You are on page 1of 14

Advances in Water Resources 32 (2009) 723–736

Contents lists available at ScienceDirect

Advances in Water Resources


journal homepage: www.elsevier.com/locate/advwatres

The role of scaling laws in upscaling


Brian D. Wood *
Oregon State University, School of Chemical, Biological, and Environmental Engineering, 103 Gleeson Hall, Corvallis, OR 97330, USA

a r t i c l e i n f o a b s t r a c t

Article history: In this work the process of coarse-graining in complex subsurface hydrologic systems is discussed, with a
Received 7 March 2008 particular effort made to examine the difference between the mathematical process of averaging as dis-
Received in revised form 4 August 2008 tinct from the process of upscaling. It is possible to show that the process of averaging itself is not suffi-
Accepted 19 August 2008
cient to reduce the number of degrees of freedom required to describe a complex heterogeneous system.
Available online 13 September 2008
The reduction of the number of degrees of freedom is accomplished entirely by the subsequent scaling
laws that one assumes are valid filters for eliminating redundant (or low value) information in the sys-
Keywords:
tem. A specific example is given for the upscaling of the effective dispersion tensor in a randomly heter-
Upscaling
Averaging
ogeneous porous medium. It is shown that: (1) generally, an averaged description can be developed, but
Coarse-graining the system does not contain any less information than the original problem; (2) by adopting a number of
Transport scaling laws, a nonlocal in time and in space formulation can be obtained that is identical to results
Subsurface hydrology obtained previously, and (3) by making one additional assumption, a local macroscale transport equation
Stochastic methods can be obtained.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction imations (or in the terminology that will be used for the remainder
of this paper, the scaling laws) that we impose on the structure of
In 1945, the celebrated mathematician and physicist Wiener the heterogeneous system that allow a useful solution to be devel-
and his colleague Rosenblueth wrote [1]: oped rather than the application of any particular mathematical
framework.
No substantial part of the universe is so simple that it can be
The purpose of this paper is to examine the distinction between
grasped and controlled without abstraction. Abstraction con-
averaging (the formal part of the coarse-graining process) and
sists in replacing the part of the universe under observation by
upscaling (which involves both mathematics and the intellectual
a model of similar but simpler structure. Models, formal or intel-
or conceptual part of the modeling process). Averaging can be
lectual . . . are thus a central necessity of scientific procedure.
done, in principle, for any heterogeneous subsurface system,
Hydrologists, as a group, have developed some very sophisti- regardless of its complexity, or the spatial variation of its proper-
cated mathematical models of complex hydrologic systems. As a ties. However, averaging itself does not necessarily result in a
group, we also put a substantial amount of faith in the mathemat- reduction in the number of degrees of freedom (or, equivalently,
ical (and associated numerical) components of our models of these the information content) of the system. It is the imposition of var-
systems. But mathematics is only the formal part, using the lan- ious assumptions, which are termed scaling laws in this work, that
guage of Wiener, by which we describe a more fundamental con- actually accomplish a reduction in the number of degrees of free-
ceptual model of a system. Because of its apparent rigor and dom. The case of nonreactive transport is considered, but the argu-
complexity, we are subject to misleading ourselves into thinking ments made in no way depend upon the presence (or lack of)
that the mathematics is the most important part of the model. This reactions in the system. A conventional convection–dispersion
perspective is probably most prevalent in stochastic hydrology, form at the Darcy scale (also called the microscale in this work) is
where the sophisticated mathematical representations required adopted as the staring point, but, again, this is done only for sim-
to describe the coarse-graining of complex systems is presented plicity in the presentation. None of the discussion depends upon
with exacting detail, while simultaneously often neglecting the conventional convection–dispersion equation being the micro-
(although hardly purposefully) the more physical interpretation scale representation, and any other linear form (e.g. fractional
of what is being accomplished. However, in most cases, it is almost derivative formulations) would be equally valid starting points,
certainly the physical interpretation, and in particular the approx- and would yield equivalent results.
To begin the discussion, it is useful to think about what types of
* Fax: +1 541 737 3099. systems are scalable. The ideal gas law is a good example of such a
E-mail address: brian.wood@oregonstate.edu system. Suppose that one would like to know the pressure of 1 mol

0309-1708/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advwatres.2008.08.015
724 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

of an ideal gas with a given volume (V) and temperature (T). Sup- sented, and a direct solution at the microscale is given in terms
pose additionally that the gas is not too far from atmospheric pres- of the corresponding Green’s functions. In Section 3, the mechanics
sure, so that we do not have to worry about diluteness or nearly of volume averaging are described, and an unclosed macroscale
solid-like behavior. Assuming that the molecules can be treated equation is developed by application of the averaging operations.
as classical objects, then the physical laws that are applicable are In Section 4, the closure problem is described, and, again, a direct
just Newton’s laws of motion. For this system, there is a large num- solution in terms of Green’s functions is given; the problems and
ber of degrees of freedom (for a mole of gas, 6:02  1023 molecules, limitations to such a solution are discussed. The concept of scaling
times 3 space coordinates, times 3 momentum coordinates), which laws is developed extensively in Section 5, and the influence of
is far too many degrees of freedom to be used in any practical com- each of the scaling laws listed above is explored in detail. Of partic-
putation. One can, however, hope that there is significant redun- ular interest is the development of nonlocal and local macroscale
dancy in these degrees of freedom, and attempt to find an transport equations, which differ only in the assumption of a single
upscaled balance equation that contains fewer degrees of freedom scaling law (law 6 below, which requires the separation of the
(and, correspondingly, less total information). This turns out to be macroscopic and averaging volume scales). In Section 6, there is
possible. The scaling laws in this case are (i) the interparticle po- some brief discussion, and some conclusions are presented in Sec-
tential energy can be neglected relative to the kinetic energy, (ii) tion 7.
the velocities are Maxwell distributed or, equivalently, that all va-
lid microstates are equally probable (cf. [2, Chapter 2, 3, Chapter 2. Microscale description
39]), (iii) the statistics of the system are spatially homogeneous
(i.e., they do not change depending upon location within the vol- 2.1. Microscale balance equations
ume), and (iv) some kind of ergodic principle holds that allows
one to exchange averages over an ensemble of velocities with aver- We consider a porous medium (Fig. 1) domain V 0 (with volume
ages of the velocities over a volume. Adopting these scaling laws, V0 ) bounded by the surface A0 (with surface area A0 ). For this dis-
one can use Newton’s laws to reduce the roughly 54  1023 degrees cussion, the term microscale will be adopted to indicate the Darcy-
of freedom down to essentially 1 degree of freedom (the pressure), scale description of flow and transport in the porous medium with
in the form of the ideal gas law, PV ¼ NkT (where N is the number support region (or domain) V D and associated characteristic length
of molecules, and k is the Boltzmann constant). Of course, there has ‘. The term macroscale is then associated with region Vx having
been a tremendous loss of information in this upscaling, but this characteristic length ‘M . The terms microscale and macroscale are
lost information generally has very little value in practical convenient labels that serve to indicate the relationship between
applications. two different observational scales of the phenomena of interest.
One generally does not think of the ideal gas law as being in any These two terms are frequently used in two-scale systems regard-
way approximate. Primarily, this is because the scaling laws are less of the actual dimensions of the characteristic lengths ‘ and ‘M .
very robust. Although not exactly true (if any such notion can actu- Note additionally that it is not necessary that all phenomena
ally be defined), the statistics of a mole of ideal gas at near atmo- occurring within the macroscale volume V 0 have the same
spheric pressure are such that the pressure being different from Darcy-scale support; for this work, however, will assume that this
that predicted by the ideal gas law is overwhelmingly unlikely. is true.
The variance around the mean pressure is so small that measurable The microscale equations governing transport of a nonreactive
deviations from the ideal gas law essentially do not occur. In the chemical species under steady incompressible flow within the do-
subsurface, however, we have a very different situation. Although main V 0 are given by
we can certainly use statistical methods to attempt to describe
oc
the processes in the subsurface, the probability that our assump- ¼ r  ðD  rcÞ  r  ðvcÞ in V 0 ; ð1Þ
ot
tions (or scaling laws) are incorrect are not all that remote. Hydro-
BC1:  n  ðvc þ D  rcÞ ¼ Fðx; tÞ on A0 ; ð2Þ
logic systems might more accurately be described as quasi-scalable
in the sense that the scaling laws themselves are seldom as robust IC c ¼ CðxÞ at t ¼ 0; ð3Þ
as they are for, say, statistical mechanical systems. This puts r  v ¼ 0 in V 0 ;
researchers in this field in the difficult position of developing sta- where v ¼ K  rh and h ¼ p  qgz; ð4Þ
tistical models where the uncertainty associated with such models
BC2: hðx; tÞ ¼ HðxÞ on A0 : ð5Þ
is very high. Under these circumstances, one should be very critical
of the scaling laws that are adopted, yet, curiously, most scaling Here, c is the concentration, D is the Darcy-scale total disper-
laws are imposed with very little fanfare (and, at times, as pointed sion (diffusion plus hydrodynamic dispersion) tensor, v is the
out by Rosenblueth and Wiener [1], even the researchers them- Darcy-scale pore-water (or seepage) velocity, K is the Darcy-scale
selves may be unaware that a scaling law has been imposed, or hydraulic conductivity tensor, and h is the hydraulic head. The
how powerful the ramifications of such assumptions may be). abbreviations BC and IC have been used for ‘boundary condition’
The objective of this paper is to examine the various scaling and ‘initial condition’, respectively. Because variations in the poros-
laws that are imposed during upscaling of mass transport in heter- ity, e, are usually much smaller than the variations in the hydraulic
ogeneous porous media. The idea is to start with an upscaled conductivity, K, variations in the porosity have be neglected for this
expression where essentially no assumptions have been imposed, development. Note that this is not at all necessary [4], but it signif-
and then to add scaling laws sequentially to gradually simplify icantly simplifies the analysis and makes the essential points of this
the problem. Six different scaling laws are examined, as follows: work a little easier to illustrate.
(1) influence of boundaries on the REV, (2) statistical homogeneity In many works on stochastic subsurface hydrology, it is as-
of the parameter fields, (3) separation of the microscopic and mac- sumed that the statistics of the velocity field can be determined
roscopic length scales, (4) quasi-ergodicity, (5) smallness of the from the statistics of the hydraulic conductivity field. Although this
variance (or coefficient of variation) of the parameter fields, and is indeed true for the case where the hydraulic conductivity field
(6) separation of the macroscopic and averaging volume length can be treated as if it contained only small perturbations (i.e., the
scales. variance of the log-transformed and normalized conductivity field
The paper is organized as follows. In Section 2, the microscale is much smaller than unity), for the more general case where the
(Darcy scale) mass and momentum balance equations are pre- hydraulic conductivity may have a large variance it is not clear that
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 725

Fig. 1. The characteristic scales of a two region medium.

such a correspondence can be made. In this light, we treat the tion fields. For the general case, the Green’s functions are nearly as
velocity field as a parameter field that is determined by directly difficult to compute as is the concentration, c, in the original micro-
inverting the fluid momentum balance for the medium. The scale mass balance equations.
disadvantage here is that one still has to determine the flow field
for the porous medium of interest, and this does not lead to as 3. Volume averaging
much of a reduction in the numbers of degrees of freedom for
the analysis as a whole. This issue will be addressed in additional For the method of volume averaging, upscaling is accomplished
detail in the material that follows, but we will proceed under the by forming weighted averages over physical volumes of porous
assumption that the velocity field can be treated as a known media rather than computing averages over an ensemble of identi-
parameter field. cally prepared systems. Volume averaging has been cast in the con-
text of distributions, in which the density or weighting functions
2.2. Integral solution are interpreted as instrument response functions [7]. Such instru-
ment weighting functions might, in principle, vary in space and
Although Eqs. (1)–(5) are complex, these equations could in time as well in response to physical and chemical heterogeneity
principle be solved directly to obtain the microscale fluid velocity and/or the changing composition (e.g. water content, concentra-
and concentration fields. The momentum balance equation (flow) tions of chemical species) of the porous media. In this work, we
is an elliptic one, and the mass balance equation (transport) is par- will use uniform weighting functions, and this results in classical
abolic. Assuming some mild conditions on the hydraulic conductiv- averages of the form
ity field (and, hence, the velocity field), general solutions to these Z
1
two equations exist in integral form. These solutions can be ex- hcijx;t ¼ cðy; tÞdy; ð8Þ
V y2V x
pressed in the form [5, Section 0.8.1]
Z  where V is the volume of the domain of integration, V x , and it is
vðxÞ ¼ K  r HðyÞGh ðx; yÞdA0 dy ; ð6Þ understood that V x is always contained in V 0 (i.e., the averaging vol-
y2V 0
Z Z ume cannot extend beyond the boundaries of the macroscale vol-
t
ume V 0 ). The volume V is then defined by
cðx; tÞ ¼ Fðy; sÞGc ðx; y; t; sÞdA0 ðy  ys Þdy ds Z
0 y2V 0
Z V¼ wðyÞdy: ð9Þ
þ CðyÞGc ðx; y; t; 0Þdy; ð7Þ y2V x
y2V 0
Here wðyÞ is the indicator function given by
where Gh and Gc represent the Green’s functions for the flow and 
1 for y 2 V x ;
transport equations, respectively, and dA0 is a delta function associ- wðxÞ ¼ ð10Þ
0 otherwise:
ated with the surface of the macroscopic volume, A0 (cf. [6]), and ys
is a vector pointing to the surface A0 . Although formally this solu- Typically, we think of the averaging volume V x as having some kind
tion can be expressed in terms of Green’s functions, these functions of regular geometry (for example, spherical) so that the volume V is
are in general as complex as the microscale velocity and concentra- constant, except perhaps near the boundaries of the volume.
726 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

It is also useful to define the deviation from the average concen- Using the definition of the operator L, this is equivalent to (after
tration. The average concentration described above is defined un- some rearrangement).
iquely for every point in the domain V 0 ; the microscale Unclosed macroscale mass balance
concentration, c, is also defined uniquely pointwise. The deviation
ohci
concentration is, then, just the (unique) field defined by the differ- ¼ r  ½hD  rhcii þ hD  r~ci  r  ½hvhcii þ hv~ci: ð17Þ
ot
ence between the microscale and average concentrations, i.e.
Eq. (17) is a macroscale integro-differential balance for mass. This
cðx; tÞ ¼ hcijx;t þ ~cðx; tÞ: ð11Þ
upscaled equation has the features that (1) it now represents the
With these definitions in mind, it is tempting to attempt to deter- mass balance at the scale of the support volume, V x , and (2) it is
mine the behavior of the macroscale concentration, hci, by averag- nonlocal in that the average of an average quantity depends on
ing the solution given by Eq. (7) directly. This is certainly possible points in space that are outside the domain of the averaging oper-
to do, and applying the definition of the average to Eq. (7) yields ator itself (Fig. 2). This expression is also unclosed, in that it involves
Z Z the deviation quantities of the dependent variable, ~c, which are un-
t
hcijx;t ¼ Fðy; sÞ known. Developing a means of representing the unknown devia-
0 y2V 0 tions in terms of known parameters and the dependent variable,
 Z 
1 hci, is known as closing the problem, and this process is described
 Gc ðz; y; t; sÞdA0 ðy  ys Þdz dy dt in the following section.
V z2V x
Z  Z 
1
þ CðyÞ Gc ðz; y; t; 0Þdz dy: ð12Þ 4. Closure
y2V 0 V z2V x

Although this equation shows that the average concentration can be 4.1. Development of the deviation equations
expressed in terms of an integral equation involving the initial and
boundary conditions, the microscale Green’s function Gc is still un- For this problem, the process of closure involves predicting the
known. This presents several essential problems with continuing unknown statistics of the perturbation terms that appear in Eq.
this line of investigation, as follows: (1) determining the microscale (17). Because in this work, the velocity field (and the average and
Green’s function is nearly as difficult to solve as the original set of perturbation fields associated with it) is treated as a parameter
differential equations given by Eqs. (1)–(5), (2) the microscale field, the closure problem amounts to predicting the quantity ~c.
Green’s function depends (in general) nonlinearly upon the micro- On the basis of the definitions of the average and the deviation
scale velocity vector and dispersion tensor fields, and (3) it is not concentrations, it is possible to develop a balance equation for
obvious how the average of Green’s function can be computed from the concentration deviations by subtracting the average mass bal-
quantities that are measurable in the field (e.g. such as the second- ance (Eq. (17)) from the original Darcy-scale mass balance (Eq. (1)).
order spatial statistics of the hydraulic conductivity). The concentration decomposition can be used to rewrite the
Darcy-scale mass balance equation in the form
3.1. Averaging the mass balance equation oc
¼ r  ½D  rhci þ D  r~c  r  ½vhci þ v~c: ð18Þ
ot
As an alternative to integrating the mass transport equations
and then averaging this solution, it can be advantageous to directly Subtracting Eq. (17) from Eq. (18) gives an integro-differential equa-
average the differential equations describing transport. One signif- tion for the deviation concentrations of the form
icant benefit of this approach is that the averaging process has a
o~c
more apparent physical interpretation. þ r  ðv~cÞ  r  ðD  r~cÞ ¼ r  ½D  rhci  hD  rhcii
ot
To make the averaging process itself a little easier to follow, Eq.
 r  ½vhci  hvhcii  r  ½hD  r~ci
(1) can be written in operator form as
 hv~ci: ð19Þ
oc
¼ LðcÞ; ð13Þ
ot Note that the dependent variable terms appear on the left-hand side
where L is a linear operator defined by of this expression, and the right-hand side represents ‘source’ terms
that drive the deviation equation; terms involving averages of the
LðcÞ ¼ r  ðD  rcÞ  r  ðvcÞ: ð14Þ concentration deviations are also included here. To emphasize this
Note that L could be any linear operator, including one that in- point, the deviation equation can be written as
volved derivatives of fractional order. Essentially all of the discus- o~c
sion in this paper rely only on the linearity of the operator, and  r  ðD  r~cÞ þ r  ðv~cÞ ¼ r  ð~f 1 þ ~f 2 þ IÞ; ð20Þ
ot
the main points regarding scaling laws (described below) do not
otherwise depend on the particular form of the operator. The con- where
vection–dispersion operator has been chosen to describe the ~f 1 ¼ ½D  rhci  hD  rhcii; ð21Þ
microscale mass balance because of relevance, familiarity, and
~f 2 ¼ ½vhci  hvhcii; ð22Þ
convenience.
Using the deviation defined by Eq. (11), Eq. (13) can be written I ¼ ½hD  r~ci  hv~ci: ð23Þ
as
In this form, ~f 1 , and ~f 2 represent the difference between a flux and
oc its average (in analogy with the deviations proposed previously for
¼ LðhciÞ þ Lð~cÞ ð15Þ
ot concentration and velocity), whereas the vector I represents a col-
lection of integral terms that involve the dependent variable, ~c. In
by linearity of the operator L. It is straightforward to apply the
most conventional treatments of volume averaging (e.g. [8, Chapter
average (defined by Eq. (8)) to this equation, and the result is
3]), many of these terms are neglected on the basis of length-scale
ohci constraints. Such simplifications will be discussed further in the
¼ hLðhciÞi þ hLð~cÞi: ð16Þ
ot sections that follow.
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 727

Fig. 2. The average about a point x (taken as the location of the centroid of the averaging volume) involves integration over all points in the support volume, x þ y. When the
average of an averaged quantity is computed, the support necessarily involves points that lie outside the volume with centroid x þ y þ y0 .

The complete statement of the closure problem requires Eq. ever, if one examines the results to this point in the analysis, two
(20) along with the relevant initial and boundary conditions, and features of this set of microscale–macroscale equations become
the equations describing fluid flow. Imposing these conditions, evident. First, the solution to the microscale closure problem must
the complete closure problem can be stated as follows: be determined for each point in the domain, V 0 , and this is at least
as difficult to compute as was the original microscale problem. Sec-
o~c
 r  ðD  r~cÞ þ r  ðv~cÞ ond, the solutions to the macroscale transport equation and the
ot
closure problem are coupled, indicating that the upscaled equation
¼ r  ð~f 1 þ ~f 2 þ IÞ in V 0 ; ð24Þ essentially has the same information requirements as the micro-
BC1:  n  ðv~c þ D  r~cÞ ¼ n  ðvhci þ D  rhciÞ scale equations do. This particular result has been purposeful;
þ Fðx; tÞ on A0 ; ð25Þ the point was to make it evident that averaging per se does not lead
IC ~c ¼ C0 ðxÞ at t ¼ 0; ð26Þ to a reduction in the information content of any particular prob-
lem. It is only in cases where there is some kind of redundancy
r  v ¼ 0 in V 0 ; ð27Þ
in spatial and temporal structure of the physical phenomena of
where v ¼ K  rh and h ¼ p  qgz; interest that we can reduce the information requirements of the
BC2: h ¼ HðxÞ on A0 : ð28Þ problem as a whole. To do this, we must make a very strong state-
ment about the structure of the physical phenomena of interest;
The solution to this closure problem can be given formally as
this process can be thought of as imposing a scaling law.
Z t Z
~cðx; tÞ ¼ Fðy; sÞ  ½Gc ðx; y; t; sÞ  hGc ijx;y;t;s dA0 ðy
0 y2V 0 5. Scaling laws
Z
 ys Þdy ds þ CðyÞ½Gc ðx; y; t; 0Þ  hGc ijx;y;t;0 dy: ð29Þ A scaling law is essentially an axiomatic statement about the
y2V 0
spatial and/or temporal structure of the phenomena that are ac-
Note that, in general, a closed-form expression for GC is not avail- counted for in the microscale balance equations. The assumption
able, and this solution would have to be computed numerically of a particular scaling law is one of the most significant steps in
(with the associated PDEs that define Green’s function). This indi- developing the simplified upscaled equations that we would like
cates that generally, upscaling methods in complex media are nec- to use in applications, but it is curiously under-recognized how sig-
essarily numerically-based (cf. [9]). nificant these assumptions are. Additionally, it is not always clear
To this point no approximations of any kind have been imposed what steps in the analysis introduce information about the struc-
on the upscaled equations or the associated closure problem. How- ture of the phenomena being examined. As an example, it is
728 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

frequently assumed that one can treat the finite fields (e.g. hydrau- within some allowable error bound, it is exchangeable with any
lic conductivity, velocity, concentration) associated with the phys- other suitable such volume, and (2) the process structure must
ical transport processes as if they were infinite (free space) fields. not be strongly dependent upon the conditions at the boundary
Although this is a reasonable assumption in many cases, it does of the volume. The first of these two criteria requires that the
make a very strong statement about the influence of the bound- assumptions of quasi-homogeneous and quasi-ergodic conditions
aries of the domain of interest. In essence, it says that the boundary are valid, as applied in the sense of Christakos [13, Section 2.6]
conditions for the finite domain do not strongly influence the (see also the discussion by Beran [14]), and is discussed in detail
transport within the domain. in Section 5.3 below. The second of these two requirements states
Familiar scaling laws are generally statistical ones. As an exam- that an integral measure of the system (which is generally a mac-
ple, second-order spatial stationarity in the velocity field is often roscopic scale parameter, such as the effective dispersion tensor or
assumed, and this is often coupled with the idea that volume the effective hydraulic conductivity) cannot depend strongly upon
and ensemble averages may be freely interchanged. This assump- the conditions that occur on the boundary of the REV. If the inte-
tion immediately reduces the number of degrees of freedom re- gral measures did depend upon the conditions at the boundary,
quired, because it essentially states that (1) the effective then the macroscale description of the system would be coupled
parameter fields that are developed on the basis of the statistics to the state of adjacent REVs, and, hence, to the entire system. This
of the microscale phenomena are spatially constant, and (2) the ex- would indicate completely nonlocal behavior, and the concept of
act details of the microscale are not necessary, but, rather, the sta- an REV would have little meaning in such a system.
tistics of the microscale are sufficient. Clearly this scaling law Periodic boundary conditions are often adopted as a reasonable
dramatically reduces the amount of information needed; because proxy for this second condition. The idea here is that, if the system
the microscale is sufficiently described by its second-order statis- is not strongly influenced by information from the boundaries,
tics, the exact details of the microscale become immaterial (how- then periodic boundaries are (one possible) suitable condition,
ever, knowledge of its statistics is essential). especially in light of the fact that they are weaker conditions for
Other examples of scaling laws might include assuming that the mass transport problem than, for example, Dirichlet conditions.
self-similar behavior in the structure of the physical processes, or They also provide a significant advantage when using finite Fourier
that a periodic model of the system sufficiently captures the struc- transforms in the analysis of the system of equations; however,
ture of the field. Or, some combinations of these approximations other boundary conditions may be more suitable for other kinds
can be made, for example, representing the system as a finite-sized of problems (e.g. for an example of boundary conditions used for
self-similar field with periodic boundaries. upscaling single phase flow, see Ref. [9]). An integral solution to
Generally, scaling laws must say something useful about (1) the the closure problem specified above can be developed in the Fou-
internal structure of the field, and (2) the influences of the external rier domain if we first apply the assumption that we can consider
structure (in other words, what happens at the boundary of the finite volumes of the media with periodic boundary conditions. The
field). Without being able to say something about these two fea- revised closure problem takes the form
tures of the field, it is not likely that any significant reduction of o~c
the number of degrees of freedom (the information content) of  r  ðD  r~cÞ þ r  ðv~cÞ
ot
the problem will be accomplished. To begin the discussion of the
¼ r  ð~f 1 þ ~f 2 þ IÞ in V0 ; ð30Þ
influence of scaling laws on the coarse-graining process, the fol-
lowing definition is proposed. BC1: ~cðr; tÞ ¼ ~cðr þ li ; tÞ on A0 ð31Þ
IC ~c ¼ 0 at t ¼ 0; ð32Þ
Definition. A scaling law is an axiomatic statement about the
r  v ¼ 0 in V0 with v ¼ K  ðrp  qgÞ; ð33Þ
structure of a field for the purposes of identifying redundant
information that might be eliminated during the upscaling process. BC2: v~ðrÞ ¼ v
~ðr þ li Þ on A0 ; ð34Þ
Scaling laws must be consistent with what is possible or Constraint hv~i ¼ 0 in V0 ð35Þ
observable for the field in question. BC3: pðr þ li Þ ¼ pðrÞ þ djk Dp on A0 ;
Note that a scaling law should be thought of as being the best pðx0 Þ ¼ p0 ; x0 2 A0 ; i ¼ 1; 2; 3: ð36Þ
description of what is known. However, such laws are essentially Here li is a lattice vector defining the periodic structure of the con-
impossible to prove valid, although they can be invalidated by centration deviation, pressure, and velocity fields. At one face
observations, necessitating the need for an improved scaling law (where djk ¼ 1; j indexes all control faces, and k indicates the spec-
(cf. [10,11]). ified control face) a constant Dp is added to the periodic pressure
For this work, we will make progress by assuming that the sys- so that there is a net pressure gradient driving flow. Note that in
tem can be treated by examining a finite volume of media and the closure problem above, we have assumed that there is no influ-
imposing periodic boundaries. Note that this essentially incorpo- ence of the initial condition on the effective dispersion tensor. This
rates the infinite-field case, since one can always let take the limit approximation may not be strictly true, and should be explored
of the result as the period of the volume under consideration tends further.
toward infinity [12]. With this approximation, we can continue our The solution can be found by finite Fourier transform tech-
analysis in the context of finite Fourier transforms. niques [15,16, Section 192], which is a generalization of Fourier
series techniques. For a function f, the finite Fourier transform,
5.1. Scaling law 1: Boundary influence Fðf Þ, is defined by
Z  
ip
In upscaling, the term representative elementary (or unit) vol- f ðkÞ ¼ F½f ðxÞ ¼ f ðyÞ exp  k  y dy ð37Þ
ume is frequently encountered, but rarely discussed in significant y2Vðx;pÞ p
detail. It is clear that an REV is representative in some sense, but and the inverse is defined by
it is not always clear what in particular are the conditions for ‘rep- 1 X
resentativeness’. Here, we propose that two features are required f ðxÞ ¼ F1 ½f ðkÞ ¼ 3
f ðkÞ expðik  xÞ: ð38Þ
ð2pÞ k2Z3
for a volume to be representative: (1) the volume must capture en-
ough of the essential structure of the physical process (as qualified For this notation, Vðx; pÞ indicates the bounds of the periodic vol-
by some integral measure of the system) in the domain such that, ume (a cube with sides of length 2p), and Z3 indicates the set of
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 729

integer vectors of the form z ¼ ðzx ; zy ; zz Þ. For periodic volumes, requirements on the structure of the physical processes that we
many of the finite Fourier transforms are identical to their free are interested in.
space analogues. Note that Fourier analysis has been used effec-
tively by several researchers (e.g. [17]) to examine the details of 5.2. Scaling law 2: Statistical homogeneity of the parameter fields
the coarse-graining procedure.
In the development that follows, we follow a similar analysis We often think of statistical homogeneity (or spatial stationa-
presented previously by Wood et al. [12]. To begin, the finite Fou- rity) in the context of second-order normal or log-normal spatial
rier transform of Eq. (30) is given by statistics. But, there is no requirement that these are the only sta-
tistical models that we can think of as being spatially stationary.
d~c 1 X p2 0 0 0 0 0
¼ k  Dðk Þ  k ~cðk  k Þdk For example, one could employ the idea of a power-law distribu-
dt ð2pÞ3 0 3 p2 k 2Z tion of statistics for some system property, and then insist that
1X ik0 p ikp ~ the statistics of this distribution be spatially stationary.
 ðf 1 þ ~f 2 þ IÞ: ð39Þ
0 0 0
 3
 vðk Þ~cðk  k Þdk  To start, we define two additional decompositions for the veloc-
ð2pÞ k0 2Z3 p p
ity vector and dispersion tensor fields as follows:
Note that we have used the convolution property for finite Fourier ~ðxÞ;
vðxÞ ¼ v0 þ v ð44Þ
transforms [15] which follows directly from Parseval’s identity
e
DðxÞ ¼ D0 þ DðxÞ; ð45Þ
Z  
ip
fgðkÞ ¼ f ðyÞgðyÞ exp  k  y dy where hvi ¼ v0 and hDi ¼ D0 are constants in space. Note that this
y2Vðx;pÞ p
latter condition can be relaxed somewhat to admit non-constant
1 X
¼
0
f ðkÞgðk  k Þ: ð40Þ mean fields when there is a separation of length scales (see Section
ð2pÞ3 k0 2Z3 5.3 below). This is consistent with the concept of quasi-stationarity
(cf. [13]), and this has been explored in detail by Whitaker [8, Chap-
An implicit solution to Eq. (39) can be found by integrating the ter 1].
equation once. The result is a formula for the deviation concentra- The closure problem can now be written
tion of the form
2 o~c e  r~cÞ  r  ðv
Z t 0 ¼t X p2 0  r  ðD0  r~cÞ þ r  ðhvi~cÞ ¼ r  ð D ~~cÞ
~cðk; tÞ ¼  4 1 0 0 0
k  Dðk Þ  k ~cðk  k Þdk
0 ot
t 0 ¼0
3
ð2pÞ k0 2Z3 p2 þ r  ð~f 1 þ ~f 2 þ IÞ; ð46Þ
3
X ik0 p
1 ikp ~ where now
 ðf 1 þ ~f 2 þ IÞ5dt 0 :
0 0 0
þ  vðk Þ~cðk  k Þdk þ
3
ð2pÞ k0 2Z3 p p ~f ¼ ½D  rhci þ D e  rhci  D0  rhhcii  h D e  rhcii; ð47Þ
1 0
~f 2 ¼ ½v0 hci þ v
~hci  v0 hhcii  hv~hcii; ð48Þ
ð41Þ
The transform, ~c is inverted back to real space by e
I ¼ ½D0  rh~ci þ h D  r~ci  v0 h~ci  hv
~~ci: ð49Þ
1 X Additional progress can be made by noting that, from Eq. (11), we
~cðx; tÞ ¼ ~cðk; tÞ expðik  xÞ: ð42Þ
ð2pÞ3 k2Z 3
have
hhcii ¼ hci  h~ci: ð50Þ
The application of periodic conditions has helped the situation
some in that now there is no explicit dependence upon the conditions Combining Eqs. (46)–(50) results in a dramatic simplification to the
at the boundaries of the REV as there was for Green’s function solu- closure problem
tion given by Eq. (29). However, the result is still very complex, and,
o~c e  r~ci
e  r~c  h D
it still does not reduce the number of degrees of freedom involved  r  ðD0  r~cÞ þ r  ðv0 ~cÞ ¼ r  ½ D
ot
in the upscaled solution. This is apparent because the Fourier trans-
 r  ½v~~c  hv e  rhci
~~ci þ r  ½ D
form of the Darcy-scale dispersion tensor and velocity vector fields
are required, and this represents complete microscale information. e
 h D  rhcii  r  ½v ~hci  hv~hcii:
Additionally, the macroscale transport equation itself is given ð51Þ
by
The macroscale transport equation can also be revised using the
ohci approximations above and the relationship hhcii ¼ hci  h~ci
¼ r  ½hD  rhcii þ hD  r~ci  r  ½hvhcii þ hv~ci: ð43Þ
ot ohci e  rhcii þ h D
¼ r  ½D0  rhci þ h D e  r~ci
This equation is still nonlocal, and it is not at all clear that it will take ot
a conventional form (i.e., one that is proportional to hci and its gra-  r  ½v0 hci þ hv
~hcii þ hv
~~ci:
dients) once ~c has been eliminated via Eq. (41); in other words, the
Although nothing at this point prevents us from developing an im-
microscale and macroscale solutions appear to be fully coupled.
plicit solution for the deviation concentration given by Eq. (51), as
The reason that the microscale and macroscale solutions re-
was done for the more general case in Section 5.1, it is apparent that
main fully coupled is that no assumptions regarding the structure
any implicit solution will require the complete solution for ~c at the
of the interior of the domain have yet been adopted. So far the
microscale. Therefore, the microscale and macroscale solutions are
concept of redundancy in the structure of the physical processes
still fully coupled, and no reduction in the degrees of freedom is
within the REV has not yet be imposed. If some notion of redun-
yet possible.
dancy cannot be developed, then there is little hope for reducing
the amount of information required to describe the state of the
system in some compact sense. Thus, the averaging itself is not 5.3. Scaling law 3: Separation of length scales I
sufficient to provide a simpler macroscale solution to the prob-
lem. Although a macroscale solution can be developed, it does In order to develop an explicit representation for ~c, we need to
not represent any reduction in the information content of the have a means of eliminating products of the ~c field with other
solution. To make further progress, we must impose additional spatially-dependent fields (such as v e As a first step in that
~ and D).
730 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

direction, we consider the scaling law referred to generally as These ideas can be used to simplify the closure problem as follows.
separation of length scales. To define what is meant by a separation First, note that the following two order-of-magnitude estimates can
of length scales, we first, define the following correlation functions be made:
for any microscale field property, m, and macroscale field property, Z !
M, by 1 ~ ‘3
v
Z ~hcii ¼
hv ~hcidy ¼ O
v hci; ð60Þ
1 1 V y2VðxÞ r 30
RMM ðx; yÞ ¼ ½MðxÞ  hMijx  !
rM ðxÞrM ðyÞ V y2VðxÞ
Z e 3
e  rhcii ¼ 1 e  rhcidy ¼ O D‘
 ½Mðx  yÞ  hMijðxyÞ dy; ð52Þ hD D  rhci: ð61Þ
V y2VðxÞ r 30
Z
1 1
RMm ðx; yÞ ¼ ½MðxÞ  hMijx 
rM ðxÞrm ðyÞ V y2VðxÞ Then note that we have the following restrictions:
 ½mðx  yÞ  hmijðxyÞ dy: ð53Þ ~hcii  v
hv ~hci; ð62Þ
Note that, although we have imposed statistical spatial homogeneity e
h D  rhcii  D e  rhci: ð63Þ
(or, equivalently, stationarity) for the parameter fields, we have not
yet imposed any particular conditions on the concentration field. On the basis of the estimates developed above, these restrictions are
Therefore, we have maintained, in general, a dependence on both valid when
x and y. ‘  L; ð64Þ
For these correlation functions, the following integral scales can
be defined: which is true when the conditions of separation of length scales is
Z valid. The closure problem then takes the form
IMM ðxÞ ¼ RMM ðx; yÞdy; ð54Þ
y2VðxÞ
o~c e  r~c  h D
e  r~ci
Z  r  ðD0  r~cÞ þ r  ðv0 ~cÞ ¼ r  ½ D
ot
IMm ðxÞ ¼ RMm ðx; yÞdy: ð55Þ  r  ½v~~c  hv e  rhciÞ
~~ci þ r  ð D
y2VðxÞ
~
 r  ðvhciÞ: ð65Þ
With these definitions, the separation of length scales is defined as
follows. The macroscale transport equation can also be simplified on the ba-
Definition. Suppose that a macroscale property MðxÞ has integral sis of the scaling laws presented in this section. For these simplifi-
scale IMM ¼ LðxÞ, and that the integral scale for the correlation of cations, we need to adopt restrictions similar to the inequalities
MðxÞ and a microscale property, mðxÞ, is IMm ¼ ‘ðxÞ, with ‘  L. For given by (62) and (63).
an averaging support scale of characteristic length r0 , there is said ~hcii  v0 hci;
hv ð66Þ
to be a separation of length scales between the macroscale e  rhcii  D0  rhci:
hD ð67Þ
property, MðxÞ and the microscale property mðxÞ, when
‘ðxÞ Note that these constraints are always met when the more restric-
 1: ð56Þ
LðxÞ tive constraints given by (62) and (63) are. The new macroscale
A separation of length scales does not always occur in natural and form of the transport equation is
engineered systems (e.g. fractal systems would not have this char- ohci e  r~ci  r  ½v0 hci þ hv
¼ r  ½D0  rhci þ h D ~~ci: ð68Þ
acteristic), but for many systems this approximation is a reasonable ot
one. When a separation of length scales exists, we might also hope
Although the closure problem (Eq. (65)) and the macroscale trans-
that the correlation functions are not strongly dependent upon spa-
port equation (Eq. (68)) are simplified compared with those pre-
tial location. Although we have already imposed this for correla-
sented in Sections 5.1 and 5.2, there are still two complicating
tions of the velocity vector and dispersion tensor fields, it is not
issues. First, as with the previous solution, an explicit analytical
automatically true for quantities that involve the concentration.
solution to this problem does not appear to be possible. Second, be-
We make the following definition to impose that condition. e appear in the dif-
~ and D
cause the bare parameter field deviations v
ferential equation, the solution to both ~c and hci still depends
Definition. Under the conditions given in (56), when the macro-
pointwise upon these microscale fields. This latter problem is miti-
scale variable is the concentration field hci, then the macroscale
gated partly by the fact that it is not ~c that appears in the macro-
concentration changes slowly in space relative to the microscale
scale transport equation (Eq. (68)) but rather the average
parameter field. If we also have the condition that e  r~ci and hv
~~ci.
quantities h D
OðrIMm Þ 6 OðrIMM Þ; ð57Þ
then we can consider the concentration field and its cross-correla- 5.4. Scaling law 4: Quasi-ergodicity and the existence of an REV
tions with the parameter fields (second order) quasi-homogeneous
(cf. [13, Chapter 2.7, 18, Chapter 1.10]). There is an additional issue regarding the correlation functions;
they will be most useful if their value does not depend strongly
Using order-of-magnitude estimates, we can think of this upon the size of the averaging volume. To accomplish this, we im-
restriction (n.b., for a discussion of restrictions versus constraints pose a condition of quasi-ergodicity (cf. [13]).
see [19]) as being roughly the condition
Definition. (Second order) quasi-ergodicity is a scaling law that
DIMm DIMM
6 : ð58Þ assumes that the support of an averaging volume is large enough
‘ L
such that it fully samples the (second order) statistics of the
Using DIMm ¼ rIMm and DIMM ¼ rIMM , where rIMm and rIMm are the
parameter of interest with a high degree of confidence. Under such
standard deviations of the two integral scale fields, then this restric-
conditions, one can assume that the mean and covariance (or
tion can be put in the form of the constraint
correlation) structures do not depend strongly on the size of the
rI M m ‘ averaging volume. A quasi-ergodic system must also be quasi-
6 : ð59Þ
rI M M L homogeneous.
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 731

The constraint associated with this definition is o~c e  rhciÞ  r  ðv


 r  ðD0  r~cÞ þ r  ðv0 ~cÞ ¼ r  ð D ~hciÞ; ð76Þ
ot
‘  r0 ; ð69Þ
BC1 ~cðr; tÞ ¼ ~cðr þ li ; tÞ on A0 ; ð77Þ
where r 0 is the characteristic length associated with the averaging IC ~c ¼ 0 at t ¼ 0: ð78Þ
volume. We will assume that the parameter fields, the concentra-
tion field, and all cross-statistics are quasi-ergodic. The solution to this problem can be found using finite Fourier trans-
The assumption of spatial quasi-stationarity and quasi-ergo- form methods, as described above, and the approach follows closely
dicity would allow us to eliminate some degrees of freedom if the derivation presented first by Deng et al. [21] and subsequently
it were also true that the characteristic size of the averaging vol- in the volume averaging context by Wood et al. [12], Appendix B. In
ume were sufficiently small relative to the characteristic size of contrast to the latter work, however, we will not assume that the
the entire system, V 0 . However, we have not yet imposed any gradient of the average concentration can be treated as a constant.
such condition; we will consider such a restriction below. As it Following [12], the finite Fourier transform of Eq. (76) takes the
stands, the size of the averaging volume necessary for quasi-ergo- form:
dic conditions to hold could conceivably be as large as the volume d~c
of interest itself. Because we have no explicit solution for ~c, we þ b0 ðkÞ  k~c ¼ Gðk; tÞ; ð79Þ
dt
need to compute ~c at the microscale for each point within the
averaging volume. In that case, we would have not yet have where
achieved any reduction in the number of degrees of freedom re- p p2
b0 ðkÞ  k ¼ iv0  k þ 2 k  D0  k; ð80Þ
quired to describe the system. p p
Z
Gðk; tÞ ¼ e
ðr  ð DðyÞ  rhcijy;t Þ  v
~ðyÞ  rhcijy;t Þ
5.5. Scaling law 5: Smallness of variance in the field properties
y2Vðx;pÞ
 
ip
A simplification that is frequently employed, but has not as yet  exp  k  y dy: ð81Þ
p
been adopted, is the assumption that the variance of the parameter
field properties is in some sense small. For the development here, In Eq. (81), we have used the incompressibility condition for the
this means that the deviations quantities are small relative to their flow field. Using the convolution property defined above, we can ex-
means. In other words, we impose the conditions press G as
kv~k 1 X
 1; ð70Þ Gðk; tÞ ¼ ~ 0 Þ  k0 hcij 0 ;
bðk ð82Þ
3 kk ;t
kv0 k ð2pÞ k0 2Z3
e
k Dk
 1; ð71Þ where here
kD0 k  
~ 0 Þ  k0 ¼ ip v
bðk
p2 e 0 Þ  k0 :
~ðk0 Þ  k0 þ k0  Dðk ð83Þ
where k  k is an appropriate norm for the velocity vector and dis- p p2
persion tensor fields. The usual norm for the deviation quantities
is taken to be the standard deviation associated with these fields, A single integration (via an integrating factor method) of Eq. (79)
whereas the usual norm for the average quantities, v0 and D0 , are yields the result (following the notation of [22,12])
given by conventional inner products. These constraints can be ex- Z t 0 ¼t
pressed by a restriction in terms of the coefficient of variation of the ~cðk; tÞ ¼ Bðk; t  t0 ÞGðk; t 0 Þdt 0 ; ð84Þ
t 0 ¼0
two fields as follows:
r~ where
v
pffiffiffiffiffiffiffiffiffiffiffiffiffi  1; ð72Þ
v0  v0 Bðk; t  t 0 Þ ¼ exp½b0 ðkÞ  kðt  t 0 Þ: ð85Þ
reD
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  1: ð73Þ The inversion of ~c is given by Eq. (42)
1 Z
2 0
D : D0 t 0 ¼t X
1
~cðx; tÞ ¼ 3
Bðk; t  t 0 ÞGðk; t 0 Þ expðik  xÞdt 0 : ð86Þ
Here rv~ and r e are the largest components of the standard devi- t0 ¼0 ð2pÞ k2Z 3
D
ation for the velocity vector and dispersion tensor, respectively.
Note that when (72) is valid and the distribution of hydraulic con- Noting that the inverse transforms of B and G, respectively are
ductivity is log-normally distributed, it is possible to relate the 1 X
velocity variance directly to the variance of the (logarithm of Bðy; t  t0 Þ ¼ Bðk; t  t0 Þ expðik  yÞ; ð87Þ
ð2pÞ3 k2Z
the) hydraulic conductivity, as shown previously by Gelhar and
1 X
Axness [20]. Gðy; tÞ ¼ Gðk; tÞ expðik  yÞ: ð88Þ
When the constraints given by the inequalities (72) and (73) are ð2pÞ3 k2Z
met, the closure problem can be simplified to a form that admits an
By Eqs. (85) and (82), we can express B and G, respectively by
explicit solution. In particular, these constraints allow one to make
the restrictions 1 X
Bðy; t  t0 Þ ¼ 3
exp½b0 ðkÞ  kðt  t 0 Þ expðik  yÞ; ð89Þ
e  r~c; D0  r~c  h D
e  r~ci; ð2pÞ 3
D0  r~c  D ð74Þ k2Z

~ ~ ~ ~ ~ ~ 1 X 1 X
~ 0 Þ  k0 hcij 0 expðik  yÞ:
v0 c  vc; v0 c  hvci: ð75Þ Gðy; tÞ ¼ bðk kk ;t ð90Þ
3 3
ð2pÞ k2Z ð2pÞ 0
k 2Z 3
Under these conditions, we are able to develop a balance equation
for the deviations that (1) involves only constant parameters in From the convolution theorem (applied, in this case, in the inverse),
terms involving ~c, and (2) involves source terms that are simple the expression for G can be reduced to the form
functions of a deviation of a parameter field and the average or gra- " #
1 X 0 0
dient of the average concentration. The resulting closure problem Gðy; tÞ ¼ ~
bðk Þ expðik  yÞ  rhcijy;t : ð91Þ
takes the form ð2pÞ3 k0 2Z
732 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

Finally, by Parseval’s identity [15], Eq. (86) is then ume must be small relative to the large length scale, L. The impli-
Z Z t0 ¼t cation here is that the nonlocal transport equation given by Eqs.
~cðx; tÞ ¼ F1 ½Bðy; t  t0 ÞF1 ½expðik  xÞGðy; t0 Þdt0 dy (95)–(97) is appropriate when the size of the averaging volume
y2VðxÞ t 0 ¼0 is comparable to the length scale associated with appreciable
Z Z t0 ¼t
changes in hci.
¼ Bðy; t  t0 ÞGðx þ y; t 0 Þdt 0 dy As a special case of this expression, if we consider the condi-
y2VðxÞ t 0 ¼0
Z Z tions where De  0 and we set
t0 ¼t
¼ Bðy; t  t0 Þ 1 X
y2VðxÞ t 0 ¼0 Bðx  y; t  t 0 Þ ¼ exp½b0 ðkÞ  kðt  t 0 Þ exp½ik
" # ð2pÞ 3
1 X k2Z3
~ 0 Þ exp½ik0  ðx þ yÞ  rxþy hcij 0
 bðk xþy;t 0 dt dy;  ðx  yÞ; ð98Þ
ð2pÞ3 0
k 2Z

ð92Þ then Eqs. (95) and (97) take the form

where ohci
¼ r  ðD0  rhciÞ  v0  rhci þ r  hv ~~ci; ð99Þ
ot
1 X Z Z t0 ¼t
Bðy; t  t 0 Þ ¼ exp½b0 ðkÞ  kðt  t 0 Þ expðik  yÞ: ð93Þ
ð2pÞ3 ~ðxÞi ¼
h~cðx; tÞv Bðx  y; t  t 0 Þ
k2Z3 y2VðxÞ t 0 ¼0

Here, the symbol rxþy is used for the gradient operator to prevent  hv ~ðxÞi  ry hcijy;t0 dt0 dy:
~ðyÞv ð100Þ
any confusion about the variable of dependence. Finally, making a
Note this last form is identical to the case presented previously by
transformation of variables and noting that the last term in Eq.
~ this solution can be written in Deng and Cushman [22] and Deng et al. [21] from the ensemble
(92) is the inverse transform of b,
averaging perspective. Although the results in those works were
the form
developed using ensemble averaging methods, scaling law 3
Z Z t 0 ¼t
(imposing quasi-ergodicity) makes the ensemble and volume aver-
~cðx; tÞ ¼ Bðx  y; t  t0 Þ  ½v
~ðyÞ  ry hcijy;t0
ages interchangeable. Note that some investigators have developed
y2VðxÞ t 0 ¼0
e methods that predict a probabilistic interpretation of the concentra-
þ r  ð DðyÞ  ry hcijy;t0 Þdt0 dy: ð94Þ
tion (which they relate informally to information loss), and under
Scaling law 4 allows us to eliminate terms that are small, relative to such conditions one may avoid imposing scaling laws 3 and 4 (cf.
other terms, due to the imposed constraint of small variance (or, in [23]). The notion of physically realized concentration that arises
our case, coefficient of variation) of the velocity vector and disper- in such macroscale balance laws, however, must be abandoned in
sion tensor fields. Although this does allow some simplifications to favor of a probabilistic interpretation.
the closure problem, it does not allow any further simplification to The upscaled equations given by Eqs. (95)–(97) or Eqs. (99) and
the macroscale equation. Recall, the macroscale equation is of the (100) are at last a representation that will reduce the number of
form degrees of freedom required to represent the system, provided that
the statistics of the parameter fields can be estimated. The reason
ohci e  r~ci  r  ½v0 hci þ hv that these expressions successfully accomplish the reduction of
¼ r  ½D0  rhci þ h D ~~ci: ð95Þ
ot information required is that they are the first ones in which the
To close this problem, we need to be able to eliminate ~c in this necessary statistics needed to parameterize the solution have been
expression. Using Eq. (94), we can compute the quantities explicitly decoupled from the average concentration.
e  r~ci and hv
hD e
~~ci by first multiplying by DðxÞ ~ðxÞ, respectively,
and v
and then averaging. The result is 5.6. Scaling law 6: Smoothness of the average concentration field
Z Z t 0 ¼t
(separation of length scales II)
e  r~ci ¼
hD e  rx Bðx  y; t  t 0 Þv
hD ~ðyÞi  ry hcijy;t0
y2VðxÞ t0 ¼0 As a final approximation, we consider the case where there is a
e
þ h DðxÞ e
 rx Bðx  y; t  t0 Þðry  DðyÞÞi  ry hcijy;t0 separation of length scales between the size of the averaging vol-
e e ume, r 0 , and the macroscale correlation length, L. We start with
þ h DðxÞ  rx Bðx  y; t  t0 Þ DðyÞi : r2y hcijy;t0 dt 0 dy; ð96Þ e  0).
Eq. (100) (still assuming, without loss of generality, that D
Z Z t0 ¼t
The question is, under what conditions can the gradient of the
~ðxÞi ¼
h~cðx; tÞv Bðx  y; t  t0 Þhv
~ðyÞv~ðxÞi  ry hcijy;t0
y2VðxÞ t0 ¼0 average concentration be treated as a constant for the computation
e of the integral in Eq. (100). Deng and Cushman [22], considered
þ Bðx  y; t  t0 Þhðry  DðyÞÞv~ðxÞi  ry hcijy;t0
this question, and indicated that the concentration can be removed
e v
þ Bðx  y; t  t0 Þh DðyÞ ~ðxÞi : r2y hcijy;t0 dt 0 dy: ð97Þ under the conditions where the characteristic length scale of the
velocity autocovariance is much smaller than the characteristic
In Eq. (97) we have defined the following outer products: v ~v~ ¼ ti tj
ev length scale for the gradient of the average concentration. Quintard
and D ~ ¼ Dij tk . Note that by combining Eqs. (95)–(97), the explicit
and Whitaker [24] also considered this question in the context of
dependence upon the microscale concentration, ~c, has been elimi-
volume averaging, and concluded that the approximation was va-
nated. The right-hand side of Eqs. (96) and (97) depends only upon
lid when r0  L. To see this latter result, one begins by generating
(1) the statistics (and cross-statistics) of the velocity vector and dis-
a Taylor series for rhci as follows:
persion tensor fields, (2) the kernel operator B, and (3) the average
concentration hci and its first two spatial derivatives. Therefore, the ry hcijy;t0 ¼ rx hcijx;t0 þ ðy  xÞ  rx rx hcijx;t0 þ    : ð101Þ
macroscale transport equation is explicitly closed with respect to
After substituting this into the integral expression given by Eq.
the dependent variable hci.
(100), an order-of magnitude estimate of the second term in the
This result is an explicit nonlocal theory that allows the predic-
expansion yields
tion of the evolution of hci. The assumptions required for this the-
r
ory to be valid are essentially that scaling laws 1–4 be valid. We ðy  xÞ  rx rx hcijx;t0 ¼ O
0
rx hcijx;t0 : ð102Þ
have not, to this point, assumed that the size of the averaging vol- L
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 733

 
Within the integral, then, terms in the Taylor series higher than first L
T ¼ O : ð114Þ
order can be dropped whenever hvi
ðy  xÞ  rx rx hcijx;t0  rx hcijx;t0 ð103Þ Then, it can be seen that the time scale inequality posed by Eq.
(112) is equivalent to the separation of length scales imposed
or, equivalently,
previously
r0
 1: ð104Þ ‘
L  1: ð115Þ
L
Note that combining this constraint with those given by the
The separation of length scales given by Eq. (105) allows the mac-
inequalities (56) and (69) gives the classical constraint adopted in
roscale equation to be expressed in the classical form
a wide variety of homogenization problems (cf. [18, Chapter
1.10,8, Chapter 1.3, 6, Chapter 4]) ohci
¼ r  ðDeff  rhciÞ  v0  rhci; ð116Þ
ot
‘  r 0  L: ð105Þ
where the total (effective) dispersion tensor is given by
With the approximations developed above, Eq. (100) is given by Z Z t0 ¼t
Z Z t 0 ¼t Deff ðx; tÞ ¼ D0 þ Bðx  y; t  t 0 Þ
~ðxÞi ¼
h~cðx; tÞv Bðx  y; t  t0 Þhv
~ðyÞv
~ðxÞi y2VðxÞ t 0 ¼0
y2VðxÞ t 0 ¼0 ~ðxÞidt0 dy:
~ðyÞv
 hv ð117Þ
 rx hcijx;t0 dt 0 dy: ð106Þ
Finally, note that because of the previously imposed assumption of
Some simplification has been achieved here; however, to obtain a statistical homogeneity (stationarity), we must have the condition
local equation, we must establish the conditions under which the that the velocity correlation function depends only upon the rela-
average concentration can be removed from the remaining time tive displacement, i.e.,
integral. To do this, we begin by constructing a Taylor series in time
~ðyÞv
Rðx  yÞ ¼ hv ~ðxÞi: ð118Þ
(similar to the space expansion above)
o In this case, the total dispersion tensor is a function of only time, as
rx hcijx;t0 ¼ rx hcijx;t þ ðt0  tÞ rx hcijx;t þ    : ð107Þ shown, for example, by Deng et al. [21]; explicitly, this is given by
ot
(after a small amount of algebra and using the symmetry of B and R)
Substituting this result into Eq. (106) gives Z Z
t 0 ¼t
Z Z t 0 ¼t Deff ðtÞ ¼ D0 þ Bðz; t  t0 ÞRðzÞdz dt0 : ð119Þ
~ðxÞi ¼
h~cðx; tÞv Bðx  y; t  t0 Þhv ~ðxÞidt 0 dy
~ðyÞv t 0 ¼0 z2VðxÞ
y2VðxÞ t 0 ¼0
Z Z t 0 ¼t
 rhcijx;t þ Bðx  y; t  t0 Þ 6. Discussion
y2VðxÞ t0 ¼0

o It is interesting to note that it is possible to generate a com-


 hv ~ðxÞiðt 0  tÞdt0 dy 
~ðyÞv rhcijx;t þ    : ð108Þ
ot pletely general upscaled equation for solute transport in heteroge-
To obtain a conventional local equation, we would like to impose neous media that applies regardless of the structure of the
the restriction heterogeneity. An expression for the averaged solute transport
Z equation and its associated closure is given by Eqs. (17) and (29),
t 0 ¼t
Bðx  y; t  t 0 Þhv ~ðxÞidt0  rhcijx;t
~ðyÞv respectively. Although the development of an averaged equation
t 0 ¼0 under these substantially nonrestrictive conditions is possible,
Z t0 ¼t the developments above show that the averaged equation itself
o
 Bðx  y; t  t 0 Þhv ~ðxÞiðt 0  tÞdt0 
~ðyÞv rhcijx;t : ð109Þ does not result in any reduction of the numbers of degrees of free-
t 0 ¼0 ot
dom required to describe the system. This is an interesting result,
Note that the first integral can be approximated by because it suggests that the averaging process itself does not sim-
Z t 0 ¼t plify the description of suitably complicated systems.
Bðx  y; t  t 0 Þhv ~ðxÞidt0  rhcijx;t ¼ Oðr2v~ t rhciÞ;
~ðyÞv ð110Þ What, then, does reduce the complexity of the description of a
t 0 ¼0
heterogeneous system? The results above suggest that it is the
where t is the integral time scale for the velocity correlation. Sim- scaling laws that are imposed rather than anything specific about
ilarly, the second integral can be estimated by the mathematical method employed. In fact, it appears to be the
Z t 0 ¼t case that successful upscaling is a function not of the particular
o
Bðx  y; t  t 0 Þhv ~ðxÞiðt 0  tÞdt0  rhcijx;t
~ðyÞv mathematical framework adopted for analysis, but, rather, the
t 0 ¼0 ot temporal spatial structure of the fields associated with the mass
 
rhci transport process.
¼ O r2v~ ðt Þ2 ; ð111Þ
T In this work, six different scaling laws have been introduced in
the process of coarse-graining the mass transport in a heteroge-
where T represents the integral time scale for the macroscopic con-
neous system. Five of these scaling laws (influence of boundaries,
centration field. Using these two estimates, the inequality given by
statistical homogeneity, separation of length scales, ergodicity,
Eq. (109) can be expressed by the simple constraint
smallness of the variance [or coefficient of variation]) were re-
t quired before any reduction in the number of degrees of freedom
 1: ð112Þ
T could be seen. With the assumption of these five scaling laws, it
It is interesting to note that these time scales can be estimated in was possible to generate a space and time nonlocal macroscale
terms of the corresponding (spatial) integral scales as follows : model where the number of degrees of freedom required to de-
  scribe the system were reduced. Interestingly, the only difference
‘ between the nonlocal model and the local one was the require-
t ¼ O ; ð113Þ
hvi ment that the average concentration field be smooth enough (i.e.,
734 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

r0  L). Under these conditions, the nonlocal model can be con- As an example, consider a 3-dimensional discrete representa-
verted to a local one. The other five scaling laws were required tion of a subsurface hydrologic system with N d ¼ 10003 (1  109 )
for both models. computational nodes at the microscale, and assume that one has
It is useful to consider a meta-model of the upscaling process it- perfect microscale information. If one were to adopt a 27 point fi-
self. For the purposes of discussion, it is easier to think about the nite difference stencil, then a finite difference representation of
discretized version of these equations for developing notions of this system at the microscale would require 27  109 degrees of
information content. However, formal definitions of information freedom. Given the complexity of geologic systems, such a number
content of continuous fields do exist (e.g. [25,26,27]), and such def- of degrees of freedom is not at all unrealistic (cf. [28]). If a set of
initions could also be adopted for describing the reduction of infor- scaling laws allowed the development of a macroscale model, with
mation in these systems as various scaling laws were adopted. a factor of 50 reduction in the number of degrees of freedom in
In Fig. 3, a graphical representation of the upscaling process is each direction, then the macroscale system would have
presented, with a focus on organizing the material in terms of N D ¼ 8  103 compute nodes.
information content. The symbols N; R, and M can be interpreted For a nonlocal model, the numerical representation involves a
as the information content (or degrees of freedom) of the original full matrix; this results because the value at each point in the dis-
microscale model, the redundant microscale information, and the crete system is coupled to the value of all other points in the sys-
macroscale model, respectively. It is useful to note that the concept tem through the convolution integral terms. Thus the macroscale
of information redundancy can be defined for both deterministic nonlocal model would contain almost 512  109 degrees of free-
and stochastic systems. Although explicit definitions for redun- dom. This actually represents an increase in the number of degrees
dancy exist, for this work it is sufficient to think of redundancy of freedom required, even though this model is technically a
as meaning that a statistical (rather than deterministic) represen- coarse-grained one. If a nonlocal model were absolutely required
tation of the structure of an REV is sufficient. The important thing (i.e., the scaling law associated with developing a local model
to note here is that under these conditions, scaling laws can be was not valid), then it may actually be more efficient to solve the
used to eliminate redundant information. When this is the case, microscale model directly for such a case.
then R > 0, and some reduction of information content may be If, however, a local coarse-grained model were valid, then a sig-
possible. If no scaling law is valid, then R ¼ 0, and no reduction nificant reduction in the number of degrees of freedom might be
of information content is possible by averaging. recognized. Continuing with the example initiated above, if a 27
Note that the information content of the closure problem is gi- point finite difference scheme were again assumed for the local
ven by ðN  RÞ, indicating the reduction in information content in macroscale model, then there would be a total of only 216  103
the solution to the closure problem itself. However, the macroscale degrees of freedom. This represents a dramatic reduction in the
equation also contains degrees of freedom, and these serve to in- information content of the coarse-grained system as compared
crease the total in spite of the reduction of information accom- with the microscale.
plished during closure. This is not an immaterial point; in The reduction of the number of degrees of freedom, then, is not
principle, it is possible to successfully reduce the number of degrees necessarily guaranteed even for cases where large reductions in
of freedom in a problem at the closure level, but still obtain a result the information content at the closure level (the redundant infor-
that does not reduce the overall number of degrees of freedom. mation, R) is possible. It is necessary that one also considers the

Fig. 3. A meta-model for the upscaling process. The information content (or, equivalently, number of degrees of freedom) at each stage are listed in the upper portion of each
box.
B.D. Wood / Advances in Water Resources 32 (2009) 723–736 735

impact of the form of the macroscale model to assure that the rela- for averaging than it does the assumptions (scaling laws)
tionship ðN þ M  RÞ < N (or, equivalently, that ðM  RÞ < 0) holds, adopted when developing the upscaled model. This is
indicating that a reduction of information content can actually be important especially in subsurface hydrology, because there
realized. seems to be a particular (although hardly purposeful) ten-
It should be noted that this entire discussion is predicated on dency to subvert the conceptual model in favor of emphasiz-
the idea of reducing the number of degrees of freedom when per- ing the mathematical methods that are used.
fect microscale information is available. Of course, this is not ordi- (2) Developing averaged equations for mass transport is gener-
narily the case for characterization of natural subsurface systems, ally possible for systems of arbitrary complexity (nonphysi-
and this presents an interesting dilemma. It may be reasonable cal pathological parameter fields being excepted), with
to use the nonlocal model discussed above because it relies on essentially no assumptions being imposed. Correspondingly,
the statistics of the field (which may be available from measure- however, there is also, in the general case, no reduction in
ments) rather than microscale knowledge of field properties the information content of the macroscale representation,
(which, in general, are not available). However, because of the and so the averaged result has no practical value.
assumptions of quasi-stationarity and quasi-ergodicity that must (3) For the problem of nonreactive solute transport in heteroge-
be imposed even for the nonlocal model, in principle any micro- neous media, five different scaling laws were identified as
scale structure realized using the measured statistics should be being necessary (in this approach, anyway) to develop a
equivalent to any other, so the exact knowledge of the true micro- nonlocal representation that is equivalent to that proposed
scale structure is technically not necessary. This is an issue that previously by Deng et al. [21]. These scaling laws were: (i)
should be explored in additional detail in further studies. influence of boundaries, (ii) statistical homogeneity, (iii)
separation of length scales, (iv) ergodicity, and (v) smallness
7. Conclusions of the variance (or coefficient of variation). In particular, the
separation of length scales and ergodicity requirements
In the same paper from which the quote in the introduction was could be related to length-scale constraints of the form
drawn, Rosenblueth and Wiener [1] noted ‘‘The best model of a cat
‘  L; ‘  r0 : ð120Þ
is another cat, and preferably the same cat”. Although this was
meant to be a somewhat lighthearted comment in an otherwise Note that this is not intended to imply that the coarse-
philosophically challenging paper (and serves to continue the graining of all operators are subject to these constraints;
physicists’ apparent obsession with metaphorical cats [29]), it also these restrictions are required only if one wants a closed non-
contains an important comment about the modeling process itself. local form for the upscaling of a convection–dispersion
All models of real systems (except for the exceptional case where operator.
the model is the system itself) represent abstractions of one sort (4) Interestingly, the primary difference between the nonlocal
or another, and this is especially true of all mathematical models and local models developed in this work is the assumption
of real systems. of one scaling law. This assumption requires that the size
When one has very robust scaling laws, such as is the case for of the averaging volume be small compared with the large
many statistical mechanical systems like the ideal gas at near length scale, L. Coupling this with the constraints given by
atmospheric pressure and temperature, then there is hope for gen- the inequalities in (120), a local model was applicable under
erating a coarse-grained mathematical model that is free from the length-scale constraints
large amounts of uncertainty. For these kinds of systems, scaling
‘  r 0  L: ð121Þ
laws such as statistical homogeneity and ergodicity are so over-
whelmingly likely to be true that deviations from these laws are (5) An upscaled mathematical model that has the ability to rep-
essentially never observed. Using the feline-oriented terminology resent a complex system usually does so at the expense of
of Rosenblueth and Weiner [1], the argument would be that any also requiring more parameterization and characterization
cat is in fact as good model for another, because they will all yield in order to get the model to work. An example comparing
the same observable result with a probability approaching unity. the number of degrees of freedom required to represent a
Hydrological models are of a very different character than the macroscale nonlocal and local model, respectively was exam-
intrinsically scalable models alluded to above. The number of de- ined. Although the nonlocal model had fewer assumptions
grees of freedom in subsurface hydrologic systems (measured, (scaling laws) associated with it, it also required more infor-
perhaps, by something like the number of integral scales con- mation. An example of a finite difference approximation to
tained in a representative volume) is generally much smaller than these two macroscale models indicated that for a system
for the robust systems mentioned previously. The terminology with, say, N D computing nodes and using 27 point template,
quasi-scalable has been suggested for these systems, because it the ratio of the number of degrees of freedom of the nonlocal
is not usually true that the scaling laws are overwhelmingly likely to local mode was on the order of N D =27. As N D becomes
to be inviolate. Although upscaling can still be conducted for such large (say, millions of compute nodes), then the difference
systems, it is extremely important that the scaling laws and other in the number of degrees of freedom between the two mod-
assumptions that are invoked be made as explicit as is possible. It els can become quite large. This does not imply that nonlocal
is also important to recognize that the results of such models are models are always less efficient than directly solving the
possibly subject to substantial uncertainty. These issues have microscale equations, only that there may be relevant cases
been identified previously by other researchers (e.g. [30–32]), where an upscaled model yields a solution that actually
but the identification of scaling laws imposed in various analyses increases the number of degrees of freedom required (when
of heterogeneous subsurface systems is still usually murky at compared to the direct microscale problem).
best.
The primary results of this paper can be summarized as follows: It seems appropriate here to discuss the relative benefits of pro-
moting particular microscale and macroscale models for represent-
(1) The ability to develop a coarse-grained equation for describ- ing transport in heterogeneous porous media. No particular
ing transport in heterogeneous porous media has much less representation of a physical process (e.g. normal, log-normal,
to do with the particulars of the mathematical methods used power law distributed, gamma distributed) a priori is intrinsically
736 B.D. Wood / Advances in Water Resources 32 (2009) 723–736

better than another. Promoting a particular model without first [12] Wood B, Cherblanc F, Quintard M, Whitaker S. Volume averaging for
determining the effective dispersion tensor: Closure using periodic unit cells.
carefully addressing what question has been asked (or, equiva-
Water Resour Res 2003;39(8):1210. doi:10.1029/2002WR001723.
lently, what the information needs from the model are) is some- [13] Christakos G. Random field models in earth sciences. San Diego: Academic
what nonsensical. This is consistent with an evolving body of Press, Inc.; 1992.
work that suggests that given a choice among models, all other things [14] Beran M. Statistical continuum theories. Monographs in statistical physics and
thermodynamics, vol. 9. New York: Wiley; 1968.
being equivalent, the simpler model is a better choice. This statement [15] Bachman G, Narici L, Beckenstein E. Fourier and wavelet analysis. New
has been known for years in science as Ockham’s Razor, but there is York: Springer; 2000.
now recognition that this concept can be developed from a more [16] Churchill RV. Operational mathematics. 3rd ed. New York: McGraw-Hill; 1972.
[17] Attinger S. Generalized coarse graining procedures for flow in porous media.
fundamental assessment of the information content of the various Comput Geosci 2003;7:253–73.
alternatives (cf. [33,34]). It seems likely that research on coupling [18] Dagan G. Flow and transport in porous formations. Berlin: Springer-Verlag;
complex hydrologic models with the concepts from information 1989. The über omnitome of the Dagan approaches. Contains some derivation
of the SC approach [chapter 5].
theory and utility theory may yield some useful results that could [19] Whitaker. Levels of simplification: the use of assumptions, restrictions, and
help to establish more formally establish what constitutes a good constraints in engineering analysis. Chem Eng Edu 1988;22:104–8.
model for a particular complex hydrologic application. [20] Gelhar L, Axness C. Three-dimensional stochastic analysis of macrodispersion
in aquifers. Water Resour Res 1983;19:161–80.
[21] Deng F-W, Cushman JH, Delleur J. A fast Fourier transform stochastic analysis
Acknowledgements of the contaminant transport process. Water Resour Res 1993;29:3241–7.
[22] Deng F-W, Cushman JH. Comparison of moments for classical-, quasi-, and
convolution-Fickian dispersion of a conservative tracer. Water Resour Res
This research was supported in part by the National Science 1995;31:1147–9.
Foundation, Collaboration in Mathematical Geosciences (CMG) [23] Morales Casique E, Neuman S, Guadagnini A. Nonlocal and localized analyses
of nonreactive solute transport in bounded randomly heterogeneous porous
Program under Grant 0724865-EAR. I thank Francisco Valdes-Para-
media: theoretical framework. Adv Water Resour 2006;29:1238–12553.
da and three anonymous reviewers for their helpful comments on [24] Quintard M, Whitaker S. Transport in ordered and disordered porous media: II.
this manuscript. Generalized volume averaging. Transp Porous Media 1994;14:179–206.
[25] Shannon C, Weaver W. The mathematical theory of communi-
cation. Urbana: The University of Illinois Press; 1949.
References [26] Wiener N. Entropy and information. Proceedings of the symposia in applied
mathematics, vol. 2. Providence (RI): American Mathematical Society; 1950. p.
[1] Rosenblueth A, Wiener N. The role of models in science. Philos Sci 305.
1945;12(4):316–21. [27] Gel’fand I, Yaglom A. Computation of the amount of information about a
[2] Garrod C. Statistical mechanics and thermodynamics. New York: Oxford; 1995. stochastic function contained in another such function. Uspekhi
[3] Feynman R, Leighton R, Sands M. The Feynman lectures on physics, vol. Matematicheskikh Nauk 1957;12(1):3–52 [[in Russian]; A translation
I. Redwood City (CA): Addison-Wesley Publishing; 1963. appears in American Mathematical Society Translations, Series 2, 12:199–
[4] Cherblanc F, Ahmadi A, Quintard M. Two-medium description of dispersion in 246].
heterogeneous porous media: calculation of macroscopic properties. Water [28] Scheibe TD. Characterization of the spatial structuring of natural porous media
Resour Res 2003;39(6):1154. doi:10.1029/2002WR001559. and its impacts on subsurface flow and transport. PhD, Stanford; 1993.
[5] Polyanin A. Handbook of linear partial differential equations for engineers and [29] Schrödinger E. Die gegenwartige situation in der quantenmechanik [The
scientists. Chapman and Hall/CRC; 2002. present situation in quantum mechanics]. Naturwissenschaftern
[6] Gray W, Leijnse A, Kolar R, Blain C. Mathematical tools for changing spatial 1935;23:807–12. 823, 844–9 [English Translation: John D. Trimmer,
scales in the analysis of physical systems. Boca Raton (FL): CRC Press; 1993. Proceedings of the American philosophical society, vol. 124; 1980. p. 323–38].
[7] Baveye P, Sposito G. The operational significance of the continuum hypothesis [30] Dagan G. Transport in heterogeneous porous formations: spatial moments,
in the theory of water movement through soils and aquifers. Water Resour Res ergodicity, and effective dispersion. Water Resour Res 1990;26(6):1281–90.
1984;20:521–30. [31] Fiori A. Concentration fluctuations in transport by groundwater: comparison
[8] Whitaker S. The method of volume averaging, theory and applications of between theory and field experiments. Water Resour Res 1999;35(1):105–12.
transport in porous media. Dordrecht: Kluwer; 1999. [32] Fiori A. On the influence of pore-scale dispersion in nonergodic transport in
[9] Wen X, Chen Y, Durlofsky L. Efficient three-dimensional implementation of heterogeneous formations. Transp Porous Media 1998;30:57–73.
local–global upscaling for reservoir simulation. SPE J 2006;11:443–53. [33] Frieden B. Science from Fisher information. Cambridge: Cambridge University
[10] Popper K. The logic of scientific discovery. New York: Basic Books; 1959. Press; 2004.
[11] Russell B. The problems of philosophy. Home University Library of modern [34] Jaynes E. Probability theory: the logic of science. Cambridge: Cambridge
knowledge, vol. 35. New York: Holt and Company; 1912. University Press; 2003.

You might also like