the Q, P relation is obtained directly in capillary rheometry, and that GN is derived from it using the
Rabinowitz differential relation, so that the former could in that case be regarded as more basic). If we
take a Carreau model (see [13], index) for the viscosity function (and this covers a wide variety of elastic
and inelastic fluids in uniform steady simple shear flow) we can show that the breadth of the transition
region from low to high Newtonian asymptotes is wider for a pipe (Q, P relation) than for uniform shear
(GN relation) and wider still for a bundle of pipes having a distribution of diameters. Thus, in general,
it can be shown that the apparent viscosity c will depend on
(i) the parameters in the model for GN ;
(ii) the geometrical parameters defining each tube;
(iii) the statistical parameters defining the distribution of tube sizes and shapes.
We also consider (Section A.3) bundles of identical tubes with diameters varying along their length.
Not surprisingly, we can show that, for model systems having the same values for m and k, significant
variations can arise in the predicted apparent viscosity. There is an advantage in this, as we can hope to
represent any given medium, as far as Darcy pressure-drop alone is concerned, by a very simple capillary
model with very few parameters, for all non-Newtonian fluids. We shall see later that this is not necessarily
sufficient to represent other characteristics of the medium, such as dispersion/residence-time distribution.
More elaborate network models, typically on a square or cubical grid, have been used by several authors,
often in connection with detailed modelling of two- or three-phase flow. Some of them have nodes of
zero volume connected by circular pipes, and are thus obvious extensions of the tubular models described
above; others have most of the porosity contained in pores connected by throats which account for all of
the pressure-drop. We shall not attempt in this article to generate any general results for either of these
network models, though we shall provide in Section A.4 some preliminary results for power-law fluids in
a square grid network. It is worth noting one curiosity: a regular square or cubical grid of equal capillaries
behaves isotropically in the case of Newtonian fluids, but not for shear-thinning fluids; thus, when some
authors line up an edge of the square with the flow direction, and others line up a diagonal, they will obtain
different predictions for most non-Newtonian fluids. This cautionary tale suggests that the geometry of
any real porous medium plays a greater role in determining flow behaviour of complex non-Newtonian
fluids than of linear fluids.
The most realistic model geometries are described topologically rather than metrically. Pores are
connected to one another by a variable number of links that can have different lengths; a subset of these
pores are associated with the inlet and outlet points or faces. For calculation purposes, they are no more
inconvenient than regular grids.
4.1.1. The power-law fluid
Many fluids display a significant power-law region in their viscosity function (4.3), as do the Carreau
models. The approximate form used for analysis over this region can be written as

GN = c

E
Ec

n1

n>0

(4.7)

456

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

We now show that for any porous medium the constitutive function defined in Eq. (4.5) is given by
 n1
E
= Cn
(4.8)
Ec
where Cn is a constant characteristic of the porous medium and of the power-law index n. The proof
supposes that the steady velocity field u(xpore ) and stress field T(xpore ) + p(xpore )I is known for some
particular value of (Q, P) in a cylindrical sample of the medium. We then note that the velocity field fu
and associated stress field f n (T + pI) satisfies the equations of motion at pore-scale, this corresponding
to (fQ, fn P) at the sample scale. Eq. (4.8) follows directly.
4.2. Thixotropic fluids
We use the term thixotropic to describe an extension of generalised Newtonian fluids in which the
viscosity is determined by the history of the rate-of-strain (many real fluids display such character when
subject to non-constant shear rate in a shear viscometer). Generically and formally, this can be written as
t

thix (t) = GN (E(t)) M (E(s))


s=

(4.9)

where M is a dimensionless functional over past times s. The simplest evolutionary equation for M is
1M
dM
=
dt
t1

(4.10)

with t1 a relaxation time for the fluid in question. An alternative form, supposedly more physically
intuitive, has been written as


0
thix
dthix (t)
=
1
g(E)
(4.11)
dt
t1
0
where 0 is the zero-shear-rate viscosity and g(E) is a dimensionless destruction function. The corresponding steady-shear-flow value of thix , GN , would be given by
GN (E) = 0 {1 g(E)}

(4.12)

from which Eq. (4.10) is trivially re-derived.


This rheological model can be used with the tubular models for the porous medium. It is at once
apparent that the (Q, P) relation in any physical sample will depend, even in steady-state flow, upon the
inlet value of thix and the dimensionless residence-time
tres = mLsample

Asample
Qt1

(4.13)

If tres 1, then the fluid will appear to be Newtonian (assuming that it has been perfectly mixed before
entry to the tube), with the entering value of viscosity, while if tres  1, it will behave as a generalised
Newtonian fluid in an infinitely long domain. For intermediate values, the apparent sample viscosity
will be dependent upon the flow-rate and the sample length. In terms of the discussion of Section 2, the
(min)
response within the averaging volume characterised by lsample
will depend upon other averaging volumes

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

457

(max)
(upstream). For a medium that is homogeneous up to the scale lblock
, a criterion for treating a thixotropic
fluid as a generalised Newtonian fluid is thus:

ml(max)
block
1
uDarcy t1

(4.14)

However, this is not as simple a statement as it seems, because it cannot be assumed that the viscosity
represented by GN in Eq. (4.9) will be directly relevant in Darcy flow, nor that the relevant Darcy
form will be as simply obtained as Eq. (4.8) even when the steady-state viscosity is given by Eq. (4.7).
The difficulties involved in predicting the Darcy behaviour are just as great as those mentioned later in
connection with dispersion.
For cases where Eq. (4.14) is not obeyed, one may expect that in Eq. (4.5) will be given by a functional
t

= MDarcy (|uDarcy (xsample )|),


s=

dxsample (s)
= uDarcy (xsample (s))
ds

(4.15)

Given M as defined by Eq. (4.9), and a sufficiently representative tubular model for the porous medium,
MDarcy can in principle be calculated for non-uniform flow fields.
For the particular case of simple bundles of tubes which we can suppose to be of infinite length, it is
relatively simple to consider unsteady flows, so that uDarcy becomes a function of s and not of xsample .
MDarcy then follows fairly simply from M(E), as explained in Section A.2. Whether such an approach
would bear any relation to reality is another matter.
4.3. Elastic fluids
These are fluids which can be represented neither in the form Eqs. (4.1)(4.3), with a scalar, nor in
the form (4.9). Formally, we can write a functional relation for the rheological equation of state:
t

T(t, a) = T {E(s, a)}


s=

(4.16)

where E(s, a) represents the past history of the rate of strain of the particle that is at a at time t, for a
general memory-fluid. Typical examples are the evolutionary models of Oldroyd, based on the Jaumann
derivative or the upper- and lower-convected Oldroyd derivatives, or the BKZ and Wagner integral models
(see, e.g. [7], p. 345 et seq. and 436 et seq. or [13], index). At the pore-scale, flow fields are strongly
affected by elastic forces which cause the principal directions of stress and rate-of-strain not to be parallel,
as they have to be for pseudo-plastic and thixotropic fluids. However, at the Darcy-scale, the relationship
(4.15) is expected to remain relevant, because all complexities at the pore-scale will be subsumed in the
scalar functional MDarcy .
Most elastic fluids are characterised by spectra of relaxation (and retardation) time scales, which can
be used as in Eq. (4.13) to determine whether history dependence is important or not. If it is not, it is
tempting to use the viscometric function for viscosity as defining an equivalent generalised Newtonian
fluid for obtaining the effective Darcy viscosity at any flow-rate. Experimental measurements have not
been conclusive: on theoretical grounds, it can be argued that the very high differences-of-normal-stress
and Trouton ratios associated with polymeric fluids will produce increasing values of apparent Darcy
(based on MDarcy ) when flow channels in the porous medium are of rapidly-changing cross-section; in
some cases, such sharp rises at the highest flow-rates used have been observed and called shear-thickening

458

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

by analogy with the behaviour of fluids, such as suspensions, in which the steady-state-shear viscosity
similarly rises at high shear rates; however, in other cases, the apparent Darcy appears to decrease as
predicted by use of a viscometric (Eeff ) to describe the elastico-viscous fluid.
These two apparently conflicting sets of observations have their counterpart in the way pressure-drops
can be calculated for various network models. If junctions between tubes of different cross-section,
including entry and exit points to large pores, are approximated by discontinuities in cross-section, then
potentially dominant entry/exit pressure losses are predicted; if however, all cross-sectional changes are
taken to be smooth, and the lubrication approximation is invoked, no elastic losses are predicted. Reality
probably lies between these two extremes. Little has been done so far to analyse this possibility.
5. The dispersion tensor for non-Newtonian fluids
5.1. Origins of dispersion
If we consider a uniform fluid flowing through a single long circularcylindrical tube of radius a (typical
of micro-models), then a simply-calculable velocity distribution u(r), r the radial co-ordinate, arises for
any generalised Newtonian fluid. If at any time, t = 0 say, a tracer of concentration c0 is added to the
entering fluid, it will be advected by the fluid. At a later time t > 0, the tracer particles will have moved
a distance
xtr = u(r)(t te )

(5.1)

down the tube, where r represents the radial position at which they entered at time te . It may easily be
shown for a Newtonian fluid that the cross-sectional average concentration in the tube will be

c0 (1 x/L), x L
c =
(5.2)
0,
x>L
where
L = 2ut

(5.3)

and c0 is a constant.
This is the result of pure advection and neglects the effect of molecular diffusivity D on the tracer,
whose magnitude relative to advection is often associated with the Peclet number (Pe) for mass transport
Pe(m) =

ua

(5.4)

This is usually thought to be large (but for u = 104 m/s, a = 105 m, D = 109 m2 /s, values not
irrelevant to possible micro-models, Pe(m) = 1). However, a more informative dimensionless parameter
is the mass Graetz number (Gz)
Gz(m) =

ua
2
Dl

(5.5)

where l is the length of the tube; this recognises that the time taken to be advected along the tube is
O(l/u)
and the time to diffuse across the tube is O(a 2 /D); Gz(m) is thus a ratio of times, small Gz

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

459

denoting diffusion dominance, large Gz denoting advective dominance. Taylor has shown that the effect
of diffusion dominance, achieved by making l  a, is to lead to an apparent tracer diffusivity
DTaylor =

a 2 u 2
48D

(5.6)

in the x-direction (this can be rationalised in dimensional terms by considering the mean advective
length L/2 achieved at time a2 /D using Eq. (5.3); we now equate the advective flux 1/2uc
0 at x = L/2
to an apparent diffusive flux based on the concentration gradient (5.2) achieved at this position,
i.e. to DTaylor c0 /(2ua
2 /D), so that they sum to the total flow uc
0 , giving DTaylor u 2 a 2 /D as in
Eq. (5.6)).
These arguments are based on pore-scale kinematics, which we believe will remain important if we are
to predict dispersive effects at a Darcy-scale, as in Eq. (2.6) for a porous medium. Clearly D in Eqs. (2.5)
and (2.6) does not have the same form as DTaylor in Eq. (5.6), while an advective effect like Eq. (5.2) is
not apparent. Why is this?
The answer almost certainly lies in the fact that the continuous streamlines (fixed in space) that must
characterise steady flow in a porous medium sample (almost) the full range of actual point values
characterising the velocity field in an extended porous medium: for part of their length, they will pass
through regions of high velocity at the centre of throats and for other parts they will be arbitrarily close to
stagnation points. Thus, the mean velocity in the direction of the pressure gradient along any streamline
will differ from uDarcy /m by an amount that decreases with streamline length; equivalently the distance
travelled along any streamline will grow more and more linearly with time. The stochastic nature of the
geometry of most porous media ensures that these departures approximate to random perturbations and so
a diffusive term arises in the Darcy-scale continuum relation for mass transport, while there is no purely
advective variation, scaling linearly as in Eq. (5.3). This process is called mechanical dispersion and is
2
dominant even when Gz(m)
sample based on l = l sample and a = k is large. The addition of cross-streamline
diffusion (based on D) will tend to decrease the values of || and In Eq. (2.6), although as Aris has
shown a relatively small value of Pe(m) will mean that a term D has to be added to DTaylor and DI/T(m) to
D, where T(m) the tortuosity for mass diffusion is of order unity.
If the fluid is Newtonian, then unsteadiness in the overall flow will not affect mechanical dispersion;
low Pe effects will scale accordingly.
The first detailed papers on this subject were published by Saffman [11], though they were largely
qualitative; because he effectively assumed complete mixing at the inlet to each capillary the work
reported inevitably led to estimates for dispersion lengths scaled by the length of the capillaries and so
represented a version of the chemical engineers notion of a residence-time distribution in a sequence of
mixers.
5.2. Effect of fluid rheology on dispersion
It is quite clear that the main effect will be associated with its effect on the pore-scale velocity field,
and hence, on the steady streamline pattern. If we consider a long tube (of length L), then a result similar
to Eq. (5.6):
DTaylor = NnN

u2 a 2
D

(5.7)

460

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

will arise, but where NnN will be a function of the rheology of the fluid and the mean flow-rate (Darcy
velocity u); if we can characterise the fluid rheology for simple shear flow by a set of characteristic times
[tn ], n = (1, 2, . . . , N), tn > tn1 :
 
ut
n
NnN = NnN
(5.8)
a
This scalar relation will be applicable provided
Pe(m) 

ut
N
,
a

i.e.

a2
t

(5.9)

which implies that


ut
N
1
L

(5.10)

which is effectively the criterion for thixotropic or elastic fluids to behave as generalised Newtonian fluids
in Darcy flow, when L is identified with lsample .
However, this is quite different from the effect on || and , which in principle has to be separately
calculated. There is very little information in the literature about this aspect. We have noticed in our work
that the effective persistence length of a streamline before it is mixed with neighbouring streamlines
increases as the fluid becomes more shear-thinning.
5.3. Fingering in unstable displacement
This is an effect that is observed in all cases of displacement of high viscosity fluid by one of lower
viscosity. It is usually treated as a continuum phenomenon, i.e. one where the scale of the fingers of
low viscosity fluid advancing into the higher-viscosity fluid is at least of the order of the Darcy scale,
and is often compared with the (more carefully investigated) fingering observed in HeleShaw cells
(involving low-Reynolds number (Re) flow between two closely-spaced parallel flat plates). It can be
observed in both miscible and immiscible displacements, i.e. when the displacing fluid is miscible or
immiscible with the displaced fluid. For miscible fluids, ultimate mixing, i.e. mixing at the pore-scale,
depends upon molecular diffusion of one fluid into the other, or of the viscosifying agent when the solvent
is the same in both fluids; for immiscible fluids, interfacial tension will act directly at the pore-scale, and
through capillary pressure gradients at the Darcy-scale (interfacial tension is the crucial factor in most
HeleShaw cell flows). Both mixing and interfacial tension will affect the detailed mechanics of fingering.
We concentrate here on miscible fingering.
Detailed study of pore-scale models shows that fingering can occur at all scales (specifically, those
below the Darcy-scale) in the sense that the less viscous fluid must move faster than the more viscous
fluid when subject to the same pressure gradient, and evidently so when diffusion can be neglected. Thus,
advective extension of the region over which mean 3 concentration or saturation varies, of the type given
by Eqs. (5.2) and (5.3), can be expected. Molecular diffusion will now play a significant role in converting
this advective process into a Taylor-like diffusion process, where a in Eq. (5.6) must be re-interpreted as
3

The mean is to be taken as an areal average over planes orthogonal to the direction of displacement, the area cutting many
fingers.

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

461

finger width w f . The changeover between advected and diffusive behaviour will occur at lengths of the
order uw
f2 /D at most, since lateral dispersion will add to the molecular contribution envisaged by Taylor,
and so, as the distance covered by any front increases, the characteristic width of fingers can be expected
to increase, until lateral restrictions constrain them. We have not been able to discover any definitive
treatment of this phenomenon, even for the highly homogeneous case where l block  l sample , lf (lf being
the length of the fingers) and where the fluids are both Newtonian; in real porous media, the fingering
pattern is often determined by heterogeneity, so that wf tends to be linked to the correlation scale of
m(xblock ) as well as to the distance travelled by the front, Xfront . The one parameter of clear importance
is the ratio of effective viscosities:
1
(5.11)
M12 =
2
Pseudo-plasticity will have its most direct effect in altering M12 where 1 and 2 are now to be interpreted
as GN1 (Eeff1 ) and GN2 (Eeff2 ). A difficulty arises at once because fingering itself suggests that E will be
larger for the less viscous than the more viscous fluid; in the miscible case, where the absolute magnitude,
e.g. Cn in Eq. (4.8), rather than the shear-rate index, e.g. n in Eqs. (4.7) and (4.8), is altered by concentration
of the viscosifying agent, fingering will be intensified by pseudo-plasticity, e.g. n < 1.
Thixotropy, allied to pseudo-plasticity with n < 1, might be expected to stabilise large well-established
fingers against tip splitting or lateral fingering instabilities, though there is no evidence for this that we
know of. Elastic fluids, that show apparent shear thickening in the Darcy representation, e.g. of the type
described approximately by
(Eeff ) = 0 (1 + cST De2 )

(5.12)

where
1
telas
De = u lpore

(5.13)

cST is a dimensionless parameter, and telas the characteristic relaxation time, can be expected to be
relatively stabilised against the fingering instability, particularly if De based on the mean displacement
velocity leads to cST De2 1 for at least one of the fluids.
6. Multi-phase flow of non-Newtonian fluids
6.1. The BuckleyLeverett approximation
Capillarity acts to distribute two or more phases within the pore space at the pore-scale, so that at rest
they attain equilibrium in a time of order k1/2 / (for as large as 1 Pa s, k as large as 1012 m2 and
as small as 103 Pa m, the equilibration time will be <1 s). For most oil and gas reservoirs containing
residual brine, the equilibration time for samples large enough to make Darcys law valid becomes of
order lsample / (if lsample is 102 m, is 102 Pa m and is 103 Pa s, then the equilibration time is still
<1 s) and is such that continuum equations assuming local equilibrium can be used.
For large-scale flooding experiments aimed at making O(1) changes to the saturations in an extensive porous medium, the distance l over which the saturation changes from its initial to its final
value is often large enough for the expected capillary pressure gradients caused by the saturation

462

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

gradients to be small compared to the viscous pressure gradients 4 given by the modification (Eq. (2.7))
to Darcys law, and so at this scale capillarity might be neglected in the sense that saturationrather
than the distribution of the phases at pore-scale and minimum Darcy-scaleis determined by mass
(I )
transport governed by the relative mobilities krel
/(I ) while all the p(I ) can be taken equal at any given
point.
The set of Eqs. (2.7) and (2.8) together with

s (I ) = 1
(6.1)
for the s(I ) , u(I ) and p is known as the BuckleyLeverett approximation and is used in many reservoir
calculations.
6.2. Relative mobilities for non-Newtonian fluids
Since the equilibrium distribution of the phases at any given saturations is determined by capillarity
and wettability, it may be assumed that non-Newtonian rheology is not significant at that level. It will of
course affect the rate of attainment of equilibrium, and in cases where the equilibrium is itself known to be
dependent on saturation history, it may change the final state. However, these variations are not specific
to non-Newtonian fluids and are similarly affected by a varying viscosity ratio for purely Newtonian
fluids: as they are not generally considered important in the latter case, they may be neglected in the
former.
(I )
The interesting question is whether a knowledge of krel
(as a function of saturation) for Newtonian
fluids and ofc ,
defined in Eq. (4.5), or alternatively of MDarcy , defined in Eq. (4.15) for simple shear
(I )
/(I ) ,
flow in the same medium, is sufficient to provide effective values for the relative mobilities krel
(I )
(I )
now to be functions of s and |u | in the case of non-Newtonian fluids.
To try answering this question, we note that the pore space available for flow of phase I at any given
saturations should, according to the argument given in the first paragraph of Section 6.2, be that which is
determined by capillary and wettability forces, and therefore not rheology dependent. These pore spaces
will not be the whole pore space, but their basic geometrical characteristics in terms of flow resistance
(I )
can be supposedly given by krel
. To derive the relevant value of (I ) , we note that the effective flow-rate
in the reduced pore space now available to phase I can be taken as |u(I ) |s(I ) , and use this value as |u| in
the already known ,
or MDarcy .
6.3. Dispersion and fingering in multi-phase flows involving non-Newtonian fluids
In practice, these are important issues, primarily because complex treatment fluids tend to be very
viscous and so back-flow of reservoir fluids involves a large adverse mobility ratio, and because pore-scale
mixing of inhomogeneous fluids can lead to unexpected and rapid variations in fluid properties and hence,
to sharp gradients in continuum parameters describing the flow. Almost all models used are specific to
particular observed phenomena; they tend to be empirical rather than rigorous.
A suitable ratio is given by k1/2 /lufor = 102 , k 1/2 = 106 , = 103 , l = 102 , u = 106 , all in SI units, this gives a
value of 0.1.
4

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

463

7. Other effects in flow of complex fluids through porous media


The model media introduced in Section 4and the Appendix A were described and analysed for simplicity
in the context of homogeneous fluids with bulk properties subject to a no-slip boundary condition at all
matrix or tube surfaces.
In practice, many of the fluids having non-Newtonian rheology are multi-component fluids with various
chemical species in solution; these can be ionic, polymeric and/or surfactant in various combinations.
They can interact with the matrix at the pore walls; when more than one phase is present, they can
exchange between phases or adsorb at mobile interfaces. These mass transfers can have long or short
time scales relative to those mentioned earlier. They can alter the interfacial tension between phases, the
wettability and slip at the matrix surfaces, and the bulk rheology of the fluid phases.
All of these complex properties can, in principle, be studied outside the porous medium, providing
the required solidliquid interfaces can be prepared at laboratory scale. Their effect on micro-scale
models can be studied directly by physical measurements at a pore-scale, or by suitable algorithms in
simulations. Experiments on samples of porous media can provide information on their Darcy-scale
effects. It does not follow, however, that observed results on real porous media could be predicted
from micro-models selected to match behaviour for fluids not exhibiting these complex properties; these
additional properties and their effects may well lead to further conditions that the micro-model would
have to meet.

8. Discussion and conclusions


Despite many attempts to derive continuum models from pore-scale information, whether formal
and theoretical or directly experimental, there is little agreement on how to predict the behaviour
of rheologically-complex fluids when they flow through porous media. Since most applications of
continuum theory use Darcys law or its straightforward extensions, most procedures for prediction
attempt to do so by prescribing an effective viscosity for single-phase flow and relative permeabilities in multi-phase flow. Data on the steady-state viscosity function as a function of shear rate,
obtained from bulk experiments, shifts the choice of effective viscosity to the choice of an effective
shear rate, which is assumed to depend on flow-rate, permeability and porosity in a simple and uniform
fashion.
Given that the pore space in a porous medium is geometrically very complex with different media displaying very different structures, it seems unlikely that simple universal algorithms for deriving effective
viscosities from bulk data will be successful. Introduction of a detailed stochastic description of real pore
geometry might be a way of parametrising the porous medium, but the labour of doing this, particularly
when the significance of any such parameters is still obscure, makes it an unattractive option, at this
stage. The alternative that has been used is to simplify the geometry by using micro-models, analogues
based on interconnecting, easily parameterised, unit channels placed on a regular lattice. This digitised
representation is ideally suited to numerical (computational) techniques, which allow direct averaging
of micro-model quantities to yield values of continuum (Darcy-scale) variables. Simple exact flow
solutions based on exact fluid rheology as well as exact wall effects exist for unit channels and so
the passage from pore-scale to continuum-scale variables becomes purely algorithmic, and thus, feasible
on modern computers.

464

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

The problem then becomes one of choosing the right distribution of components for the analogue
geometry. The proposal that we make here is that this distribution should be constrained to satisfy
as many observations as can be made on the same or equivalent samples of a given porous medium
as possible. These observations should involve a limited number of test fluids covering the required
range of different rheological behaviour, over such a range of steady flow-rates that the full rheological
complexity of each fluid be sampled, using tracer tests to measure dispersion lengths and static overall
molecular-diffusion effects to yield one measure of tortuosity. Comparison of forward predictions for
continuum properties using a candidate micro-model (for the sample porous medium concerned) should
suggest optimal ways of meeting the required constraints, namely the observed data. We have decided
that the best procedure is to disregard dispersion in the first instance, and to restrict attention to obtaining
models that successfully predict pressure-drops; this has the pragmatic advantage that pressure-drops and
flow-rates are much easier to measure accurately than residence-time distributions, while the dependence
of the latter on sample length can be more significant than for pressure gradients. Extending the model
to cover dispersion depends crucially on mixing at nodes: it remains to be discovered whether a single
simple network model can predict dispersion lengths for fluids with widely different rheologies using
simple algorithmic rules for particle tracking. This is described briefly in Section A.4.
Once an optimal micro-model has been derived, it could then be tested on a new fluid in similar flow
conditions, and then be used to predict unsteady displacement flows of miscible fluids. If predictions
matched observations for both tests, then confidence in the model would be sufficient to use it for routine
predictions until it could be shown to fail significantly in some particular application involving a hitherto
insignificant physical or chemical process. It would then be timely to increase the number of constraints
on the initial model to include the new physical or chemical process.
One of the intermediate steps that has been interposed successfully is the use of small-scale physical
embodiments of the micro-models to confirm experimentally that the correct mechanics, physics and
chemistry was being used at the micro, i.e. unit, scale in the computational model. Comparisons can be
undertaken both at the detailed single-unit scale, and at the continuum, i.e. model-averaged, scale. This
is of particular importance when trying to model capillarity in two- and three-phase flows and in tracking
fronts in displacement flows. A related approach is to use tomographic techniques to follow flows in bead
or sand packs of suitable materials at a sub-Darcy-scale. These partially-averaged measurements on real
media provide an independent and potentially crucial test of micro-scale simulators, both computational
and physical. They are expected to provide crucial information about dispersion, and how it should be
modelled.
A further issue which has to be mentioned is that most macro-scale flows in porous media are affected by
heterogeneity and anisotropy. So continuum models for prediction of such flows have to include reliable
methods for up-scaling when computer simulators are used. As far as the effects of non-Newtonian
behaviour are concerned, the arguments given in the previous sections can almost always be adapted to
apply to up-scaling techniques developed for simple Newtonian or two-phase flows.
For further reading see [16,810,12].

Appendix A. Micro-models
In this appendix, we consider simple micro-models that can be used to make predictions for Darcy-scale
behaviour. For simplicity, at this stage, we shall consider micro-models based on networks of intercon-

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

465

necting capillary tubes of fixed length, but different diameter; without loss of generality they can be placed
on a finite cubical (square, or linear) lattice taken to represent a minimal REV. Nodes, i.e. points where
up to 6 capillaries meet (4 for two-dimensional, 2 for one-dimensional models), will be assumed to have
no volume. A porosity can be calculated for such a system.
For single-phase flow of a Newtonian fluid, an instantaneous linear flow-rate/pressure gradient
relationship, for flow between opposite faces of a cubical REV held at constant pressures, can readily be calculated, given the diameters of all the capillary units within the REV, based on Poiseuilles law
for flow in a single capillary. This provides a computed permeability. Modifications to cover exit/entry
pressure losses associated with nodes can also easily be made. Analogous calculations for generalised
Newtonian fluids will be more complicated (but perfectly feasible), because the analogue of Poiseuilles
law is flow-rate dependent, leading to a flow-rate and rheology-dependent effective Darcy viscosity.
For thixotropic or elastic fluids the entry state (given by the past history of flow) at each capillary entry
has to be specified. This is a non-trivial matter if exact histories of fluid elements are to be calculated
according to Eqs (4.8) and (4.9). Some simplification can be achieved if (a) a dimensionless residence-time
in a single capillary based on Eq. (4.13) using the capillary length instead of sample length is either very
large or very small, and (b) fluid entering a node is assumed to be fully mixed at nodes. This gives
either a generalised Newtonian response in each capillary (dealt with in the previous paragraph), or
a Newtonian response that is slowly-changing. By imposing periodicity on the micro-scale model, an
effective steady-state Darcy viscosity can be obtained by iteration using the exit state from one REV as
the entry state for the next. This yields a behaviour that is analogous to that represented by Eq. (4.15).
For elastic fluids, exit/entry losses (at nodes) that increase much more rapidly with flow-rate than the
simple-shear capillary-flow losses can be included.
Algorithms to predict dispersion are more difficult to devise if they are to be realistic, particularly if
the mass Pe and Gz based on a single capillary are neither very large nor very small. The results given in
Section A.4 are effectively based on small Gz.
Simple capillary models are not thought to be sufficiently realistic when modelling two- or three-phase
flow in real porous media, for which various pore/throat models have been used. Polygonal cross-sections
for throats allow a range of steady saturations within them, thus, leading to direct predictions of relative
permeability.
A.1. Flow in a single capillary
For Newtonian fluid of viscosity N , Poiseuilles law can be written as
N =

d 4 G
,
128q(d, G)

G = p

(A.1)

where d is the capillary diameter, q is the flow-rate at a pressure gradient G, and where the distinction
between p and p will no longer be made. For a generalised Newtonian fluid satisfying Eq. (4.3), the
corresponding relation is
GN (sw ) =

w
d 4 G/32
=
sw
{3qGN (d, G) + G(dqGN /dG)}

(A.2)

where w is the wall shear stress and s w the wall shear rate. Normally, we think of GN (s w ) as the
basic rheological function. However, for our purposes, it is simpler to use the function qGN (d, G) or its

466

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

inverse GGN (q, d) as the primary rheological function characterising our fluid (this is the case in capillary
viscometry anyway). For example, for a power-law fluid, the relation (4.7) is replaced by


q0
qGN =
(A.3)
G1/n
1/n
G0
1/n

where q0 /G0 is a single dimensional constant, provided by any pair of values (q, G) for the capillary
concerned.
The advection and diffusion caused by flow through such a capillary has been analysed for a Newtonian
fluid in Section 5.1. Very similar calculations can be carried out for any GN fluid. The definitions of the
Pe and Gz can be retained.
A.2. Flow in a parallel bundle of capillaries (variable-bundle model)
We consider now a set of capillaries of variable length 5 lI and diameter dI , (dI +1 dI , lI +1 lI if
dI +1 = dI ), I = 1, . . . , N. We take them to be representative, at some level, of unidirectional Darcy flow
in a homogeneous isotropic medium. We, thus, suppose them to represent the pore space in a cubical
REV of length l lI (to allow for tortuosity) for all I, each one starting on one face and ending on the
opposite face. The porosity of the REV will be
 d 2 lI
I
mVB =
(A.4)
3
4l
I
For a given pressure-drop P across the entry and exit faces, the total flow-rate Q for a GN fluid will be



dI , P
QVB =
(A.5)
qGN
l
I
I
and so the ratio k/ in Eq. (2.1) can be readily obtained for any particular fluid characterised by qGN . If
we choose a Newtonian fluid, of arbitrary viscosity N and specify the set dI , lI , l, then mVB and kVB
can be obtained. If m and k are known for a particular sample, then these represent two scalar constraints
on the set characterising the bundle. Noting that k1/2 has dimensions of length, a suitable dimensionless
formulation would use the quantities
dI
lI
dI = 1/2 ,
lI =
(A.6)
k
l
as model parameters and
u VB =

kQVB
,
2
l qGN (k 1/2 , PVB / l)

p VB =

PVB
lGGN (QVB / l 2 m, k 1/2 )

as the corresponding Darcy-scale variables, all expected to be O(1). Eq. (A.4) becomes
 

k
m=
dI2 lI
4l 2
I
5

(A.7)

(A.8)

The requirement that all capillaries should have the same length introduced above can be maintained by supposing that each
of the capillaries is composed of a large but variable number of equal basic capillaries.

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

467

Since we expect k/ l 2 = O(02 ), where 0 is defined by Eq. (3.4), the sum I will be over a large number
of capillaries. For formal calculations it is convenient to replace the sum by an integral over a distribution
which are then parameterised by a small number of dimensionless
of capillary diameters d and lengths l,
constants. A simple example is given by
l)
= N1 exp{(d d0 )2 } exp{(l l0 )2 },
prob(d,

l0 > 1

where N1 is a number yet to be specified. Eq. (A.8) is then replaced by




4ml2
2

= N1
l exp{(l l0 ) } dl
d 2 exp{(d d0 )2 } dd
k
1
0

(A.9)

(A.10)

which can be taken to define N1 in terms of 4ml2 /k and l0 , d0 . The dimensionless analogue of Eq. (A.5)
provides a further constraint on the distribution given by Eq. (A.9). For a Newtonian fluid, linearity
implies that the integral will be computable and it will introduce a single relation between l0 and d0 . If
we were then to carry out a permeability test on a known power-law fluid, this would provide a further
constraint using the integral form of Eq. (A.5), and so l0 and d0 would at best be uniquely defined, the
alternative being that no physically-acceptable solution were possible.
What is clear is that in general no simply defined capillary bundle model would satisfy flow measurements made using a large number of different generalised Newtonian fluids in equivalent samples of the
same real porous medium. To do so would require careful choice of a large number of parameter pairs
dI , lI .
It is tempting to try to match dispersion data also. The residence-time distribution for any given parallel
bundle model can be calculated. If the flow is purely advective, then the arguments of Section 5.1 can
be used to give a result similar to Eq. (5.2) with c representing a cross-sectional average concentration
across all the tubes at any 0 < x < l. For any given tube we take xI = x lI ; c will be a function of
1 (x/umax t) where umax is the maximum value of uI max /lI , uI max being the maximum velocity in
tube I.
If the mean Gz
QVB k
Gz(m)
(A.11)
VB =
l3D
is very small, then Taylor diffusion is relevant in each tube, from which a single-tube and hence a
parallel-bundle residence-time can be calculated. If the distribution of dI and lI are sufficiently close
to some d0 ,l0 , then the residence-time distribution would be dominated by some mean Taylor diffusion
coefficient
(m)
DTaylor

Q2VB k
l4D

(A.12)

which as we have argued in Section 5.1 would not be consistent with the observed dispersion tensors D
proportional to mean velocity QVB /l2 as in Eq. (2.6). Broadening the (dI , lI ) distribution would accentuate
mean advective spread, but it is unlikely that this would provide the result (2.6).
On the other hand, if the REV dimension l of the micro-model were taken to be many times smaller than
the length lsample over which D were estimated, then the required dispersive model could be attained by
supposing all the flow in the micro-model REV to be instantaneously mixed each time it passed through
the exit plane of the next identical REV model. This would not make any alteration to permeability

468

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

calculations, but would determine a particular value of l/k1/2 that best fit observations. This argument
would be strongest if the relevant Gz(m)
VB , based on the l fitting the sample dispersion measurements,
were large.
Thixotropy could be included equally coarsely by supposing that the rheology was uniform at any
cross-section in a tube, and that it was slowly varying in any tube I according to Eq. (4.10) with (d/dt)I =
u I (d/dx)I and EI = 2u I /dI . The mixing leading to dispersion would provide a volume-averaged value
based on the exit values MI for one REV as the entry values for the next. The calculation of
for M,
pressure-drops and flow-rates for each tube I in any given REV would be more complex than in the case
of a GN fluid.
Elasticity is difficult to introduce in parallel tube bundles because there is no change in cross-section of
any of the channels; this could be done by introducing, purely to account for elastic entry and exit losses,
a specified number of large pores along the length of each capillary.
A.3. Flow in a bundle of equal capillary tubes of slowly-varying diameter (equal-bundle model)
This represents an alternative way of matching permeability, porosity and tortuosity measurements in
samples. Each of the tubes is now defined by a total length l0 l, composed

in the digitised form by


a connected set of N capillaries of diameter dI and length lI . By definition I lI = l0 . The porosity is
defined by
n  2
mEB =
(A.13)
lI dI
4l0 I
where n is the number of tubes/unit area. If lI  dI , lubrication theory as implied by Poiseuilles law
can be applied here to each of the N component capillaries. qEB is now constant for all tubes and so

PEB =
lI GGn (qEB , dI ),
(A.14)
QEB = NqEB
This relation obviously covers all GN fluids. The arguments given in Section A.2 concerning thixotropic
fluids can be applied to this case without significant alteration.
Elastic pressure losses can be introduced directly at each change in cross-section within any tube, i.e.
after each (lI , dI ) unit. If a known loss can be attributed to any given fluid in passing from one diameter dI
to another dI +1 at any given qEB , then the additional pressure losses in a REV of length l can be calculated.
These losses which we can write formally as Pelas (d1 , d2 , qEB ) would yield

PEB =
{lI GGN (qEB , dI ) + Pelas (dI , dI +1 , qEB )}
(A.15)
I

instead of Eq. (A.14). This is the most obvious advantage of the equal-bundle (EB) over the variable-bundle
(VB) model. Conversion into an integral form for the elastic case would be rather elaborate and unphysical,
though it would be reasonable for inelastic fluids.
Clearly, a combination of the two models would provide extra flexibility in matching data and as we
shall see might be a useful simplification, for calculation purposes, of the network model described in
the next section.

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

469

A.4. Flow in a network of capillaries


The capillary-bundle models described in Sections A.2 and A.3 are potentially perfectly adequate for
one-dimensional Darcy flows of single-phase fluids in homogeneous isotropic porous media. As argued at
the end of Section A.2, even dispersion data could be matched when l l sample ; the mixing process given is
easily applied to give D|| . A more elaborate process involving the output from laterally-neighbouring REVs
mixing in specified proportions with the central REV to form the input for the next axially-neighbouring
REV would be needed to model D . The use of tortuous (lI > l; l0 > l) capillaries in the (VB, EB)
models can in principle match pure molecular diffusion data for tortuosity. Given rheological data in the
(q, G) form for any specified fluid, the computational demands for forward prediction (calculation of the
P, Q relation for given dI , lI , l) and inversion (calculation of optimal dI , lI for given P, Q) are as modest
as they could be.
The case for using networks, whether two- or three-dimensional, is not clear at this point. For simplicity,
regular networks with equally-spaced nodes suggest themselves. On simple geometrical grounds, it is the
multiple connectivity of the nodes that seems best to fit with pore-scale flow: tortuosity, now defined in
terms of mean lengths of flow paths, is achieved by paths of least resistance having lateral as well as
axial steps; mixing at nodes provides a simple approach to dispersion based on large Pe covering both
axial and lateral components; it provides far greater flexibility in modelling two- or three-phase flow; its
greatest advantage lies in the possibility of following fingering at a pseudo-pore-scale. The concomitant
of greater flexibility is increased computational demands, particularly for forward prediction in the case
of highly-non-linear GN fluids, when repeated iteration of matrix solutions is involved to give a single
P, Q value; this imposes a corresponding penalty on the even more computationally-intensive inversion
process.
The case for networks must, therefore, be based on using micro-models to represent multi-phase flow
and small-scale fingering in porous media, the details of which are beyond the scope of this article. Some
preliminary results for power-law fluid flow in square networks are given in Section A.4.1.
A.4.1. Flow of a power-law fluid in a two-dimensional capillary network
We consider a power-law fluid, as defined in Eq. (4.7), and a regular two-dimensional capillary
network mapped on a rectangular Cartesian grid (co-ordination number 4). The radius of each capillary of the network is randomly chosen from a capillary-radius probability distribution function (pdf).
For this network to be an effective model of a particular natural porous medium, for which flow data
for the power-law fluid under consideration are assumed to be available, the first parameter to
reproduce is the permeability k. Clearly, an infinite number of the pdf parameters can give rise to the
same value of k. One particular choice would be to have all the capillaries with the same radius r1
(model 1), while another particular choice would be to have them distributed between r2 min and r2 max
(model 2). As we will see in the following, we cannot expect both models to match experimental data
obtained with the power-law fluid. Therefore, k is not sufficient to describe the flow of the power-law
fluid.
We are still left with a large number of degrees of freedom to attempt matching of additional data.
Parameters from the pdf, geometry of the network, co-ordination number, length of capillaries suggest
themselves. In order to gain insight into which parameters matter, it is useful to study the forward
problem whereby a given power-law fluid flows into a given network. The two-dimensional networks and
associated boundary conditions, which have been studied here, are illustrated in Fig. 1. The methodology

470

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

Fig. 1. Illustration of two-dimensional networks considered in this study. Capillaries are mapped onto a rectangular and regular
grid of 256 (flow direction) by 32 capillaries. Each capillary has a length of 200 m. Capillary radii are picked randomly according
to one of the five pdf plotted on the right. A uniform inlet pressure is applied at both inlet and outlet, flow-rate is then computed
for different values of the inlet pressure, giving rise to apparent rheograms in porous media when measuring viscosity with
Eq. (2.1). Distribution 4 corresponds to the case where all capillaries have the same radius equal to 10 m.

for solving flow in such networks in described in ([12], p. 195). In the following, we present some general
observations and measurements made with the numerical networks:
1. It can easily be proved and observed numerically (Fig. 2), that a power-law fluid still behaves as a
power-law fluid at the network scale, and the apparent power-law exponent is the same as the bulk one:
u (2p)1/n

Fig. 2. Example of apparent rheology of three different power-law fluids as determined by a network model (distribution 1).
The value of the slopes for these three lines is exactly the value of the bulk power-law exponent of the corresponding fluid. The
apparent viscosity is the value of the viscosity determined by Darcys law (2.1) when everything else is known. Shear rates are
linearly approximated from Darcy velocities through Eq. (4.4) with a fixed value of .

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

471

Fig. 3. Evolution of flow paths with power exponent n, n = 1 (top, Newtonian case), n = 0.1 (middle), n = 0.01 (bottom),
network based on distribution 2. These pictures are generated by dropping particles at the inlet and following them according to
the mean velocity in each capillary. At a junction between capillaries, the probability of going into a given capillary is proportional
to the flow-rate in this capillary. Each of these pictures has been obtained by using 10,000 particles, the colour indicating the
frequency of passing through a given capillary. White indicates that a large number of particles have passed through this capillary
(more than 1000) while black indicates 0 occurrences.

Fig. 4. Tortuosity distributions along of flow paths for different pdf, in the Newtonian case and for n = 0.01. Tortuosity is
measured by following a particle as it flows in the network and by recording the length of its flow path. This graphs shows that
different pdf give rise to different tortuosity distributions and that power-law fluids follow paths of lower tortuosity as illustrated
in Fig. 3.

472

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

Fig. 5. Evolution of the correction factor (as defined in Eq. (4.4)). This factor is used to correct the Darcy velocity in order to
calculate what the shear-rate would be for the measured Darcy viscosity. This shear rate is called the apparent shear-rate.

2. The tortuosity and number of flow paths decreases with decreasing power-law exponent n, as illustrated
in Figs. 3 and 4.
3. The correction factor , as defined in Eq. (4.4), varies with both pdf and power-law exponent n (Fig. 5).
4. The higher the mean tortuosity of the network (i.e. the mean tortuosity of flow paths for any Newtonian
fluid), the higher the correction factor (Figs. 4 and 5).
5. For a given network, the effective pdf, as determined by following particles along flow paths, does not
vary significantly with the power-law exponent (Fig. 6).

Fig. 6. By recording the radius of each capillary visited by a particle, an effective pdf can be determined illustrating which
radii have been the most frequented. This is such an example with distribution 3 showing that this effective pdf does not vary
significantly with the power-law exponent.

J.R.A. Pearson, P.M.J. Tardy / J. Non-Newtonian Fluid Mech. 102 (2002) 447473

473

From these observations, we conclude that the paths followed by power-law fluids have a pathwise
permeability very close to that of the paths followed by Newtonian fluids. Power-law fluids seem to follow
paths of least tortuosity and not of least resistance in the permeability sense. It is now clear that model
1 (or any capillary bundle model without tortuosity) is not a good choice as the tortuosity of its flow
paths is always unity, whatever the fluid. The mediums tortuosity appears to be an important property
that all models have to satisfy if they are meant to simulate flow of a power-law fluid; this is in line with
the model (Eq. (A.2)). This preliminary study suggests that an understanding of the relationship between
the effective pdf, which does not vary significantly with power-exponent, the initial geometrical pdf and
the tortuosity distribution of a porous medium, will enable one to constrain network models capable of
modelling flow of Newtonian, power-law and possibly all generalised Newtonian fluids.
References
[1] P.M. Adler, Porous Media, Butterworth-Heineman, London, 1992.
[2] P.M. Adler, H. Brenner, Transport processes in spatially periodic capillary networks: Parts IIII, Phys. Chem. Hydrodyn. 5
(1984) 245297.
[3] R. Aris, Mathematical Modelling, Dispersion in Flow, Academic Press, London, 1999 (Chapter 7).
[4] G.I. Barenblatt, V.M. Entov, V.M. Ryzhik, Theory of Fluid Flows Through Natural Rocks, Kluwer Academic Publishers,
Dordrecht, 1990.
[5] J. Bear, Y. Bachmat, Introduction to Modelling of Transport Phenomena in Porous Media, Kluwer Academic Publishers,
Dordrecht, 1990.
[6] J. Bear, J.M. Buchlin, Modelling and Applications of Transport Phenomena in Porous Media, Kluwer Academic Publishers,
Dordrecht, 1991.
[7] R.B. Bird, R.C. Armstrong, O. Hassager, Dynamics of Polymeric Liquids, Vol. 1, 2nd Edition, Wiley, New York, 1987.
[8] F.A.L. Dullien, Porous Media, Academic Press, New York, 1992.
[9] C.M. Marle, Multiphase Flow in Porous Media, Technip, Paris, 1981.
[10] G. de Marsily, Quantitative Hydrogeology, Academic Press, New York, 1986.
[11] P.S. Saffman, A theory of dispersion in a porous medium, J. Fluid Mech. 6 (1959) 321349;
P.S. Saffman, Dispersion due to molecular diffusion and macroscopic mixing in flow through a network of capillaries, J.
Fluid Mech. 7 (1959) 1 94208.
[12] K.S. Sorbie, Polymer-improved Oil Recovery, Blackie, Glasgow, 1991.
[13] R.I. Tanner, Engineering Rheology, 2nd Edition, Oxford, 1988.

Вам также может понравиться