PB87-141479
     STOCHASTIC PREDICTION OF  DISPERSIVE CONTAMINANT
     TRANSPORT
     Department of Civil Engineering
     Cambridge, MA
     Dec  86
EJBD      ^^=^^^E^^^^=^==E=^^^^^^^=
ARCHIVE
EPA
600-
2-

114                     U.S. DEPARTMENT OF COMMERCE
                     National Technical Information Service

-------
                                                                           P887-141479
                                                                    EPA/600/2-86/114
                                                                    December 1986
                   STOCHASTIC PREDICTION OF DISPERSIVE  CONTAMINANT TRANSPORT
                                              by

                           Efstratios G. Vomvoris and  Lynn  W.  Gelhar
                             Massachusetts Institute of  Technology
                                Cambridge, Massachusetts 02139
 f*
 3"
—                                   Cooperative Agreement
4                                       CR-81 1135-01-2
                                        Project  Officer

                                        Joseph F.  Keely
                              Applications and Assistance Branch
                       Robert S. Kerr Environmental  Research Laboratory
                                      Ada, Oklahoma  74820
                        ROBERT S.  KERR ENVIRONMENTAL RESEARCH LABORATORY
                              OFFICE OF RESEARCH AND DEVELOPMENT
                             U.S.  ENVIRONMENTAL PROTECTION AGENCY
                                         ADA, OK 74820
                                REPRODUCED BY
                                    U S. DEPARTMENT OF COMMERCE
                                           NATIONAL TECHNICAL
                                          INFORMATION SERVICE
                                          SPRINGFIELD. VA 22161

-------
                                  TECHNICAL REPORT DATA
                           (Please read Instructions on the reverse before completing/
 REPORT NO.
  EPA/600/2-86/114
            3 RECIPIENT'S ACCESSION NO, 1-
                  -'    *  --   *• '
 TITLE AND SUBTITLE
 STOCHASTIC PREDICTION OF DISPERSIVE CONTAMINANT
 TRANSPORT
            5 REPORT DATE
                December 1986
            6. PERFORMING ORGANIZATION CODE
 AUTHORIS)

 Efstratios G.  Vomvoris and Lynn W. Gelhar
                                                          8 PERFORMING ORGANIZATION REPORT NO
 PERFORMING ORGANIZATION NAME AND ADDRESS
 Massachusetts  Institute of Technology
 Department  of  Civil  Engineering
 Cambridge,  Massachusetts  02139
                                                           10 PROGRAM ELEMENT NO
                     CBPC1A
             11 CONTRACT/GRANT NO
                   CR-811135
2 SPONSORING AGENCY NAME AND ADDRESS
 Robert S. Kerr Environmental Research Lab. - Ada,  OK
 Office of Research and Development
 U.S.  Environmental  Protection Agency
 Ada,  Oklahoma   74820	
             13. TYPE OF REPORT AND PERIOD COVERED
              Final  (09/01/83 - 04/30/86)
             14 SPONSORING AGENCY CODE
                   EPA/600/15
5 SUPPLEMENTARY NOTES
6 ABSTRACT
       The  objective of this research  agreement was to develop mathematical  models
  to  quantify the concentration variability observed in field measurements of
  concentration plumes.  The concentration variability is attributed  mainly to
  spatial  heterogeneity of the hydraulic  conductivity field.  Since there is limited
  information about actual distributions  of hydraulic conductivity  in a given site,
  the log-hydraulic conductivity  is modeled as a three-dimensional  anisotropic
  stationary random process.  It  is shown that the concentration  variance, which
  measures  the intensity of the variations, is proportional to the mean concentra-
  tion gradient and to the variance and correlation scales of the log-hydraulic
  conductivity; it is inversely proportional to the local dispersivity values.  The
  main contribution to the concentration  variability results from the coefficients
  corresponding to the longitudinal mean  concentration gradients.  Analysis of a
  portion of the data from the Canadian Forces Base, Borden, Ontario, shows the
  applicability of the developed  results  in field situations.  Simple analytical
  examples  demonstrate the way to use  the results in a predictive context.
                               KEY WORDS AND DOCUMENT ANALYSIS
                  DESCRIPTORS
                                              b.lDENTIFIERS/OPEN ENDED TERMS
                                                                           COSATi field/Croup
18. DISTRIBUTION STATEMENT

   RELEASE TO PUBLIC.
19 SECURITY CLASS iTIut Reporti
   UNCLASSIFIED	
                                                                         21 NO 01
104
                                              20 SECURITY CLASS (Tinspage/
                           22 PRICE
                                                 UNCLASSIFIED
EPA Form 2220-1 (R.». 4-77)   PREVIOUS EDITION is OBSOLETE

-------
                        DISCLAIMER
 "The information in this document has been funded wholly
 or in part by the United States Environmental Protection
 Agency under Cooperative Agreement CR-811135-01-2 to the
 Massachusetts Institute of Technology.  It has been subjected
- to the Agency's peer and administrative review, and it has
 been approved for publication as an EPA document."
 Mention of trade names or commercial products does not constitute
 endorsement or recommendation for use.
                             ii

-------
                                  FOREWORD
     EPA is charged by Congress to protect the Nation's land, air and water
systems.  Under a mandate of national environmental laws focused on air and
water quality, solid waste management and the control of toxic substances,
pesticides, noise and radiation, the Agency strives to formulate and imple-
ment actions which lead to a compatible balance between human activities and
the ability of natural systems to support and nurture life.

     The Robert S. Kerr Environmental Research Laboratory is the Agency's
center of expertise for investigation of the soil and subsurface environment.
Personnel at the Laboratory are responsible for management of research pro-
grams to:  (a) determine the fate, transport and transformation rates of
pollutants'in the soil, the unsaturated and the saturated zones of the
subsurface environment; (b) define the processes to be used in character-
izing the soil and subsurface environment as a receptor of pollutants; (c)
develop techniques for predicting the effect of pollutants on ground water,
soil, and indigenous organisms; and (d) define and demonstrate the applica-
bility and limitations of using natural processes, indigenous to the soil
and subsurface environment, for the protection of this resource.

     This report contributes to that knowledge which is essential in order
for EPA to establish and enforce pollution control standards, predict the
extent of existing contamination, and to develop ways to restore the
environment, when contaminated, in a reasonable and cost-effective manner.
                                       Clinton W. Hall
                                       Director
                                       Robert S. Kerr Environmental
                                         Research Laboratory
                                     ill

-------
                                    CONTENTS
Abstract	    vi
Figures	   vii
Tables	    ix
Symbols 	     x

Section
    1.  Introduction  	     1
    2.  Conclusions 	     3
    3.  Recommendations 	     4
    4.  Natural Variability and the Stochastic Approach  	     5
    5.  Analysis of Concentration Variability 	    15
    6.  Application Examples  	    38
    7.  Field Data Analysis:  The CFB Borden Site	    47
    8.  Discussion of Results	    67

References	    °^
Appendices
    A.  The Mathematics of the Approximate Analytical Evaluations  ...    74
    B.  The Integration Over Wave-Number	    84
    C.  Transformation to Polar Coordinates 	    88
    D.  Mean Concentrations Fitted at the Borden Site Data	    91
            Preceding page blank

-------
                                   ABSTRACT
    The objective of this research agreement was  to develop mathematical
models in order to quantify the concentration variability observed  in  field
measurements of concentration plumes.

    The concentration variability is attributed mainly  to the  spatial
heterogeneity of the hydraulic conductivity field.  Since there  is  limited
information about the actual distribution of the  hydraulic conductivity  in a
given site, the log-hydraulic conductivity is modeled as a three-dimensional
anisotropic stationary random process.  It is characterized by its  mean  and
its covariance function or spectrum.  Implementing Darcy's equation and  the
convective-dispersive equation at the local scale, relationships  among the
concentration variations and characteristic parameters  of the  porous medium
are found.  Approximate analytical expressions are also developed.

    It is shown that the concentration variance,  which  measures  the intensity
of the variations, is proportional to the mean concentration gradient  and  to
the variance and'correlation-scales of the log-hydraulic conductivity; it. is
inversely proportional., to the local dispersivity  values.  The  main
contribution to the c6ncentration variability results from the coefficients
corresponding to the longitudinal mean concentration gradients.

    There is a very high correlation in the concentration values  along the
direction of the flow, which is still under investigation.  The  orientation
of the hydraulic conductivity principal axes does not have a significant
effect on the concentration variance values.

    Analysis of a portion of the data from the CFB Borden site shows the
applicability of the developed results in field situations.  Simple analytical
examples demonstrate the way to use the results in a predictive  context.

    This report was submitted in fulfillment of Cooperative Agreement  No.
CR-811135-01-2 by the Massachusetts Institute of  Technology under the
sponsorship of the U.S. Environmental Protection  Agency.  This report  covers
a period from September 1, 1983 to April 30, 1986, and  work was  completed as
of April 8, 1986.
                                       vi

-------
                                     FIGURES

                                                                         Page

          (a) Hydraulic conductivity data from laboratory cores
              from a borehole in the Mt. Simon aquifer Illinois
              (Bakr, 1976) 	    6
          (b) Hydraulic conductivity profile from a core in the
              CFB Borden site, Canada (Sudicky, 1985b) 	    6

4.2           Concentration measurements from the Borden site.
              (a) Depth averaged data (b) Vertical profiles
              (Freyberg et al. 1983 and personal communication)   ....    8

4.3           Schematic graph of a stationary one dimensional
              covariance function of fcnK showing the correlation
              scale \	10

4.4           Schematic graph and comparison of the measured,
              predicted with field scale dispersion and predicted
              with laboratory scale dispersion concentration profiles   .   10

5.1           Coordinate systems for the case X2 = X2*; the
              dashed elipse represents schematically the e"1
              level of the correlation function	13

5.2           Comparison of sample autocorrelation function (crosses)
              for hydraulic conductivity (from grain-size analysis)
              with theoretical autocorrelation function (solid  line)
              (From the CRNL, Ontario experiment, Moltyaner, 1985)  ...   24

5.3           Comparison of Che exponential correlation function  and
              the "hole-type" correlation function 	   25

5.4           Comparison of the variance for an anisotropic aquifer
              (correlation scales l^, 12* ^ and an equivalent
              isotropic (correlation scale i. = (41^2^3)   )	29

5.5           Comparison of the approximate analytical expressions
              and the numerical results for the TJJ coefficients  ....   30
              (a) TH(.) along the x direction for different e
              (b) T22(«) along the x direction
              (c) T22(«) along the y direction

5.6           Test of the validity of the approximate result for
              the variance TH (1, 1, e, 1;0) with respect  to  the
              e values	32

6.1           Normalized mean concentration and standard deviation  at
              x = 50 ra, z = 0 ra	39

6.2           Isoconcentration lines and standard deviations at  z = 0  m   41


                                      vii

-------
                                                                         Page

           Correlation function for a point at y = z = 0 m	42
           (a) Correlation along x
           (b) Correlation along lines in the x-y plane at
               various angles with x
           (c) Lag-correlation contours in the x-y plane

6.4        Normalized concentration and one standard deviation
           interval along a longitudinal section (line coordinates
           y = z = 0)	44

6.5        Normalized concentration and one standard deviation
           interval along a vertical section (line coordinates
           x = xc - center of mass of plume, y = 0)	45
6.6       • Normalized concentration and two standard deviation
           interval along another vertical section (line
           coordinates x = xc - 5 m, z = 1 ra)	46

7.1        Instrumentation, location of cores and hydraulic head
           distribution on August 10, 1984 (solid line) and November
           15, 1984 ( 	 ) at the tracer-test site (after Sudicky
           1985b).  The circled samplers show the locations of  the
           wells used in this analysis.  V denotes the well used for
           the vertical section analysis 	  48

7.2        Isoconcentration contours at depth z = -5.31 ra from  the
           Borden site data (original data from 0. Freyberg, personal
           communication)  	  49

7.3        Schematic presentation of the methods used  for the
           estimation of the mean concentrations	58

7.4        The measured, fitted and plus-minus one standard deviation
           concentration profiles along the longitudinal section at
           z = -5.31 m	59

7.5        The mean concentration curves fitted at the wells on the
           longitudinal section (c in the vertical vs. z)	60

7.6        The measured, fitted and plus-minus one standard deviation
           concentration profiles along the vertical section at (x,y)
           =  (38.3,  16.08) 	  64

7.7        The values plotted in Fig. 7.6 shown in a semi-logarthmic
           plot	65

C.I        Schematic graph of the coordinate transformation geometry  .  .  90
                                     viii

-------
                                   TABLES
                                                                         Page

              Approximate analytical results for the
              concentration covariance 	  27

5.2           Effect of the stratification orientation on  the
              concentration variance 	  34

7.1           Existing Information on the Borden site parameters  ....  48

7.2           Effect of the input parameters on effective
              hydraulic conductivity 	  53

7.3           Effect of the input parameters on the
              macrodispersiviti.es	55

7.4           Parameters used in the variance estimation	56

7.5           Concentration gradients in the vertical section
              at (x,y) = (38.3m., 16.08m.) 	  56

7.6           Estimation of the range of variations  in the
              longitudinal section	»	62

7.7           Estimation of the range of variations  in the vertical
              section	"	63

D.I           Parameters used to fit mean concentrations at the wells   .  91
                                       ix

-------
                                LIST OF SYMBOLS




A               auxiliary variable defined in Eq. (5.27)

All» A22» A33   macrodispersivities along axes x, y, and z

B               auxiliary variable defined in Eq. (5.27)

c               concentration of the transported solute (volumetric)

dZfOO          Fourier-Stieltjes wave amplitude of the random process  f(x_)


E[ • ]            expectation operator

E..             local bulk dispersion coefficient tensor


E               coefficient defined in Eq. (A.I)

F               mean of log-hydraulic conductivity

f(x)            perturbation of log-hydraulic conductivity  field
                and generic symbol for a random  field  in Section 4.
G               mean concentration gradient along x.; G. =        .
 j                                                 J   J          J

K..             coefficients used in the evaluation of K. . ;
 1J             defined in Eq. (7.11)                    1J

J               hydraulic gradient in direction  x.


K               hydraulic conductivity

1C               defined from exp[E(lnK)j


1C               effective hydraulic conductivities

K(t)            auxiliary variable defined  in  Eq. (5.16)

k^               the three-dimensional wave  number vector;  (kj ,  k2 ,  k3)

i               characteristic correlation  scales of  the  hole  type
 i              exponential covariance  function

-------
m(x)            mean of a  random  field


m               rate of mass  injection


M               total mass  injected


n               porosity


q               specific discharge in the  x   direction


R_,(x,, x_)     covariance  function  of  the random field  (f(x_)


Sf,(2c. , £9)     spectrum of the random  field  (f(:c)


T           "    nondimensional coefficients used  in  the  evaluation of
 JP             the concentration covariance;  defined  in Eq.  (5.33)

 i
T.              dimensional coefficients used  in  the evaluation of
 JP             the concentration covariance;  defined  in Eq.  (5.33)


V               mean pore  velocity


£               (xj, x2, x3)  = (x, y, z)   coordinate vector and
                coordinate  axes in the  three-dimensional space


x               location of center of mass of  a plume


cu , cu,          longitudinal  and  transverse local dispersivities


6               coefficient defined  in  Eq. (A.I)


Y               flow factor = q/K J.


f^il' 2i2^       semi-variogram of a  random process


 ij             Kronecker delta, equals 1  if  i =  j


e               defined from a,/i^


z               defined in  Eq. (A.I)


6               auxiliary variable defined in  Eq. (5.16)


K               ratio of transverse  to  longitudinal  local dispersivity


X               correlation scale of the negative exponential
                covariance function
                                        xi

-------
PI , p2          aspect ratios of the covariance of the £nK
                defined in Eq. (5.33c)


p  (xj, Xj)     correlation function of the random process g(x)
 SB

 2
o               the variance of a random field g(x)
 g                                               ~


1(1               rotation angle of the concentration plume with
                respect to the mean flow direction


                (overbar) mean of a random field


'               (prime) perturbation of a random field
                                       xil

-------
                                   SECTION  1

                                 INTRODUCTION
    Field observations of solute transport  in groundwater  indicate  that
dispersivities are scale dependent and can  range over many  orders of
magnitude.  This discrepancy between  the  laboratory  and  field  experience has
been attributed mainly to the spatial variation of the flow properties of the
natural porous earth materials.  Therefore, over the last  decade  a  substantial
portion of the research in the area of subsurface hydrology has  focused  on
understanding the effects of the spatial  heterogeneity of  natural formations
on flow and -solute transport.

    Continuously expanding computer capabilities have facilitated the use of
deterministic numerical models with parameters which may be spatially variable
over a grid and have, in principle, made  it possible to  obtain solutions that
previously could be described only qualitatively.  However, such  solutions are
limited to the specific application and are unable to generate relationships
valid over a broad range of problems.  More importantly, in order for a
numerical model to reproduce the actual detail of a  given  field situation, it
must have sufficient grid resolution  to represent the velocity field  producing
the mass transport and information about  the controlling parameters,  such as
hydraulic conductivity and porosity,  at each grid point.   While  the former may
be practical for small idealized flow systems (say over  tens of meters), the
number of'nodes required to resolve the actual three-dimensional  heterogeneity
at a more practical field scale (say  1 km)  can easily exceed the  capacity of
even the largest supercomputers.  Furthermore, the actual  measurement of the
required parameters with that degree  of resolution is a  totally impractical
task.

    It is the lack of information about the spatial  distribution  of the
controlling parameters that has led to modeling them as  realizations  of
random fields.  A random field can be viewed as a generalization  of the  random
variable concept.  Before measuring the value of a random  variable  the only
information known is a range for its  possible values (defined  by  its  mean and
variance).  In a random field it is further assumed  that there is a
correlation in space among the different  values, so  that knowledge  of the
value at one location carries information about the  possible values for  the
hydraulic conductivity at different locations in such a  way that  they fall
into the prespecified ranges (defined by  the mean and the  variance) and  have
the appropriate correlation in space  (defined by the correlation  function).
This approach is coined the stochastic approach.  The fundamental assumption
is that for each one realization the  classical laws  of flow and mass  transport
hold at some representative scale.

-------
    In Che stochastic approach, the mean  or  expected  value  behavior of the
concentration field is found to be equivalent  to  the  classical  dispersive
transport equation.  The  theory also  predicts  the  resulting field-scale
dispersion coefficients in terms of the statistical characteristics of the
hydraulic conductivity variability.   However,  the  mean  concentration
represents the average behavior of the ensemble rather  than the actual
realization of the aquifer.  Therefore, a classical transport model will
simulate a smoothed approximation of  an actual field  simulation;  the actual
concentration will vary around the smooth mean.   In order to realistically
present the results of model simulations, it is necessary to understand and
quantify this error in classical models.

    The stochastic approach can be used to quantify the  variation around the
mean concentration.  The main focus of this  research  is  the development of the
theory to describe such variability.  The degree  of variability will be
characterized by the concentration variance.   It will be shown  that the
concentration variance depends on measurable or determinable quantities
characterizing the aquifer material,  such as the  correlation scales and the
variance of the log-hydraulic conductivity,  the inclination of  the stratifi-
cation, and the local dispersivity values.   It also depends on  the mean
concentration gradient which, through the macrodispersivities,  represents the
overall effect of the aquifer heterogeneity on the mass  transport.  Section 6
shows how the mean concentration gradient can  be  estimated  in a predictive
context with numerical or analytical  models.   Section 7  demonstrates how it
can be estimated through a simple curve fitting with  existing plume
measurements, by the analysis of the  field data from  the Borden site in
Canada.

    The stochastic approach is complementary to classical deterministic
modeling.  The predicted concentration variance can be used as  a  realistic
and physically-based calibration target for  the large-scale numerical models.
Furthermore, one can use any classical numerical or analytical  model that
reproduces the general characteristics of the  contamination event  and,  in
addition, obtain a measure of the expected variability around the  model
results.   Finally, knowledge of the possible deviations  of  the  actual plume
from the  predicted one can improve substantially  the design of  monitoring
schemes.

    The present study is focusing on  quantifying  the variation  of  the
concentration measurements or model predictions,  in heterogeneous  porous
formations.  Examples demonstrate the practical application of  these results.

-------
                                    SECTION  2

                                   CONCLUSIONS

    The concentration variability resulting  from  the  heterogeneity  of  the
natural aquifers can be quantified using stochastic analysis  in  terms  of
measurable or determinable properties of the hydraulic  conductivity field.   It
can be expressed mathematically as the sum of the products  of  the concentration
gradients weighted by appropriate coefficients, which depend  on  the variance and
the correlation scales of the log-hydraulic  conductivity, the  local dispersivity
values and the orientation of the stratification.  The  dominant  coefficient
corresponds to the gradient along the mean flow and is  inversely proportional to
the local dispersivity and proportional to the correlation  scales.

    The effects of the orientation of the stratification  in a  preliminary
analysis appear to be less important, but further investigation  is  required  for
definitive conclusions.  The structure of the concentration covariance for  the
particular spectrum used results in a singular effect,  namely, high correlations
along the mean flow direction.

    Approximate analytical results have been developed  that establish  the role
of the key parameters.

    The preliminary analysis of field data from the Borden  site  shows  that  the
results of the proposed theory can be applied in  practical  situations  and
produce consistent confidence intervals for  the anticipated concentration
variability.

-------
                                    SECTION  3

                                 RECOMMENDATIONS

    A local relationship that quantifies the concentration  variability
anticipated in mass transport in heterogeneous aquifers has  been  found.   The
relationship involves the gradient of the mean concentration,  the correlation
scales of the log hydraulic conductivity field and  the local dispersivity
values.  As shown in Section 7, it can be readily used and  produce quite  tight
confidence intervals for the concentration values,  even with crude estimation
of the mean concentration field.

    The obtained concentration standard deviation can be  implemented  for  the
calibration of numerical models.  The discretization inherent  in  the  numerical
models induces a smoothing or averaging of the true concentration field and
makes it impossible to reproduce the measured values.  The  estimated  standard
deviation of the concentration can be used as a measure of  the acceptable
deviation of the numerical model results from the measured  values.

    The preliminary application of the results to a field site was very
encouraging.  Further analysis is needed to  explore the full three-dimensional
nature of the controlled plume and to follow its movement and  its response  to
the orientation of the stratification.  Important issues  such  as  spatial
averaging of the concentration field introduced by  the sampling technique and
its associated filtering properties need to  be addressed.

    The developed theory established the basic relationships among the aquifer
characteristics, the flow parameters and the concentration  variance.   However,
the effects of unsteady concentration input, initial conditions,  flow reversal,
and unknown history of injection need to be  investigated.   The singular behavior
of the concentration covariance along the mean flow direction  requires
additional investigation before accepted as  a valid descriptor of the physical
system.

    Finally the issue of the mean concentration field, central to the theme  of
all stochastic theories, should be investigated.  Methods to enhance  the  ability
to identify or estimate the mean concentration need to be developed.

-------
                                   SECTION 4

                NATURAL VARIABILITY AND THE STOCHASTIC APPROACH
THE THEORETICAL FRAMEWORK

    The conventional approach In groundwater modeling  is  Co  treat  an  aquifer
as an idealized homogeneous formation with a constant  value  for  each  physical
parameter so that the idealized aquifer behaves like the  actual  formation.   In
cases where there is predictable heterogeneity one might  divide  the aquifer
into a number of homogeneous zones, each with a different  equivalent  value  for
any used parameter.  The simple case of known anisotropy  can be  treated
incorporating suitable coordinate transformations.

    However,- the actual heterogeneity in a formation which would
conventionally be treated as homogeneous can be very pronounced.  Figure 4.1
illustrates this variability for the hydraulic conductivity  which  is
considered the most important flow property of an earth porous material.
Note that the quantity plotted is the logarithm of the hydraulic
conductivity.  The observed spatial variation is well  documented in the
literature; "... the jigs and jags of every electric log,  the spread  of
measurements from every sampling program, and common sense geological
reasoning attest to this fact." (Freeze, 1975).  The intensity of  the
variability of hydraulic conductivity can be characterized by the
estimated variance of the natural logarithm of the hydraulic conductivity
and which may be as high as 13 for some rocks (Freeze, 1975).

    In addition, due to the underlying depositional process, the hydraulic
conductivity field will typically exhibit some spatial structure.  Oftentimes
this structure manifests itself as layers or lenses and reflects directional
dependence or anisotropy.

    In a practical sense variability of the porous formation results  in
variations of the flow field and the head distributions around the ones
predicted for homogeneous systems.  It is further propagated to  the
distribution of the concentration since the latter is  directly affected by  the
flow field.  Figure 4.2 presents actual concentration  data from  the field
scale experiment at the Borden site.  Notice that intense  variation of the
measured values is observed for the vertical sections  and  an even  substantial
one for the depth averaged data, Figure 4.2a.

    The fundamental question that emerges is how can these effects be
modeled.  If it is assumed that a complete knowledge of the  hydraulic
conductivity process is available, the existing and continuously expanding
computer capabilities that permit the implementation of numerical  models with
parameters spatially variable across the grid, can produce the required
solutions.  However, at the current level of computing capabilities the
allowed resolution of the input field is not fine enough  to  capture the
anticipated effects.  Furthermore, such solutions will be  case specific and
unable to establish general relationships among the key parameters valid over
a broad range of problems.  But the "Achille's heel" of the  sophisticated
numerical modeling of the spatial heterogeneity is that for  all  practical

-------
                                       100
                                   DISTANCE, FEET
200
                   213.238
                   213.0-18 —
                   217.048
                        IU
                                                   IU
                                 10        IU

                      Hydraulic Conductivity  (cm/a)
Figure 4.1  (a) Hydraulic conductivity data  from laboratory cores  from
                a  borehole in the Mt. Simon  aquifer,  Illinois (Bakr,
                1976)
            (b) Hydraulic conductivity profile  from a core in the  CFB
                Borden  site,  Canada (Sudicky,  1985b)

-------
purposes it is impossible to acquire all  the detailed  data required to adequately
represent the actual variation of the parameters.

    The lack of information about the controlling  parameters,  in addition to the
need for analytical and general relationships,  has  led to modeling the
appropriate parameters of the porous medium as  realizations  of random fields.
The fundamental assumption is that  for each one realization,  and usually we have
only one in groundwater related problems,  the classical laws  of flow and mass
transport hold at some representative scale.

    The concept of a random field or spatial stochastic process is a
generalization of a random variable.  If £ represents  a point  in the three-
dimensional space and f(x) a random variable corresponding to  that point jt, the
set of random variables ff(Xi)» f (.£2 ) i ••• f(.2k)»  •••}  where 2^,3^, ... Xfc»
are the coordinate vectors of points 1,2,  ... k, is called a  random field and is
simply written as f(jc).

    As a random variable is completely characterized by its  probability density
function, for the complete probabilistic  characterization of  a random field f(j>0
one would need to determine the joint probability  density function of any set
fCxj), f(x2>, ... f(xji) for anv n and for  i » 22 •  •••  2Sn arbitrary points in
the spatial domain of interest.  In most  applications  the entire spatial law can
neither be identified from finite data nor is required; the  first two moments of
the distribution suffice to characterize  the random fields in  a number of
problems.  The first two moments are,

    - the mean function or first-order moment of the random field

                                E[f(x)] =  m (x)                 .       (4.1)

      where E[«] denotes mathematical expectation
    - the covariance function or second-order moment,
    The zero lag covariance, when x-  x~,  is  the  variance,  denoted as af2-
Another representation of a second-moment  is  the  variogram  (Journel and
Huijbregts, 1978) introduced in  the  field  of  geostatistics, and defined as,
    Note that Eq. (4.1) and Eq.  (4.2)  is  a  generalization of the concepts of mean
and variance of a random variable.   Eq. (4.2)  incorporates the information about
the spatial structure of the  process.   It defines  the  correlation between two
points of the field located at x and  x  ,  if  it  is  normalized by the
                        1      2
variance.

    The mean and the covariance  can  in general depend  on the location x, > as can
be seen in Eqs. (4.1) and  (4.2).  When the  mean and  covariance are independent
of jc, the process is said  to  be  (second order) stationary or statistically
homogeneous:

-------
      X

     (m)

      I !


      1C


       9


       6

       •7


       6


       5

       A


       S


       i


       1
 -6
DAY 85
          DAY 0
  9 IOO 200 "JOG 444 600
      C (mg/l)
              0  100  ISO 330  iOO ICO
                  C (mg/l)
                                                 y (m)
                                             -ll—
                                               r
                                               o  103
C (mg/l)
Figure 4.2    Concentration measurements  from the Borden  site,  (a)
              Depth  averaged data, (b)  Vertical profiles  (Freyberg et
              al,  1983  and personal communication)
                                8

-------
                       E{f(x)}  = m

                       Rff(V V = Rff(xl" V                       (4"4)

If the field is said Co be isotropic, the same correlation exists for points
that are separated by the same distance regardless of their orientation  in
space, i.e.,
                            RffQi^x^ = Rff(|-i~-2r                   (4"5)
Figure 4.3 illustrates graphically a one-dimensional stationary covariance
function.  One characteristic length scale is the separation distance at which
the correlation decreases to the e   level, designated as the  correlation
scale.  In the general stationary case the covariance can always be  expressed in
terms of the spectral density function or spectrum as follows,
where the spectral density function can be expressed as,


                       Sff 00 =  I  3 /// e ~^'L R
                                 (2ir)  -«
and k_ is the three-dimensional wave number vector.  The Fourier  transform pair
in Eqs. (4.6) and (4.7) implies that knowledge of only one  of  the  two,  spectrum
or covariance, is sufficient.  They both carry the same information,  but  it  is
displayed differently.  The covariance function has a physical meaning  that  is
easier to grasp:  how two actual values of the process under consideration,  a
certain distance apart "are expected to be correlated.  The  meaning of the
spectrum is more abstract:  if one could decompose a whole  set of  statistically
similar processes in trigonometric waves and make a histogram  of the  encountered
wave-length, that would be the spectrum.  It gives the distribution of  the
variance over the wave-number domain.

    If a process is statistically homogeneous, it will always  have a  spectrum,
and then it is possible to take advantage of the representation  theorem
[Luraley and Panofsky, (1962), Gelhar (1984)],
                                   CO
                            f(x) = /// e1^ dZf(k)


               E{dZf(k) dZ* (k')} = Sff(k)dk     , if k =  k1


                                  = 0            , if k_#  k_'            (4.8)

where dZ,(k) is  the Fourler-Stieltjes wave amplitude of  the process f(:c).
This theorem essentially  states that the process f(x)  is  composed of  a sum of
uncorrelated or  orthogonal increments in the wave number  domain.  Furthermore,
the representation  in Eq.  (4.8) is unique and  provides a  convenient formalism
for treating differential  equations involving  stationary  variables.


                                       9

-------
                                              s  x,-x2
Figure 4.3     Schematic graph of  a  stationary  one  dimensional
               covariance function of  £nK showing  the  correlation
               scale X
                                             fitted (laboratory scale
                                         **                dispersion)
                                                fitted (field  scale
                                                        dispersion)
 Figure 4.4     Schematic graph and comparison of the measured,
                predicted with field scale dispersion and  predicted
                with laboratory scale dispersion concentration profiles
                                 10

-------
    The spectral methodology will be adopted  In  the  current  work.   The_
characteristic random fields are decomposed in a known  mean  function,  g,  and a
zero-mean stationary perturbation, g1,

                            g(x) = g(x) + g'U)                          (4.9)


    Utilizing the representation theorem in the  differential equations
relating the various fields, one can relate the  spectrum of  the  dependent
variable to the spectrum of the input processes.

    Following the traditional deterministic approach, the aquifer  is modeled
as a continuum and the accepted expressions for  the  mass conservation, the
solute conservation and Darcy's equation are  assumed to hold at  the local
scale.

REVIEW OF RELATED WORK

Groundwater Flow

    Historically the problem treated initially in a  probabilistic  framework
was that of groundwater flow.  The approaches developed can  be  briefly
summarized into the following categories:

Monte Carlo Simulations—
    Realizations of sets of aquifer properties are generated through a model,
and the direct problem is deterministically solved for  each  realization.   From
the statistical analysis' of these solutions,  stochastic characteristics of the
head and flow fields can be inferred.  Freeze (1975) in his  pioneering paper
treated the problem of one-dimensional saturated flow in finite  domain where
the head at the boundaries was specified.  Spatial variation but not spatial
structure was the main limitation.  The latter was introduced later in the
work of Smith and Freeze (1979a, 1979b), where the nearest neighbor model
generator was used to generate one- and two-dimensional fields.

Perturbation or Linearization Models—
    This was the most widely used theoretical approach.  The stochastic
differential equations are solved using "small"  pertubations theory or Taylor
expansions in the space or frequency domain.  Gelhar (1984)  presents a good
review of the spectral methodology which has  been used  in a  series of  papers,
namely, Gelhar (1974, 1977), Bakr et al. (1978), Gutjar et al.  (1978), and
Mizell et al. (1980).  In the aforementioned  works head variances  and
covariances are obtained for the cases of one-,  two- and three-dimensional
flow in confined and unconflned aquifers.  Comparable results are  obtained in
Gutjahr and Gelhar (1981), Dagan (1982a), Vomvoris (1982), Kitanidis and
Vomvoris (1983), Hoeksema and ICitanidis (1984),  and  McLaughlin  (1985)  using
first order analysis in the space domain.  The advantage of  the  latter
approach is that it can be applied in bounded domains,  while the spectral
method is tailored for infinite stationary domains.   Note that  all the
results, regardless of the method used, show  that the one- and  two-dimensional
cases provide erroneously large variances, due to the unrealistically  limited
number of paths that are offered to the flow  system.


                                       11

-------
Composite Media Analysis—
    This approach was followed by  Dagan  (1979)  and  follows  closely the
self-consistent method described in  Beran  (1968).   The  hydraulic conductivity
field was represented as a collection of homogeneous  blocks with random radii
arranged at random locations.  A similar method,  that of  the embedding matrix
was presented in Chirlin and Dagan (1980).

Solute Transport in Groundwater

    In the late seventies, the probabilistic  framework  was  extended to include
the solute transport problem where the heterogeneity  of the natural formations
is believed to be more influential.  An  apparent  influence  is on the
dispersivity values corresponding  to field scale  plumes where an increase of
two or three orders of magnitude with respect  to  the  anticipated laboratory
measured values was observed.  A critical  analysis  of existing field
observations of solute transport has revealed  that  dispersivities are scale
dependent, Gelhar et al. (1985).   Furthermore,  concentration measurements
exhibit significant fluctuations around  the values  predicted from analytical
or numerical solutions.  The review  by Anderson (1979)  suggests  that the
dispersivities used in numerical models  are little  more than fitting
parameters and as such lack a sound  physical  basis.
    The central questions that the related research has focused  on are:
    - What are the governing equations for large-scale  mass transport in a
      heterogeneous aquifer?
    - How do the different scales  affect Che  characteristic parameters of what
      seems to be the same phenomenon?
A typical situation is depicted in Figure  4.A.  The two smooth curves
correspond to numerical or analytical solutions predicted with laboratory
scale dispersivities and. fitted field scale dispersivities.   The actual
measurements exhibit significant fluctuations  around  the  values  predicted by
the models.  The current work therefore  attempts  to answer  the following
question:
    - How can the concentration variability around  the  predicted solutions be
      quantified?

    The first attempt to show the  effects  of  spatial  heterogeneity was through
Monte Carlo simulations [Schwartz  (1977),  and  Smith and Schwartz (1980)].
From a numerical analysis point of view, the  solution of  the mass transport
equation is by far more difficult  than the solution of  the  flow  equation.
Instabilities in the algorithm are very  frequent  and  numerical dispersion of
the same order of magnitude as the physical dispersion  may  occur (Ababou
et al. 1985).  No quantitative conclusions were drawn from  the Monte Carlo
simulations.

    Gelhar et al (1979) quantified for the first  time the relationships
among field scale or macro-dispersivities  and  characteristic measures of the
hydraulic conductivity field.  The physical situation examined was that  of
macrodispersion in a stratified aquifer, and  the method used was spectral
analysis and Fourier-Stieltjes representation.  The authors developed the
governing equations near and far away from the  source,  demonstrated the
transition in the dispersivity until it  reached its asymptotic value and
showed the skewness of the concentration distribution near  the injection vs.


                                       12

-------
the symmetric Gaussian concentration distribution at points  far  downstream.
Some further aspects of the stratified case were discussed in Matheron  and
de Marsily (1980).  In a subsequent paper, Gelhar and Axness (1983)  treated
the general three-dimensional case of raacrodispersion in aquifers.   That  work
focused on the asymptotic dispersivity coefficient  from  the  simple case of
isotropic media to the most general one of statistically anisotropic media
with arbitrary orientation of the stratification.   The results produced
numerical values of the expected order of magnitude and compared favorably
with the scarce existing field data.

    Dagan (1984) produced similar results for  the raacrodispersivity  following
a Lagrangian approach which considers displacement  of particles  in random
velocity field generated from a random hydraulic conductivity field. This
analysis for the case of a statistically isotropic  medium yields results  for
the asymptotic dispersion coefficient which are equivalent to those  of  Gelhar
and Axness (1983).

    A detailed quantitative testing of the macrodispersivity prediction has
not been completed yet [Gelhar, 1985], but a number of carefully designed
field experiments are now underway [Freyberg et al. (1983),  LeBlanc  (1985),
and Betson et al. (1985)].  Hufschmied (1985)  has shown  that
macrodispersivities calculated from measurements of the heterogeneity of
hydraulic conductivity field are consistent with those estimated for a
cold-water plume in an aquifer in Switzerland.  Encouraging  comparisons are
also shown in Sudicky (1985b).

    The issue of concentration variation has not been yet widely recognized.
Gelhar et al. (1981) developed the concentration variance for the stratified
case in answering a comment on their 1979 paper.  They found that the
concentration variance is proportional to the  variance of the log hydraulic
conductivity, the square of the mean concentration  gradient, the fourth power
of the correlation scale of the hydraulic conductivity and inversely
proportional to the square of the local transverse  dispersivity. The
assumption of perfect layering is very restrictive  and thus  these results are
primarily of qualitative significance in demonstrating the factors which
influence the concentration variability.  These results do indicate  that
significant variations of concentration are to be expected across the layers
even after very large mean transport distances.

    Gelhar and Gutjahr (1982) have developed the equation for the
concentration variance for one-dimensional transport of  a solute undergoing
radioactive decay.  They have considered transport  in a  stream tube  with
varying cross-section, thus incorporating the  influence  of a three-dimensional
flow variation.  The results relate the concentration variance to statistical
parameters of the longitudinal velocity field.  Dependence of the
concentration variance on the square of the mean concentration gradient and
the square of the correlation scale of the velocity is observed. Due to  the
one-dimensional structure of the problem, analytical solutions for the
concentration perturbations can be obtained which could  be used  in a finite
domain.  The radioactive decay results in a finite  concentration variance even
when local mixing is eliminated by setting the local dispersivity value equal
to zero.

                                       13

-------
    Gutjahr et al. (1985) developed similar  results  following  a slightly
different methodology.  They treated the same  problem with  emphasis  on the
development of equivalent macrodispersivity  values.  Their  result  for the
concentration variance is still under investigation.

    Dagan (1982b, 1984) is the only other available  work  focusing  on the
calculation of the concentration variance.   His  approach  follows  closely the
theory of diffusion by continuous movements  [Taylor  (1921)]  referred to above
as the Lagrangian method.  In a qualitative  sense, Dagan's  results are similar
to the ones found in this research work.  One  of  the assumptions  made was that
the effects of local disperslvity can be neglected when calculating  the mean
motion of the plume.  It is thus believed that Dagan's results are applicable
in a region near the source of pollution where the convective  component of the
hydrodynamic dispersion is predominant.  In  contrast, the results  in this
study correspond to the far field downstream region.  From  the analyses of the
mean concentration, it is indicated that the local disperslvity does not have
an important effect on the macrodispersivity values.  The opposite effect is
observed when examining the concentration variance.  A final point is that the
approach of Dagan is performed for statistically  homogeneous,  isotropic media,
while the current work is based on the more  realistic case  of  anisotropic
hydraulic conductivity fields.
                                         14

-------
                                    SECTION 5

                      ANALYSIS OF CONCENTRATION VARIABILITY
THE GOVERNING EQUATIONS

    The general equation describing transport of an  ideal nonreactive
conservative solute in a saturated, rigid porous medium  is,  [Gelhar  and  Axness
(1983)]:

                            3° j. „  3c  - 1  v   l£                   (5.1)
                         n  IT+ qi^ Tx. Eij 3x.                   <5'1J

where,  n = porosity
       q. = specific discharge in  the *^ direction
        c = concentration of  the transported  solute
      E  .= local bulk dispersion coefficient  tensor.

The dispersivity values in  Eij will be  assumed  constant  at  the scale at
which Eq. (5.1) is valid.   Eq. (5.1)  incorporates the  conservation of  mass
equation for a homogeneous  fluid,

                                 1   q  = 0                            (5.2)
                                 3xt  Mi

    The  fields of  log hydraulic  conductivity,  specific discharge and
concentration will be decomposed into a mean  and a  small perturbation about
the mean.  Therefore,
                       F  + f '    (log  hydraulic conductivity)

             q   =  q.+  q'        i  = 1,  2,  3  (specific discharge)         (5.3)


             c  = "c + c1          (concentration)

 The  log  hydraulic conductivity perturbation field, f, is assumed to be a
 statistically  homogeneous random  field with zero mean and known covariance
 function.   Note that through the  Darcy equation, the specific discharge and
 the  hydraulic  conductivity are directly related, [Gelhar and Axness (1983)].
 The  concentration perturbation, c',  becomes a random field also and
 substituting Eq^_ (5.3) into Eq. (5.1)  the equations for the mean
 concentration, c, aiid  perturbation,  c1, can be found.  Without loss of
 generality, we can align the Xj axis along the mean flow, which renders


                              q -  q\  *  0 , ?2 = ?3 = 0                  (5.4)

     The  local  dispersion tensor may then be approximated in the form

                                        "aLq  0    0
                              [E
                                                   °Tq
                                       15

-------
where cu and OT are the local longitudinal  and  transverse
disperstvitiesT  Eq. (5.4) and Eq.  (5.5) can  simplify  the  equations
for the mean concentration and perturbation to:

               n |I  + q |I  + q. i£i = q OL  a!| +  q ttT  (i2! + i!|j    (5.6)
                                                1            23

                         —                  222
               3c' .   , 3c   .    3c'         3  c'        ,3 c'    3 c' ,,    ,s n
             n IT  + qi 3^  + q JZ  ~ q •L  —2  ' q  aT  (—2  + ~~2 J    (5<7)
                       , 3c'     , 3c'    -
                   = q! -2—  - q' -7—  =0


The zero mean term on  the right  hand  side  of  Eq.  (5.7)  is  of second order in
perturbations and is assumed to  be  negligible compared  to  the first order
terras on the left side.

    In Gelhar and Axness (1983)  Eq. (5.6)  and Eq.  (5.7)  are used to calculate
an expression for the  macrodispersivity tensor A  .   It  is shown that far
away from the source,  the mean  concentration  satisfies  the classical transport
equation with constant macrodispersivities.   In the  current work,  the
attention focuses on Eq. (5.7).  It will  be assumed  that the mean
concentration field is known.

    The effects of the velocity  field variations  on  the  local dispersion were
addressed in Naff (1978)'.   He retained the actual  velocity field in Eq.  (5.5)
and found that the effects  of the velocity variations on the macrodisper-
sivities were of second order.   An  extension  of his  analysis to study these
effects on the concentration variability  showed that they  were negligible in
this case also.  Consequently,  Eq.  (5.5)  is considered  a valid representation
of the local dispersion effects, at the current level of approximation.

SPECTRAL SOLUTIONS

    The equation for the concentration perturbation, Eq. (5.7), can be
easily solved if it is assumed  that the concentration perturbation field is
statistically homogeneous.   Invoking  the  Fourier-Stieltjes representation
theorem [Lumley and Panofsky (1962)]  one  can  find  relations between the
spectra of the concentration and log  hydraulic conductivity.  The  simple
case of a steady concentration  field  will  be  investigated  first.

Steady State Concentration  Field

    The perturbed quantities are represented  as,
                     00                       00
               c' =  / exp(ik.x)dZc(k); q^ =  / exp(ik.x)dZq (k)        (5.8)
                    —o»                     —oo             i

where k=(k  ,k  ,k.)  is  the wave  number vector, xfU^x^x.^) is the position

                                       16

-------
vector, and the integration is over three-dimensional wave  number  space.
Combining eq. (5.8) and eq. (5.7) and invoking the uniqueness of the  spectral
representation
                            n     o  o
                                             =G,dZ_                      (5.9)
where G.= - —=—, the mean concentration gradient,  is  taken  to  be  locally
constant.     j
                                                 *
    The complex conjugate Fourier amplitude of dZ  ft,  namely  dZ ,  can  be easily
found from Eq. (5.9).  Taking expectation of dZ d2  results  inCthe  following
spectral relationship:


                S  (k) = i  -=	s-J-i	,  0  ,                   (5.10)
where S  (_k) = the three-dimensional spectrum of  the  concentration
       cc      perturbation,
               = the three-dimensional spectrum of  the  specific  discharges
             along the axes x.and x  (j, i = 1,2,3).

    The final step is to establish the relationship between  Sq.q (k)  and
S,.(jOi the spectrum of the log-hydraulic conductivity  field.  JIn the
geneTal case, one expects aquifers to exhibit three-dimensional  anisotropic
heterogeneity with the mean flow arbitrarily oriented with respect  to the
stratification.  The coordinate systems might look  then as in  Figure  (5.1)  where
it can be observed that 'the mean hydraulic gradient vector does  not coincide
with the direction of the mean flow.  The covariance  function  has its principal
axes at the primed system.  Using the Darcy equation  for a locally  isotropic
medium, Gelhar and Axness (1983) showed that the  specific discharge spectra can
be expressed in terms of the log-conductivity spectrum  as
(5.1 la)
where j,£,m,n = 1,2,3.  (Note  that summation over m  and  n  is  implied.)
A simplified case corresponds  to statistically  isotropic £nK  field,  for
which the specific discharge spectrum becomes,
                  (k) - K  J J  (6  -
where y = 4/K Ji» and Ji ls the hydraulic gradient  in  the  x  direction.
    The covariance function of the concentration  field can be  computed  combining
Eq. (5.10) and Eq. (5.11),

                                       17

-------
                                                MEAN FLOW
                                                DIRECTION
Figure  5.1     Coordinate systems for Che  case *2 = X2'> cne

              dashed elipse  represents schematically  the e"1

              level of the correlation function
                          18

-------
where R  (£) is the covariance function of the concentration field,  at  lag
£. The variance is found by taking e = 0,
                                       00
                              Rcc(0) = / Scc(k)dk                      (5.13)
Unsteady Concentration Field
    The concentration perturbation field,  c'Ot.t),  is,  in general,  time
varying.  Again it will be assumed that the concentration field  is  locally
stationary (statistically homogeneous)  since the time-invariant  hydraulic
conductivity field is spatially homogeneous.  This  assumption is expected  to
be particularly valid away from the source.  Because the unsteady mean
concent rat io.n solution will be expressed in terms of a  moving coordinate
system, 5  ( g  = x  - qt/n, ?2 = X2' 53 =  X3^' 1C is aPPr°Priate Co express
the perturbation transport equation (Eq. (5.7))  in  that system as
          Jf 'I   +  "ill   =  v1     +  «^—  +-)      (5-7a>
          3t  'l      1 3Ci      L  8«J
where c'(£,t), i.e., _£ and c and are the independent variables.   The flow
field is steady in the fixed coordinate system so that  q'  in (5.7a)  is
represented by
                   q!(x)  =  f   e1-*- dZ  (k)                        (5.14a)
                    1 ~      i»          qi ~

but the time varying c'(^,t) will be represented as

                     CO
        c'(x.t)  =  /  ei-'-dZc (k;t,x.)
                 =  /   exp [i(k1?1 + qt/n) + k^ + k^\  dZc(k;t,ci)
                    — CO

                 =  c'(?i,t)                                          (5.14b)


Note that dZc is considered to vary slowly in space so that  the dependence
on gi is only parametric.  It should be mentioned parenthetically that the
introduction of the moving coordinate system is the more natural system  to
use, but, of course, from a mathematical point of view the two systems are
equivalent.  The moving system is introduced in order to take advantage  of the
fact that in the moving system, the concentration field will vary slowly in
time.
                                       19

-------
    Using Eqs. (5.11a)  and  (5.14b)  in (5.7a)  along with the uniqueness of the
representation theorem,

         n |   I  dZ  (_k;  t)- G.dZ  QO + iqk  dZ (k_;t)
           dt|_£C         J   °.j          1C

                            2       22
                 =   ~q  lar'ci  "*"  *]•'  2  "*"  3 ^J  "^c(_»^'                 (5.15)
Denoting,
                  dZ (k;  t)  =  p
                     C   q   , .,   A    .  2 _,_     ,.2 _,_ . 2 .,              ,,  .,.
                   Q ^ ™"  I T 1C   "T"  /v  iC   TrtCk.TiC.il              I T  L n J
                         n   I  1     L, 1     ^   2    3  '

                   K(t)   =   -  G.dZq.(k)


Eq. (5.15) can be simply  written as

      |£ I  +  0P  = K                                                   (5.17)

With initial condition  c'Ot.O)  = 0, or in the wave number domain dZ (k,0) = 0
= p(0), Eq. (5.17) can  be solved.   The solution is,

           -flt   C        Or
      p = e 0C /  K(T)  e9T  dT                                           (5.18)
                 o


      p=e"0CidZ   (k)  /  e9T G.(£,T) dT                             (5.19)
               "   qj --    o      J

    The integral in  this  last equation can not be evaluated analytically (G.'s
are complicated  functions which are unknown in the sense that they depend on  the
solution of the  concentration perturbation).   Because of the exponential term in
Eq. (5.19), the  main contribution will, for large 9, come from the region near
the upper limit, T = t.   Therefore, G.(£.;T) is expanded around the value at 7
= t. The expansion leads  to,
                              8Gi ,
       G.(£,T)   =  G.(C,T)  + r-1-   _r (x-t) + . . .                     (5.20)
        J            J         o T   I T L

where it is implied  that  the time derivative  is in the moving coordinate system
—  I  .  This expansion also has the  important advantage that it allows the
mean^transport equation to  be approximated in the form of the classical
advection-dispersion equation;  otherwise the  governing equation will be known in
the form of an integro-differential equation  wherein the macrodispersive flux is
dependent on the mean concentration gradient  at earlier times.  This feature  is
discussed by Gelhar  (1986).   Here because the goal is to develop results which
can make use of  mean solutions  based  on the classical transport equation, this
analytical approximation  will be explored in  detail.  Keeping only second-order
derivatives and  combining Eqs.  (5.19) and (5.20):
                                        20

-------
                q.   o                           =c

A change of variables,  u=t—r,  and the  limit  for  t •*•<»,  gives.
      p =  IdZq  /   e-0u [G.Or.t)  - ul |(5.0]  du                   (5.22)
Consequently, dZc(k^,t) has an approximate  analytical  expression.   From  this
point and on the procedure of the previous section  can  be  followed.   Integrating
Eq. (5.22) the complex amplitude is given  from,
      dZc(k,t)  -    dZq [G.^.t)     -j-      2  ]                      (5.23)
                        J                    59

with complex conjugate,


      dZ^ (k.t) - I dZ*_ [Cj(i.t)  \   -         „   ]                   (5.24)


The spectrum of c'(x,t)  is found by multiplying  the  complex  conjugate  amplitudes
and taking expectations.  Recall,

       E [dZ (k;t) dZ* (k;t)l  = Srr(k,t)  dk;
            c -      c  -        cc -     -                             (5>25)

       E [dZ  (k.) dZ* (k)] =  S    (k)d]c
            qj      qz         qjqa

This operation leads to  the following expression

                  i             i  i    "*Gi     i   l     3Go     l    l
      •«<*.<>  - i ^,,1 v, ?  ?   - i^ e. ? ? -  n1 BJ e- 1
                     ac   ac    Q2 Q*2


where summation over all the values of  j  and  i  is  implied.   To  simplify  notation
further denote
                                      21

-------
              .
           n  1

      B  =  J [aLk2 + aT(k2 + k2)]                  -                    (5.27)

      0  =  B + iA

      9* =  B - iA

     08* =  B2 + A2                                                      (5.28)

       222
      0* =  B  - A + 2iAB

      *2     2    2
     0 i =  B  - A - 21AB
Thus, Eq. (5.26) can be written as,


                  i2Sq,q.Trh2  1GJG*+S

                   3G.          ,         3G
                                            -    "                        <5-29'
'<" -'  "  " n2   V* A2 * B2  '"1".
The relationship between Sq.q  and Sff is presented  in  the  previous
section (Eq. 5.1 la)).  Notejtnat the inclusion of the time  variation  has
introduced three additional terras in the concentration  spectrum, as can be  seen
by the comparison oE Eq. (5.24) with Eq. (5.10).  This  implies  that the "zero"
order behavior of the unsteady case solution will be determined  from  the  steady
state.  The results will be different only in the region where  the concentration
gradients change rapidly, i.e., near the peak concentration.

THE LOG-HYDRAULIC CONDUCTIVITY SPECTRUM

    The intuitive notion for a covariance function structure dictates  that  it
should have its maximum at zero lag and decay, in some  way, as  the lag increases
dropping to zero at some distance.  A simple model that captures such  an  image
is a negative exponential.  It has been used extensively and appears  to be
consistent with the available data [(Bakr (1976) Sudicky (1985a)].  However
within the statistical limitation of field data, other  forms of  the covariance
can be equally acceptable.  Occasionally, the mathematical  structure of the
models used to describe the physical behavior require a stronger assumption
about the covariance behavior.  Bakr et al. (1978),  Gutjahr and  Gelhar (1981),
Vomvoris (1982) have used different covariance models,  which still show a decay
with distance but, in one sense, reinforce the periodic or  stationary  nature of
                                       22

-------
the modeled process.  The simplest of these  latter models  is  the  so-called "hole
type" covariance function (Bakr et al, 1978).   Its main  characteristic,  besides
the negative exponential decay, is the negative  correlation after a certain
lag.  In the wave number domain this characteristic  corresponds  to a zero value
of the spectrum at jt = 0, i.e., zero integral  scale.   Moltyaner  (1985,  p.
363-22, 23) has observed such negative correlations  in field  data (Figure 5.2).
In turbulence, where similar mathematical analysis has been used, a popular
spectrum is the following [Hinze  (1975), Aldama  (1985)].

      S(k) - k4exp (- ja2k2)                                           (5.30)


which also has zero integral scale.

    For the concentration covariance a similar zero  integral  scale log  hydraulic
conductivity covariance has to be used.  It  can  be seen  from  Eq.  (5.10)  that for
k2 = k3 =0, the kj in the denominator will  cause  the  integral to blow  up
at kj =0, when one integrates from -» to +».   This  behavior  can  be avoided if
Sq.q (0,0,0) = 0, or, from Eq. (5.11), Sff(0_)  =  0, which is the
zero-integral requirement postulated.

    A simple hole-type, three-dimensional covariance is  the following:
                                             2
                2 ,,  I  N -s
             = ff  (1-   s)e  ,
                                      2     1                            (5.31)
                                     s  =  —
                                    2 2
               42  V2*3        *!*!
      a£fV!y -   a     z                 3                               (5.32)
                      •n        (1 + ki^i^

The £i's are some characteristic scales of  the hole exponential  covariance
that can be uniquely related to the e"1 correlation scale.   The  notation \
will be reserved for the negative exponential covariance  function  correlation
scale.
                                                                     2
  Figure (5.3) compares the hole type with the exponential,  R(^)  =  a*  exp
(-£/X), when they have the same e~^ value for  the  same  distance  5.   The  two
covariance functions are practically indistinguishable.
                                        23

-------
                   RUTOCQBRELRTICN  FUNCTION
                                                         ~


          !\J

        ..L
        ...L
           |  I  J I  I I  I  I I  !  I I  I  I I  I
                           LflC OISTSNCi (CM1
Figure 5.2     Comparison of sample autocorrelation function (crosses)
               for hydraulic conductivity (from grain-size analysis)
               with theoretical autocorrelation function (solid line).
               (From the CRNL,  Ontario experiment,  Moltyaner,  1985)
                               24

-------
                                        = a-(s/3s0))exp(-(s/se»
                       3.0
                                4.0
                                        5.0
                               lag (g/x)
                                                 8.0
                                                         7.0
                                                                  9.0
Figure 5.3   Comparison of the exponential  correlation  function and
             the hole exponential  correlation  function
                          25

-------
APPROXIMATE ANALYTICAL ESTIMATIONS

    The covariance function of  the concentration  field  involves the evaluation
of the three-dimensional integral in Eq.  (5.12),  which,  for  the general case,
can not be performed in an analytical  fashion.  However,  for some simple
situations, approximate analytical expressions  can  be  found.  These expressions,
although not directly applicable  in all  practical situations,  reveal the
structure of the concentration  covariance and the influence  of the various
parameters.

    The covariance function in  Eq. (5.12) can be  written  as:

      Rcc (£ } = Tjp,e,l > GjGp                         (5.33a)

where, the Einstein's notation  for repeated indices  is  used  and

                                 K?   f.          •   S
    Tjp< WVW*'9'^ =  ~2 /   exp '-^  ,2
     J*                           q  -«              k
                  '  1 Vn    <            >  <*              (5'33b)


where the Sff (jc) has the form of  Eq.  (5.32)  and  j,  p, m,  n can take the
values 1,2,3.

    One can further  convert the integral  in  Eq.  (5.33b)  to a function of non-
dimensional variables (Appendix B) so  that Eq.  (5.33a) becomes


             Tjp(p1.P2'e''e'
                                                                               '•<**.
    (i)  The concentration variance is inversely  proportional  to the local
         dispersivity value.  This  is a physically intuitive  behavior  since the
         smaller the local dispersivity the less  the capacity  of the medium to
         attenuate developing tongues of fast moving solute.   Gelhar et al.
         (1979) and Matheron and de Marsily (1980)  have  described similar
         phenomena .
                                        26

-------
                                         Table 5.L
Case
                             APPROXIMATE ANALYTICAL SOLUTIONS

                                                 T    (E=0)
                                                                                (e=0)
1.


isocropLC i
2
°C
inis. .
2
Y
2
3

J_
KE

1
2
Y
1
9

22


2.  Anlsotroplc AnK
                        1    2  1  \_  \_
                                 e PI P2
                         2   3
                        Y
                                                 1   I  r  1   , _L_
                                                  2  6  L 2 +   _2
                                                 Y      2Pt     R
                                                                                  22
                                                                  R  -  2R
3.  Isotroplc £nK
a.  Rcc(e,0,0)
                        JL_  2_  I
                         2  3  KE
                            l-B-B2eBEU-B)]
                                                  2313
b.
                        1112
                         2 . KG 3 ?

                                                                    -16]
        (0,0,5)
                               »,£,0)
                                                  T22(e,0,0)
T2.,(0,?,0)
                                   < = — ,  R  -  1 -  p.  ,
                                       °L              l
                                                                             I  '
El(-x) = Exponential Integral,  K2(O = Modified  Bessel  function of 2nd order, *p2 = 1.

-------
(ii) The concentration variance is proportional  to  the  product  of  the
     correlation scales of the hydraulic conductivities.   Higher correlation
     scales imply that solute trapped  in a more  permeable  layer tends  to  stay
     there longer, since the formation persists  for larger  scales•   Consequently
     a randomly located measuring probe would have  higher  chances   for  a  "hit or
     miss" output.

         Vomvoris and Gelhar (1985) presented a  comparison  of the  variances  of
     two different media, an anisotropic and an  isotropic  one,  having  the same
     product of correlation scales (Figure 5.4).  The favorable comparison of
     the two results indicated the option, for variance calculations,  of
     replacing an anisotropic medium by an equivalent statistically isotropic
     one with correlation scale equal  to the geometric  mean of  the  anisotropic
     correlation scales.
(iii) The concentration covariance along the direction of  the mean  flow  has  very
      high correlation scales (Figure 5.5).  This singular  behavior is still
      under investigation.

THE NUMERICAL CALCULATION

    The equation for the concentration covariance, Eq. (5.12),  cannot be
evaluated analytically for a general combination of  the  input parameters.  A
direct three-dimensional numerical evaluation, unless very  carefully designed,
can produce unreliable results due to the irregular  behavior of the integrand.
A reduction to a two-dimensional integral can be accomplished if  one converts
Eq. (5.12) to spherical coordinates and integrates over  wave number u.   Appendix
B presents the details of the above integration.  The result is,  Eq. (B.7) and
Eq. (B.U),
   2*   ir
= /    /
w=o  g=o
                         du dg sine F  (f5,u)
                                               t
         {[A
            n
            A
             13  3
         (3+3
)]  exp[-
                J	I
                 PIB!
       exp -
where the different quantities:
                                      M2>
                                                          (5.34)
                                A2, A,E,R,B are  functions  of
and are defined in Appendix B.  The integral  in Eq.  (5.34)  is  evaluated  first
over angle 0 using the Romberg integration technique and  then  with  a  simple
trapezoidal rule over u>«
                                        28

-------
     8.
0)
c
o


.2   S.

§-  *°

£>   SJ
o   -:
        »-_-«
                               -e
        ^
          \
           v
             s — -e
         o.i
                0.8        1.6        2.4        3.2        4.0
f\j
  p
      O).

                10
      ct.'o       0.8
1.5
Figure 5.4   Comparison of Che variance for an anisotropic aquifer

             (correlation scales £1,  &2,  £3) and an equivalent

             isotropic (correlation scale i = («,,, £, ,  Jl
                               •nW-.               ±    £,
                                               --._
                              29

-------
z
o
UJ
o
u
CM

-------

-------
so
c
    sj
    CD
    g_
    s.
    OJ
    d-4.0
2      -2.4


     log (a/
                       approximatQ

                       analytical
                                     numGricai
-1.6
-0.3
                                                       -n
0.0
Figure 5.6    Test of  Che validity of the approximate result for  the

            variance Tu(l,l,e ,1;0) with respect  to the e values
                           32

-------
    The approximate analytical expressions in Table 5.1 are  evaluated  by
comparison with the numerical integration.  A comparison of  the numerical
results and the approximate analytical is shown in Figure 5.5.  The  agreement  is
very good for small values of eic = of/i-  Figure (5.6) compares the  expression
for the variance.  The analytical approximation gives identical results with the
numerical for values of eic = aj/i UP to 0.10.

    Table 5.2 shows the effects of the angles of orientation of the
stratification on the variance.  For the chosen correlation  scales the effect  of
the vertical inclination seems to be stronger.  In most cases  the values remain
of the same order of magnitude.  A notable case is Tj^, the  coefficient that
multiplies (-3c/3x1)(-3c/8x3).  Its magnitude increases up to  an order of
magnitude less than the value of TH«  Since the longitudinal  gradients are
very much milder than the vertical ones, such an increase could equalize the
effects of the longitudinal and vertical gradients in the estimation of the
variance.  Table 5.2 demonstrates the complicated interaction  of the correlation
scales, the-local dispersion coefficient and the orientation of the
stratification.  Further investigation of these relationships  is required.

    Figure 6.3, in the next section, shows the results of the  numerical
integration for the covariance of an isotropic hydraulic conductivity  field.
The presented curves are directly proportional to T^.  Very high correlation
scales along the flow direction are observed.
                                       33

-------
Table 5.2
EFFECT OF THE STRATIFICATION ORIENTATION ON
THE CONCENTRATION VARIANCE

*
0
0
0
0
5
5
5
5
10
10
10
10
15
15
15
15

e
= =5 = = =S=
0
• 5
10
15
0
5
10
15
0
5
10
15
0
5
10
15
(a) Pl
Tll
518.5
592.96
653.14
716.22
547.74
598.41
656.95
723.73
583.84
612.45
658.47
725.5
607.27
625.75
668.91
726.47
= 0.02 ,
T22
0.72
1.69
4.57
8.83
0.77
1.7
4.48
8.73
0.89
1.77
4.32
8.34
1.09
1.88
4.22
7.87
P2 = 0.2 ,
T33
0.76
3.51
11.59
24.08
0.81
3.62
11.82
24.57
0.92
3.77
11.87
24.44
1.07
3.79
11.75
23.68
e = 0.1
T12
0
0
0
0
-2.27
-1.92
-1.81
-1.41
-5.83
-5.24
-4.19
-2.5
-10.07
-9.35
-7.65
-4.
, K = 1
0
30.05
64.58
102.0
0
30.87
65.05
102.06
0
32.01
65.79
101.97
0
32.47
65.99
101.31

T23
======
0
0
0
0
0
-O.Ob
-0.07
-0.01
0
-0.27
-0.26
-0.05
0
-0.47
-0.62
-0.23
   34

-------
Table 5.2 (continued)

*
0
0
0

0
5
5
5
5
10
10
10
10
15
15
15
15

e
0
5
- 10

15
0
5
10
15
0
5
10
15
0
5
10
15
(b) p
TI
228
241
264

291
230
242
264
291
236
246
266
292
242
250
268
294
!~- 0.
1
.45
.45
.17

.39
.72
.84
.75
.74
.25
.19
.09
.59
.9
.97
.94
.08
05 ,
T22
0.
0.
2.

3.
0.
0.
2.
3.
0.
0.
1.
3.
0.
0.
1.
3.
P2 =

52
9
01

73
54
9

7
57
92
97
6
62
95
93
46
0.2
T33
0.
1.
4.

9.
0.
1.
4.
9.
0.
1.
4.
9.
0.
1.
4.
8.
i £

61
66
68

34
62
67
67
3
66
68
64
18
71
71
57
94
= 0.
Tl
0
0
0

0
-1
-0
-0
-0
-2
-2
-I
-0
-3
-3
-2
-1
1 , K
2





.07
.97
.74
.34
.29
.04
.57
.73
.49
.23
.50
.21
= 1
T13
0
12.4
25.62 •
i
39.89
0
12.44
25.63
39. 8J
0
12.51
25.62
39.62
0
12.53
25.51
39.2

T23
0
0
0

0
0
-0.05
-0.06
-0.01
0
-0.11
-0.14
-0.04
0
-0.17
-0.22
-0.09
         35

-------
Table 5.2 (continued)

*
0
0
0
0
5
5
5
5
10
10
10
10
15
15
15
15

e
0
5
10
15
0
5
10
15
0
5
10
15
0
5
10
15
(c) Pl - 0
Tll
2510.15
2587.02
2740.71
2987.12
2519.89
2592.32
2745.17
2989.86
2545.33
2609.93
2758.45
2998.03
2581.12
2637.65
2780.19
3011.38
.05 , P2 =
T22
0.57
4.23
14.75
30.86
0.72
4.32
14.71
30.64
1.14
4.61
14.61
29.99
1.8
5.04
14.42
28.96
0.2 , e
T33
0.75
11.54
42.66
90.41
0.81
11.52
42.39
39.77
0.98
11.45
41.6
87.8
1.24
11.29
40.26
84.69
= 0.01 ,
T12
0
0
0
0
-15.39
-14.74
-12.14
- 7.37
-30.74
-29.29
-24.08
-14.6
-45.66
-43.34
-35.54
-21.52
K = 1
T13
0
135.
.i 275.08
421.07
0
13^.76
274.3
419.7
0
133.95
272.14
415.73
0
132.39
268.3
408.9

T23
0
0
0
0
0
-0.76
-1.18
-0.97
0
-1.5
-2.33
-1.93
0
-2.17
-3.38
-2.81
        36

-------
         Table 5.2 (continued)




(d)  PI = p2 = 0.25 , e = 0.02  , K =  1




     T22       T33       T12       T13        T23
0
5
10
15
259.92
263.06
272.34
287.58
0.28
0.39
0.79
1.41
0.37
0.78
1.97
3.79
0
0
0
0
0
8.43
16.92
25.42
0
0
0
0
                  37

-------
                                   SECTION  6

                             APPLICATION EXAMPLES

    Two hypothetical examples will show the applicability  of  the results
developed in Section 5.  The examples focus on  the  estimation of the
concentration variances.

The Continuous Concentration Input

    If solute mass is injected at a constant  rate at  the origin (x=y=z=0)  in a
three-dimensional aquifer with constant mean  groundwater flow velocity,  the
resulting steady mean concentration distribution is approximated by the
following form when x» y,z:
                                              r2       Z    i             / ^  i \
                   c(x) = 	   m  	 exp-t
where A__, A_. = lateral macrodispersivities along y,  z
      n        = porosity
      m        = rate of mass injection
      V        = mean pore velocity (assumed along x)  =  q/n
      x_        = (x,y,z) location vector

    The components of the gradient of the concentration  are  given  from,

                                         2    2    .
                                         f    -—ill
                                             iA33 x
                                                                        (6.2)
The concentration variance is then given from  the  form  in  eq.  (5.33c),

                        2    2 2 T   $1  3i                             (6.3)
                       ffc " °f*3  ij 3x£ 3x

where the T . 's will be computed from the approximate analytical  ex-
pressions in Table 5.1.  The following parameters  are used [x^ =x^ t*2 =v >X3=Z^

      A   = 1.0 m.

      A22 = A33 = °'01 ™'
      a   = 0.01 m.
                1 m.
      af  = 1
The concentration is normalized by the quantity  c =m/[ nA   A__Vj .

    Figure 6.1 shows a cross-section of the plume at  x=50  m.  and  z=0  m.
plotting the mean concentration plus and minus a standard  deviation.   It  is a

                                       38

-------
                     actual  ccnc.?
c
0
c
a
u

o
a
o
a
=  CD •
   10
                                          3.0
Figure 6.1  Normalized  mean concentration and standard deviation at

           x - 50 m,  z = 0 m
                         39

-------
logarithmic plot which enables one to observe the increase of the coefficient
of variation at the tails of the plume.  The fluctuating line is a sketch to
illustrate a possible concentration realization.  At locations y=z=0 ra. due
to symmetry the gradients (3c/3y), (3c/3z) = 0; however, (3c/3x) # 0,
and it is responsible for the approximately 0.25 coefficient of variation
observed at this location.  Figure 6.2 shows the isoconcentration lines and
the standard deviations in a horizontal plane located at z=0 m.  This form of
plot allows one to visualize how much the contoured configuration of a plume
may vary.  Figure 6.3 shows the correlation function of the concentration for
points located at the longitudinal axis of symmetry of
the plume.  Along the flow directions extremely high correlation scales
are observed.

    Along the center line of the plume, the coefficient of variation can be
written as

             !                              0'5                        <6-4)
                                    vx <• Jo'
                             c      '
                                                                     2
and for the parameters chosen with the isotropic relationship y=exp(af/6)

                            ac    7
                            —  = —    , x in m.                      (5.5)
                            ™     X
                            c

Thus for distances higher than 100 correlation scales the coefficient of
variation will be a few percent.   Figure 6.1 shows that this should be viewed
as the minimum concentration variability at a given longitudinal distance from
the source.  The implied near source behavior with very large coefficient of
variation must be viewed as only a qualitative feature since the perturbation
approach used to develop the coefficients T.. (Section 5) was based on the
assumption that the concentration variations'1 were relatively small.

The Pulse Injection

    As in the previous case, the form of the mean equation is assumed known,
with effective parameters resulting from the aquifer heterogeneity.  The mean
concentration distribution for an instantaneous point injection at the origin
is

  zUtt).  	L	    -rrtsaai.^.^J^w-M
             f I  ••\J*^/'»  A  A  \ -^ / *       ^" 11        1 1       1 "J
            n(4irtV)   (AnA22 33'
                                          .Y
where M = mass injected

      A  , A,,, A  = macrodispersivities along x, y, z axes respectively

      V = mean pore velocity                             - -'„-
      £= (x.y.z)  = (Xj.X2.X3)

Denote Vt = x , the location of the center of mass of the plume at any
time.  The concentration gradients are given from


                                       40

-------
                                                            c- .00001
CD
t-«
X
a
                                   x  -axis
           Figure 6.2   Isoconcentration  lines and standard deviations at

                       z = 0 m

-------
5
5 2
a i
  d
   3.
                        X-AXIS
          iso-corrQlation contours
     i — o- ®- — ®- — — — — Q ---  .
                     0.3
  (D- O- -4D— —• — -©— —	
                 0.5

>	o	-o	.	
                 0. 7 ~
     (9	o	
                      0.9
   rtl.c)      4.0      0.0       12.0     111. 0

                         X-AXIS
                                              20. U
                                                         6.3  Correlation function for a point at y -  z  =  0
                                                              (a)  Correlation along x
                                                              (b)  Correlation along lines in the (x,y)  plane at
                                                                   various angles with x
                                                              (c)  Lag-correlation contours In the  (x,y) plane

-------
                                x-x
                        3c _ -     c
                        3x ~ c  2A
    The concentration variances were calculated  using  the  numerical evaluation
for an anisotropic heterogeneous aquifer with  parameters,
                                               2
              £L= 3m  , i2=  1m  ,  H3= 0.15ra  , af  =  0.5  ,  o  = 0.01m
The macrodispersivities are

                            A  = 0.3ra  , A   = A__=  0.01m


and the values for T.., T_2, and T~_ are taken from  Table  7.4.

The concentration is  normalized by

                           V M/[(4ir)3nAuA22A33]

Figure 6.4 and 6.5 show the mean plus  ana minus, a standard deviation for a
longitudinal and vertical section respectively along the axes of  symmetry of
the plume.  Note that at the maximum concentration a  =0.   This  is
because the concentration gradient  is  zero  at  this location.  The analysis was
developed assuming that the concentration gradient is  constant.   This is
clearly not true around the concentration peak,  so it  is not  expected that the
results hold at this  region.

    Figure (6.6) shows the concentration variance  in a vertical  section 5 m.
away from the concentration peak in the x direction  and  1  m.  in  the lateral.
The minimum coefficient of variation (at the peak) is  not  zero anymore.

-------
                                   Longitudinal Section
o
a
                                  42  43  44  45 46 47  48  49 50 51  52  S3 54
           -3
36 37 38  39  40
     Figure  6.4   Normalized concentration and one standard  deviation
                  interval along a longitudinal section  (line  coordinates
                  y = z = 0)

-------
                                 Vertical   Section ( x=xc )
a>
o
                      i   l   i   i   l   i   i   l   i
                -2-1.8-1 6-1 4-1.2-1-08-0.6-0.4-0.2 0
    i   i
0.2 0.4 0.6 0.8
1  1 2 1 * 1 6 1.8  2
  Figure 6.5   Normalized concentration and one standard  deviation
               interval along a vertical section (line  coordinates
               x  = xc - center of mass  of plume, y = 0)
                                  45

-------
                        Vertical  Section (x=40,z=1)
           1.8-1 fi-1 4-1.2-1-08-06-04-0.2 0  0.20.40.608  1  1.21.41.618  2
Figure 6.6  Normalized  concentration and two standard deviation
            interval along  another vertical section (line
            coordinates  x = xc - 5 m, z = 1 m)
                             46

-------
                                   SECTION 7

                   FIELD DATA ANALYSIS: THE CFB BORDEN SITE

    The results from the stochastic theory will be applied in  the  controlled
field scale experiment at Canadian Forces Base, Borden, Ontario.   This  is  a
preliminary application since the available data represent two  sampling dates
only, namely, October 25, 1982 and November 28, 1983 (Freyberg,  1984, personal
communication).  The tracer site is located in a sand quarry down-gradient
from an abandoned landfill (Sudicky, 1985).  The tracer injection  occured  on
August 23, 1982, and comprised of 10.7 kg of chloride solution  injected at an
average concentration of 892 tng/1 from a zone between 2.0 m and  3.6 m below
ground surface (Figure 7.1).  Table 7.1 summarizes part of the  existing
information about the site.

    The stochastic theory can be applied to produce estimates of:
    (i)   effective hydraulic conductivities;
    (ii)  macrodispersivities; and
    (iii) variance and covariance of the concentration deviations.

    The input data set required must include:
    (a) the mean log-hydraulic conductivity, F, and its standard deviation,
        af.
    (b) correlation scales of the log-hydraulic conductivity field, £nK;
    (c) direction of its principal axes; and
    (d) local dispersivicles.

Prior to analyzing the data a preliminary investigation of their information
content is required.
                                   Table 7.1

              EXISTING INFORMATION ON THE BORDEN SITE PARAMETERS


a2£ = 0.3 - 0.5

fc. 3 2 - 5 m; i2 s  2 - 5 m; fc_ = 0.10 m

AU s 0.30 m; AZZ =» 0.034 m; A33 - « 1* m

    = F = KG = 7.15 x 10~3 cm/s

  = -5°

  Approximately zero was assumed by Sudicky (1985) without actual  analysis
  of the vertical spreading; quantitative analysis of data from an earlier
  test at the Borden site (Gelhar et al. 1985) shows that vertical spreading
  is greater than molecular diffusion (A^ = 1 to 2 mm).

                                    47

-------
    60
    40 r
g
x
     20 L
                             1  20 .'A'
                                                           »—'-4
                                                        \
dwa
                                                 + Multilevel Somolers
                                                 ^ Injection Wells
                                                 a Piezometer
                                                 • Core Location
         -20
                   20
40
                                Y(m)
      Figure 7.1  Instrumentacion,  locacion of  cores  and hydraulic  head
                  distribution on August  10V  1984  (solid line)  and
                  November 15, 1984 (dashed line)  at  the tracer-test  site
                  (after Sudicky, 1985b).  The  circled samplers  show  the
                  locations of the wells  used in this analysis.  V
                  denotes the well used for the vertical section
                  analysis.
                                   48

-------
       MAX
       MIN
31 S
O
       CONTOURS
        1
        2
        3
        4
        5
        B
2. 5
ID
6O
100
2OO
260
vo
                                  Figure 7.2  Isoconcentration contours ac depth  z =  -5.31m,  from the
                                              Borden site data (original data  from D.  Freyberg,
                                              personal communication)

-------
Averaging Effects in Sampling

    The theory developed assumes point values of all  the  processes,  where as
it is traditional in groundwater studies, a "point" is defined  in  a
Representative Elementary Volume (REV) sense, or, at  the  same  scale  that  the
convective-dispersive equation for the mass transport holds.   However,  the
data collected are directly affected by  the resolution of the  sampling
method.  For the hydraulic conductivity  data at the Borden site, the values
were determined from cores 0.05 m. in height and 0.05 m.  in diameter.   Even at
this scale the sampling method introduces an averaging of the  true process
which results in reduced variances and increased correlation scales  (Vanmarke,
1983).  A similar analysis was performed at the hydrothermal test  site
Aefligen, in Switzerland (Hufschmied, 1985).  The so-called "variance
function", x(a)» describes the variance  reduction in  terms of  the  correlation
scale.  For a one-dimensional case the measured process is given by,


            .     z + Az/2
    f(z) =  i   /  f(z) dz                                              (7.1)
            AZ
                 z - Az/2

and its variance is:
     2      2
    J_  =  Of
     f
                                                                        (7.2)
    With the exponential model proposed by Sudicky  (1985a)  for  the  vertical
correlation and with i = O.I m, Az = 0.05 in  (the height  of  the  sample),
x(0.05) s 0.85.  Therefore, the variance of  the point values  of  the
log-hydraulic conductivity should be increased by  15%.

The Spectrum of InK

    The estimation of the concentration variance requires a spectrum that
enforces stationarity on the underlying process.   Such a spectrum has  a  zero
integral scale in three dimensions (Section  5).  The exponential covariance
function does not possess this property, and this  is the reason  that a
"hole-type" covariance function was used in  the theoretical development.   The
difference between these two covariance functions  (see Figure 5.3)  is  so  small
as to be practically indistinguishable in terms of  statistical  inference  and,
as will be shown in the following, the shape of the spectrum  has minimal
effects on the estimation of the effective hydraulic conductivities and  the
macrodispersivities.
                                        50

-------
    The important parameter is the correlation  scale  which  is  defined by the
lag corresponding to the e~l level of  the correlation function.   In the
isotropic case the correlation function used  is,


    p(s)  =  (1 - y s) e"S ,    s = |                                   (7.3)

and the e-1 level is given by s = 0.723,

       p(0.723) = e"1  ,  Cg = 11(0.723)                                 (7.4)

    Notice that the correlation scale  of the  above  function is different than
the length scale dividing the lag.  The opposite  holds  for  the exponential
correlation function.  It is easy to verify,  since  the  exponential  correlation
function is given from
    PE

that
              e"S'   , s' =f   ,                                         (7.5)
    PE(1) = e"1. 5e = X.                                                (7.6)

    Consequently, the correlation scales determined  by  Sudicky (1985b)
correspond to the X's (in the notation used  here), and  the  correlation  scale
for the hole type correlation function, which  is  used  in  this  analysis,  is
given from,
        0.723
                                                                        (7.7)
    The above relationship can be expanded  to  the  anisotropic  case  also,  where
the following will hold:

    1. =     Xi   ,  i =  1,2,3                                          (7.8)
           0.723

    Therefore, the values given  in  Sudicky  (1985)  Xi=X2=2  m> X3=0.1 m,
would correspond to

    A1 = 12 = 2.76 m, Z3  = 0.14  m                                        (7.9)

    With these preliminary remarks, required because of  the  sampling technique
used and the covariance model fitted,  the stochastic theory  can  now be  applied
to estimate the quantities mentioned in  the introduction.

Effective Hydraulic Conductivities

    Gelhar and Axness (1983) have developed the general  equations  for the
effective hydraulic conductivities  in  terms of the spectrum  of the
log-hydraulic conductivity.  The analysis evolved  around a linearized version
of the flow equation and  was heuristically generalized for the case of  af > 1.
The effective hydraulic conductivities along the principal axes  of  flow are
given from
                                      51

-------
    Ku-Kt  exptoj  ({-g.t)],  i = 1,2,3                            (7.10)

(no summation over repeated  indices), where K£ = exp [E(lnK)] , and the gi;j's

can be calculated from



    Si-=  /   ~SV~  sff  (^f)  d-                                     (7*1L)
      ^    k  k'  a


    The primed system of  axes  is aligned with the principal axes of the
hydraulic conductivity.   Freyberg (1985) has observed an approximate 5°
clockwise horizontal  inclination of the principal axes of the plume with
respect to the mean flow.  This  is an indication of horizontal anisotropy in
the hydraulic conductivities,  in contrast with the reported values of the
correlation scale, and a  counterclockwise alignment of the principal axes of
the plume with respect to the  mean flow, or, plume trajectory.  Table 7.2
summarizes a series of calculations where the input parameters were the first
seven columns, i.e. of, 4 ,  A,,  *., a = local dispersivity, * = horizontal
angle of x1 and x, 6  = vertical  angle of xj and x (where x[ = major principal
axis of hydraulic conductivities, x_= mean flow direction).  In columns 8,9,
10 the estimated values of K ,/K., K  /'<    and K3,/X do not show significant
variations and compare quite well witn  tne values of 1.18 and 0.85 reported by
Sudicky (1985).  The  high values correspond to variances not considered by
Sudicky (1985), i.e., of  = 1.5,  and a" = 1.0.  It can also be observed
that the anisotropy in the effective hydraulic conductivities both in the
horizontal as well as the vertical plane is very mild even for aspect ratios
of the original field of  ^/^ = 5/0.5  = 10.

    Note that the coefficients gtj , Eq.  (7.11), needed for the estimation of
the effective hydraulic conductivities  in Table 7.2 were evaluated numerically
for the hole type covariances function, Eq. (5.32).  No significant difference
from the values shown in  Gelhar and Axness  (1983) was observed.

Asymptotic Macroscopic Dispersion

    Gelhar and Axness (1983) have  developed expressions for  the asymptotic
macrodispersivities for the general case of arbitrary orientation of
stratification.  These expressions  are  quite  complicated  to  be evaluated
analytically except for the case of:

         hydraulic conductivity isotropy  in the horizontal or vertical plane,
         negligible local dispersivity,
         exponential  log-hydraulic conductivity covariance,  where approximate
         solutions can be obtained.
                                     52

-------
in
U>
                                                      Table 7.2


                             EFFECT OF THE INPUT PARAMETERS ON EFFECTIVE  CONDUCTIVITIES
i.
If
I
0.4
1
1
1
0.5
0.5
0.5
0.5
1.5
1.5
1.5
1.5
0.5
0.5
0.5
0.5
0.5
1
1
1
1
Aj (m.)

5
5
2
2
2
2
2
2
2.5
2.5
5
5
2
2
2
2
5
2.5
5
2
2
*2 (m.)

0.5
0.5
0.5
2
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
1
1
*3 (m.)

• 0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
a (m.)

0.01
0.01
0.01
0.01 '
0.01
0.01
0.01
0.01
0.01
0.001
0.01
0.001
0.01
0.001
0.01
0.001
0.001
0.001
0.001
0.001
0.001
4>

0
5
10
5
0
5
0
0
20
20
20
20
20
20
45
30
25
25
20
45
20
Q
33
5
5
0
10
0
0
5
10
0
0
0
0
0
0
0
0
0
0
0
0
0
K/ If
1 1 ' A
11 I
1.22
1.64
1.61
1.59
1.27
1.27
1.27
1.27
2.07
2.07
2.1
2.1
1.27
1.27
1.27
1.27
1.28
1.62
1.64
1.6
1.6
K / K
^11' ^«
22 4
1.14
1.4
1.41
1.59
1.18
1.18
1.18
1.18
1.66
1.66
1.65
1.65
1.18
1.18
1.18
1.18
1.18
1.4
1.39
1.52
1.52
K /K
33 I
0.87
0.72
0.72
0.65
0.85
0.85
0.85
0.85
0.61
0.61
0.61
0.61
0.85
0.85
0.85
0.85
0.85
0.72
0.72
0.67
0.67

-------
    Using the same set of input parameters  as  in Table  7.2,  a  numerical
evaluation of the asymptotic macrodispersivities was  performed assuming  the
covariance function of the inK is the hole  type, and  results are  shown on
Table 7.3.  Recall from Table 7.1 that the  estimated  longitudinal dispersivity
for the side is 30 cm. , and that there is a clockwise rotation of the plume
with respect to the mean flow.  These two characteristics will be the
calibration targets for the selection of the correlation scales.

    Table 7.3 shows the different combinations of input parameters  that
produce a longitudinal macrodispersivity of the same  order of  magnitude  as the
observed ones.  It is more difficult to match  the 5°  clockwise rotation; Note
that most of the input combinations produce angles between  1°  and 3°.  The
predicted asymptotic lateral macrodispersivities are  smaller -by 2 to 3 orders
of magnitude.  In that case the local dispersivity, which is additive, will
be the prevailing value.  Gelhar (1986) in  an  analysis  of the  developing
macrodispersivities has shown that the lateral dispersivities  attain a maximum
value and then drop to their asymptotic magnitude. Dagan (1984)  derived a
similar result following another approach.   A  plausible speculation,
therefore, is that the measured field value for A22 of  about 3 cm,  38 m
downstream from the injection point, is still  influenced by  the developing
transverse dispersion process and, hence, is much larger than  its asymptotic
value, which is the one that the spectral theory estimates.

    With  the information presented in Tables 7.1, 7.2 and 7.3, it is
considered that a representative set of values for the parameters of the CFB
Borden is as in Table 7.4.  The correlation scales have been adjusted for  the
hole type covariance.  The variance is slightly increased  to account for
sampling  averaging, and anisotropic behavior is included t'o produce the
longitudinal macrodispersivity.

Variance  of  the Concentration Field

    The theoretical analysis of Section 5 has developed the following form for
the concentration covariance,

    RCC - °f 43 T1J(p1,P2,c,K,+ ,8;£) G^.                           (7.12)

The analysis  has  shown  that  the  controlling local dispersivity is a^.
 (KE = oT/A3)  and  that the  effect of  , 6 is minimal.   Therefore as a first
approximation the variance of  the  concentration  field will  be calculated from:

     °   "  Gf  *3  Tiii'V6'''5-   ii
                                    GG
      c

-------
Ul
Ul
                                                 Table  7.3

                          EFFECT OF THE  INPUT  PARAMETERS ON MACRODISPERSIV1TIES
2
°f
0.4
1
1
1
0.5
0.5
0.5
0.5
1.5
1.5
1.5
1.5
0.5
0.5
0.5
0.5
0.5
1
1
I
I
tj.)
5
5
2
2
2
2
2
2
2.5
2.5
5
5
2
2
2
2
5
2.5
5
2
2
££.)
0.5
0.5
0.5
2
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
1
1
t£.)
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
a(m)
~~~"™
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.001
0.01
0.001
0.01
0.001
0.01
0.001
0.001
0.001
0.001
0.001
0.001
*
0
5
10
5
0
5
0
0
20
20
20
20
20
20
45
30
25
25
20
45
20
6
5
5
0
10
0
0
5
10
0
0
0
0
0
0
0
0
0
0
0
0
0
M->
0.211
0.276
0.209
0.187
0.187
0.182
0.167
0.126
0.175
0.28
0.185
0.31
0.14
0.224
0.099
0.181
0.236
0.264
0.316
0.306
0.383
Aj,
0.27
0.88
0.38
0.72
0.24
0.25
0.32
0.48
0.91
1.3
l.l
1.9
0.31
0.12
0.27
0.16
0.24
0.79
0.84
0.19
0.11
A33
0.62
1.5
0.79
0.27
0.54
0.56
0.77
0.12
1.1
0.34
1.1
• 0.33
0.76
0.19
1
0.21
0.2
0.31
0.3
0.27
0.24
A1
Al
0
-3
-2
0
0
-0
0
0
-8
-18
-1
—2
-0
-4
-2
-4
-6
-13
-15
-7
-5
2

.1
.4
.13

.56


.8

.1
.4
.94


.7
.7



.6
A1
A13
4.3
14
0
6.6
0
0
2.9
5.9
0
0
0
0
0
0
0
0
0
0
0
0
0
Aft.)
0.211
0.276
0.209
0.187
0.187
0.182
0.167
0.126
0.175
0.28
0.186
0.312
0.141
0.224
0.099
0.181
0.236
0.265
0.317
0.307
0.384
A*1
A22
0.27
0.63
3.6
0.72
0.24
0.25
0.32
0.48
0.46
0.13
0.49
0.12
0.3
0.049
0.23
0.045
0.047
0.092
0.09
0.028
0.032
A*1
A33
0.53
0.88
0.79
2.5
0.54
0.56
0.72
0.92
1.1
3.4
1.1
0.33
0.76
0.19
1.1
0.21
0.2
0.31
0.3
0.28
0.24
*
0
-0.66
-0.66
0.04
0
0
0
0
-2.9
-3.8
-3.4
-4.4
-3.8
-1
-1.2
-1.5
-1.6
-2.9
-2.8
-1.3
-0.85
n
1.2
3.1
0
2.1
0
0
1
2.7
0
0
0
0
0
0
0
0
0
0
0
0
0
    1 Units are in
                   mm.

-------
                               Table 7.4

              PARAMETERS USED  IN THE VARIANCE  ESTIMATION


a2 = 0.5
 f
1. = 3.0 m; 4, = 1.0 m; l  = 0.15 m
x  » cu =a=0.01m.
Tu = 776.09; T22 = 1.03; T33 = 0.88
                               Table 7.5

            CONCENTRATION GRADIENTS IN THE VERTICAL  SECTION
                    AT (x,y) = (38.3 m.,  16.08 m.)
z (m.)
-3.51
-3.81
-4.11
-4.41
-4.71
-5.01
-5.31
-5.61
-5.91
-6.21
-6.51
3c
3x
0.185
0.981
4.824
1.25
6.838
6.64
18.21
1.7
2.16
-13.79
-2.83
ii
ay
-.707
-4.14
-5.61
-7.77
3.64
-4.92
-1.75
1.934
9.731
20.78
20.48
*
R
0.3788
0.579
0.627
0.641
0.411
0.205
0.208
0.087
0.508
0.735
0.825
      * R = coefficient of multiple determination,  which is  equivalent to
            the correlation coefficient  in  univarlate  regression analysis
            (Neter and Wasserman,  1974).
                                   56

-------
    The values for the coefficients T   are shown  in Table  7.4.   They were
calculated using the numerical integration programs developed  under  the
general name TIJ_ND (where I, J = 1,2,3).  The next task  in the  calculation
of the variance of concentration is the estimation of  the mean gradient.   Two
sections will be shown, an approximate longitudinal one and a  vertical one
(Figure 7.1).  The approximate longitudinal section, follows a line  at
z = -5.31 m and is rotated at an angle of about  15° clockwise  from the
principal axes of the computed macrodispersivities.  Although  at this depth
the choice of this direction appears justified (Figure 7.2), it  was  mainly
necessitated by the location of the sampling wells, the objective being  to
use the original raw concentration data.  Consequently, this direction will
be considered as the x direction for the calculation of the gradients.  The
vertical gradient was estimated through analytical curves,  of  a  Gaussian type,
that were fitted to the data at each well along  x.  The gradient along y was
estimated roughly from the plotted contours.  For  the  vertical section,  at
(x,y) = (38.3 m, 16.08 m), the same Gaussian curve fitting  as  before was used
to estimate the vertical gradient, while the gradients along x and y (Table
7.5) were estimated by fitting a regression plane  through that point and its
eight surrounding ones (Neter and Wasserman, 1974).  These  procedures are
summarized in Figure 7.3.

    The longitudinal section with the fitted mean  is shown  on  Figure 7.4.
The series of Figures 7.5(a) through (i) show the  fitted  curves  in the
vertical direction along the wells.  Appendix D  describes the  parameters for
each curve.  Finally, Figure 7.2 is used to estimate the gradients along the
y direction.   All this information is collected in Table 7.6  and with the
T..'s presented at Table 7.4 the variance of the concentration can easily
be^estimated.  Figure 7.4 shows the (c + a , c - a) interval which will
include approximately 67% of all measurements ff concentration is normally
distributed.

    Table 7.7 shows the data used for the vertical section  at  (38.3, 16.08).
The gradients 3c/3x,  3c/3y are calculated from  the linear  regression as
described above.  The c +_ a  interval is plotted in Figure  7.6.   The effects
of the vertical gradients on the variance are quite strong  in  this case.   A
three-order of magnitude difference in the coefficients T..  and  T,_  is
balanced by a four-orders of magnitude difference  in the  squared gradients
(Table 7.7).  The lateral horizontal gradient has a minimal contribution.  The
logarithmic version of the previous plot shows the increase of the coefficient
of variation at the tails (Figure 7.7).
                                      57

-------
— : gradient along  this direction  was  estimated by a continuous curve fit


	: gradient along  this direction  was  estimated from contour plotting


-R  : gradients on  this plane were estimated  from linear regression
    Figure 7.3  Schematic presentation of the methods used for the
                estimation of the mean concentration
                                  58

-------
                                    L.ONG I TUD I NAL_   l_ I Nl
Ul
VO
                                                             I  '  '   '  I  '  '  '  I
                                                                       5     3O     35
                                                        X
                      Figure 7.4   The measured, fitted and plus-minus one standard
                                   deviation concentration profiles along the  longitudinal
                                   section at z = -5.31rn

-------
-334 -3.54 -3 64 -4 14 -4.44 -4 74 -504 -S.34 -5 fr» -£• ft ~« 24 -« &4 -6,64
                                                    -3.21-3 SI -381-4.11 -4 41 -4 71 -5 01 -S31-SC1-SV1 -«.21-8.51-e.Bt-7.11
-3 IB -34» -37B —4 OB -4 30 -4 «B -4 Co -£. X -5 S» -ffc. -« IB -6*
                                                      -31! -342 -372 -4 CC -132 -462 -462 -E 22 -S 52 -£. 82 -« 12 -642 -e 72
               Figure  7.5  The mean  concentration curves  fitted  at the wells on
                             the longitudinal section  (c  in the  vertical vs.  z)
                             +=fitted,O =measured

-------
-312 -342 -3.73 -4 CO -t 33 -4 «3 -4 S3 -122 -513 -S B2 -« 12 -e 42 -« 72
                                                      -343 -373 -4O5 -4.33 -4 «3 -4 B3 -123 -5.S3 -6.B3 -«.13 -« 43 —« 73
                                                   Figure 7.5    (continued)  The mean concentration
                                                                   curves  fitted at  the wells  on the
                                                                   longitudinal  section
 -3 14-3 44-3 74-4 O4-4 34-464—« 64-5 34-5 54-S b< -fl 14-844-674

-------
Table 7.6
ESTIMATION OF THE RANGE OF VARIATIONS
IN THE LONGITUIMNAL SECTION
x (in.)
27.25
31.3
35.3
38.3
40.3
42.3
44.3
46.22
48.35
50
52
y (m.)
6
9
12
16.08
17.94
19.95
22
24
25.95
28
30
z (m.)
-5.39
-5.31
-5.34
-5.31
-5.29
-5.22
-5.22
-5.23
-5.24
-5.3
-5.3
c (mg/z)
(meas. )
1.6
2.3
31.1
169
319
243
208
123
19.3
3
2
c (mg/t)
(fitted) '
2.82E-4
0.17
14.76
183.18
312.97
296.61
149.67
42.20
5.98
0.57
0.06
3c/ 3c/
0.0004
0.18
10.23
55.29
28.34
-38.11
-52.36
-23.81
-4.71
-0.57
-0.07
0
0
0
0
7
0
0
0
9.3
0
0
3c
3z
0
0
15.57
86.96
207.06
233.59
129.9
23.43
35.65
0
0
«c
0.001
0.55
30.29
163.62
86.26
115.00
155.26
70.39
14.40
1.68
0.23
0.001
0.73
45.06
346.80
399.23
411.62
304.93
112.59
20.38
0.29
0.29
«=
-0.0094
-0.38
-15.5
19.55
226.70
181.61
-5.59
-28.19
-8.42
-1.11
-0.16

-------
Table 7.7
ESTIMATION OF THE RANGE OF VARIATIONS
IN THE VERTICAL SECTION
z(m.)
-3.21
-3.51
-3.81
-4.11
-4.41
-4.71
-5.01
-5.31
-5.61
-5.91
-6.21
-6.51
-6.81
-7.11
c(mg/fc) c(mg/A) 3c/3z
(measured) (fitted)
0.4
4.2
9.2
16.4
63.6
112
254
169
306
200
132
13.1
5.6
1 ' 7.8
0.12
0.88
4.84
19.85
60.6
137.9
234.06
296.06
279.17
196.24
102.83
40.17
11.69
2.54
0.85
5.44
25.1
83.57
195.82
310.64
297.92
86.96
-191.33
-326.63
-271.85
-145.53
-53.86
-14.17
3C/3X
0
0.18
0.98
4.82
1.25
6.83
6.64
18.21
1.7
2.16
-13.79
-2.83
0
0
3c/3z
0
-0.70
-4.14
-5.61
-7.77
3.64
-4.92
-1.75
1.93
9.73
20.78
20.48
0
0
°c
0.08
0.77
3.86
. 16.52
19.90
-37.00
35.62
54.50
19.74
33.23
49.00
16.90
5.37
1.41
**«.
0.20
1.65
8.70
36.37
80.51
174.95
269.68
350.56
298.91
229.47
151.84
57.076
17.071
3.95
c -a
i
0.03
0.10
0.98
3.32
40.69
100.93
198.43
241.55
259.42
163.00
53.82
23.26
6.32
1.12

-------
                 VERTICAL.   l_ I Nl
            I             I
Figure 7.6  xiie measured, fitted and plus-minus one standard
            deviation concentration profiles along the vertical
            section at (x,y) = (38.3, 16.08)

-------
    ooo
     i  oo
o      10
QL

Z         1
LJ
O
Z
O
o         i
       O 1
VERTICAL   LINE


            1   '   '    I   '   '
                              T	T
                                         J	L
                                                       l   I   I
                                            X
            Figure 7.7  The values plotted  in  Fig.  5.6 in a serai-logarithmic
                        plot

-------
Conclusions

    The analysis of the field data demonstrates  the  validity  of  the  stochastic
theory results for the concentration variance.   The  predicted concentration
intervals defined by +_ a  are consistent with the observed fluctuations  of
concentration around a smooth mean and capture most  of  the measured  values.
The effect of the longitudinal concentration gradient in the  calculation of
the variance is dominant except when focusing on vertical sections where the
vertical concentration gradient is higher by 2 or 3  orders of magnitude.

    The estimated effective hydraulic conductivities are quite close to  the
observed geometric mean of the measured hydraulic conductivities.  The
estimated longitudinal macrodispersivity is also comparable to the one
estimated through the method of moments.

    The analysis reveals the need for the development of a formal  procedure
for the estimation of the mean.  Analysis of data from  subsequent  sampling
dates is needed to show the robustness of the results.
                                    66

-------
                                   SECTION 8

                             DISCUSSION OF RESULTS

    A simple theoretical result was developed to quantify  the  concentration
variability.  When measurements of the concentration exist,  it  can  be  thought
of as giving a measure of the range of expected variation  of the  concentration
around a smooth mean concentration distribution.  The  theory is also useful  as
a predictive tool when the solute distribution is extrapolated  for  distances
far downstream using a classical dispersive transport  model.

    The functional form of the concentration variance  is easy  to  implement.
It is a local relationship that involves  the mean concentration gradient,  the
variance and correlation scales of the log-hydraulic conductivity field,  and
the local dispersivity values.  The mean  concentration gradient can be
estimated either in a predictive context  with numerical or analytical  models,
or through a simple curve fitting with existing plume  measurements. The
former approach was demonstrated in Section 6 and the  later  by  the  analysis  of
the Borden site data in Section 7.

    The variance of the log-hydraulic conductivity  can be  measured  in  situ or,
as more information for the behavior of different types of aquifer  materials
is accumulated, it can be estimated from  information about geologically
similar sites.

    The local dispersivities can be also  found from a  laboratory  test  of  soils
from the site or estimated based on the general properties of  the soil (Klotz,
et al. 1980).  It should, be noted that the effect of the local  dispersivity  on
the variance is significant and cannot be neglected.   This behavior is in
contrast to that of the macrodispersivity which is  not sensitive  to the value
of the local dispersivity.

    An important factor needed for the variance estimation is  the correlation
scales.  Analyses such as Sudicky's (1985a,b) show  ways to estimate their
values in situ.  It is very encouraging that the estimated correlation scales
give a consistent picture for the mean plume when used to  estimate  the
macrodispersivities and the concentration variance  as  a measure of  the
variation around the mean plume.  Gelhar  (1985) presents a collection  of  the
variances and correlation scales for log-hydraulic  conductivity or
log-transmissivity estimated either from  direct analysis of  measurements  or
through inverse modeling procedures (Hoeksema and Kitanidis, 1985).

    The predicted concentration variance  is an appropriate calibration target
for the numerical models.  Other sources  of error such as  measurement  errors
or discretization errors associated with  the numerical scheme should also be
considered, but typically those effects are not expected to  be  the  predominant
ones.  Therefore, if the root mean squared difference  between the observed and
simulated concentration fields is used as an index  of  the  model error, the
calibration of the numerical model could  be considered adequate when the  model
error and the error (standard deviation)  predicted  from the  stochastic theory
are of the same magnitude.


                                        67

-------
    Although the results of the comparison with  field data  are  encouraging and
indicate the workability of the stochastic theory  to predict  concentration
variance, there are a number of approximations and/or refinements  which  need
to be evaluated before the full range of applicability of the theory  can be
established.  These features are discussed in the  following paragraphs.

    The development of the unsteady concentration  case in Section  5,  invokes
an expansion of the concentration gradient around  the current time t.  Although
only the first two terras of the expansion have been retained, additional terms
could easily be included and evaluated numerically.  Since  these  terms are
multiplied by the time derivatives of the mean concentration, it  is
expected that they will have ajninimal effect on the variance except  near
the concentration peak where ac/ax^ i = 1,2,3 changes rapidly.   The  effects
of these unsteady terms near the peak need to be explicitly evaluated  in order
to understand their significance.

    The t > o> assumption for the integration of  Eq. (5.22)  is not  required.
Eq. (5.21) could be integrated for any finite time, which would result in an
additional factor [l-exp(-et)] multiplying the complex amplitude  in Eq.
(5.23).  From the form of Eq. (5.21), it is evident that the  asymptotic  result
will overestimate the variance for finite time and consequently will  usually
be conservative.  This finite time effect should be evaluated in  future  work.

    In a heterogeneous aquifer, the local dispersivities are  expected  to be
variable in space.  The effect of local variation  of dispersivities can  be
analyzed using an approach similar to that of Naff (1978).   This  effect  may be
of significance in determining an effective local  dispersivity  to  be  used in
the calculation of the concentration variance.

    The behavior of the concentration covariance along the  mean flow  direction
is somewhat puzzling from a physical point of view.  As is  shown  in Figure
5.5a, the longitudinal correlation scale of the  concentration variation
becomes very large.  This implies that regions of  high or low deviation  from
the mean concentration will persist over very large longitudinal  distances,
say up to 10 or 100 times the correlation scale  of ink.  Note that the degree
of longitudinal persistence increases as e=a/fc,  decreases.   Although
somewhat similar behavior seems to be indicated  by other methods  (Dagan  1984),
we feel that this behavior is questionable in terms of our  intuition  about the
physical system.  Intuitively we expect a' smaller  degree of longitudinal
persistence.  This feature requires additional detailed investigation  because
the concentration covariance function is a very  important factor  in the  design
of plume monitoring networks.  Preliminary results indicate that  the  high
persistence is caused by the distribution of the spectrum energy  at high wave
numbers.
                                       68

-------
                                  REFERENCES

1.  Ababou, R., D. McLaughlin, L.W. Gelhar and A.F.B. Torapson.   Numerical
    Simulation of Saturaced/Unsaturated Flow Fields in  Randomly  Heterogeneous
    Porous Media.  Proceedings of International Association  of Hydraulic
    Research Symposium on the Stochastic Approach  to Subsurface  Flow,
    Montvillargenne, France, June 1985.

2.  Aldama, A.A. Rodriguez.  Theory and applications of two-and  three-scale
    filtering approaches for turbulent flow simulation.   Ph.D. dissertation,
    Massachusetts Inst. of Tech., 397 pp., November 1985.

3.  Anderson, M.P.  Using Models to Simulate the Movement  of Contaminants
    Through Groundwater Flow Systems.  Critical Reviews in Environmental
    Controls, 9(2), pp. 97-156,  1979.

4.  Bakr, A.A.  Stochastic Analysis of the effect  of spatial variations of
    hydraulic conductivity of groundwater flow.  Ph.D.  dissertation,  New
    Mexico Institute of Mining and Technology, 244 p, 1976.

5.  Bakr, A.A., L.W. Gelhar, A.L. Gutjahr and J.R. MacMillan.  Stochastic
    Analysis of Spatial Variability in Subsurface  Flows 1. Comparison of One-
    and Three-Dimensional Flows, Vol. 14, No. 2, pp 263-271, April  1978.

6.  Beran, M.J.  Statistical Continuum Theories, Interscience, New  York, 1963.

7.  Betson, R.P., J.M. Boggs, S.C. Young, and L.W. Gelhar.   Design  of field
    experiment to investigate ground water saturated zone  transport processes,
    Electric Power Research Institute, Inc., Interim Report  RP 2485-05, 104
    pp, 1985.

8.  Chirlin, G.R. and G. Dagan.  Theoretical head  vartograms for steady flow
    in statistically homogeneous aquifers.  Water  Resources  Res. 16(6), pp
    1001-1015, 1980.

9.  Dagan, G.  Models of groundwater flow in statistically homogeneous porous
    formations, Water Resources  Res. 15(1), 47-63, 1979.

10. Dagan, G.  Stochastic Modeling of Groundwater  Flow  by Unconditional and
    Conditional Probabilities, part 1, Conditional Simulation and the direct
    problem.  Water Resour. Res., 18, pp. 813-833, August  1982a.

11. Dagan, G.  Stochastic Modeling of Groundwater  Flow  by Unconditional and
    Conditional Probabilities.   2.  The Solute Transport.  Water Resour. Res.,
    Vol.  18, No. 4., pp. 835-848, August 1982b.

12. Dagan G.  Solute Transport in heterogeneous porous  formations.   J. Fluid*
    Mechanics, 145, pp. 151-177, 1984.                                       -*•
                                       69

-------
13. Freeze, R.A.  A Stochastic-Conceptual Analysis of One-Dimensional
    Groundwater Flow in Nonunifonn Homogeneous Media.  Water  Resour. Res.,
    Vol. 11, No. 5, pp. 725-741, October 1975.

14. Freyberg, D.L., D.M. Mackay and J.A. Cherry.  Advection and  Dispersion  in
    an Experimental Groundwater Plume.  Proceedings of the Hydraulics  Division
    Specialty Conference in Frontiers in Hydraulics Engineering,  A.S.C.E.,
    pp.  36-41, July 1983.

15. Freyberg, D.L.  A natural gradient experiment on solute transport  in  a
    sand aquifer II.  Spatial moments and the advection and dispersion of
    non-reactive tracers.  Water Resourc. Res., (in submittal) 1985.

16. Gelhar, L.W.  Stochastic Analysis of Phreatic Aquifers.   Water  Resour.
    Res., Vol. 10, No. .3, pp. 539-545, 1974.
                      . i
17. Gelhar, L.W.  Effects of Hydraulic Conductivity Variations on Groundwater
    Flows.   Second International Symposium on Stochastic  Hydraulics, Lund,
    Sweden, in Hydraulic Problems solved by Stochastic Methods,  Water  Res.
    Publications, Fort Collins, Colorado, pp. 409-431, 1977.

18. Gelhar, L.W.  Stochastic Analysis of Solute Transport  in  Saturated and
    Unsaturated Porous Media,   (in press 1986).

19. Gelhar, L.W.  Stochastic Subsurface Hydrology: from Theory to
    Applications.  Proceedings of International Association of Hydraulic
    Research Symposium on the Stochastic Approach to Subsurface  Flow,
    Montvillargenne, France, June 1985.

20. Celhar, L.W.  Stochastic Analysis of Flow in Heterogeneous Porous  Media,
    in Fundamentals of Transport Phenomena in Porous Media, J. Bear and M.Y.,
    Corapcioglou, eds., Martinus Nijhoff Publishers, Dordrecht,  The
    Netherlands, 1984.

21. Gelhar, L.W., A.L. Gutjahr and R.L. Naff.  Stochastic  Analysis  of
    Macrodispersion in a Stratified Aquifer.  Water Resour. Res., Vol.  15,
    No.  6,  pp. 1387-1397, Dec.  1979.

22. Gelhar, L.W., A.L. Gutjahr and R.L. Naff.  Reply to Comments on
    Stochastic Analysis of Macrodispersion in a Stratified Aquifer. Water
    Resour. Res., 17, pp. 1739-1740, 1981.
                                                      >.
23. Gelhar, L.W. and A.L. Gutjahr.  Stochastic Solution of the One-Dimensional
    Convective-Dispersive Equation.  Report No. H-ll, New  Mexico Institute  of
    Mining and Technology, pp.  20, 1982.

24. Gelhar, L.W. and C.L. Axness.  Three-dimensional stochastic  analysis  of
    macrodispersion in aquifers.  Water Resour. Res., 19,  161-180,  1983.
                                        70

-------
25. Gelhar, L.W., A. Mantoglou, C. Welty and K.R.  Rehfeldt.   A Review of Field
    Scale Physical Solute Transport Processes in  Saturated and Unsaturated
    Porous Media, Research Project 2485-5,  EPRI EA-4190,  Electric Power
    Research Institute, Palo Alto, California,  116 pp., August 1985.

26. Gutjahr, A.L., L.W. Gelhar, A.A. Baler and J.R.  MacMillan.   Stochastic
    Analysis of Spatial Variability in Subsurface  Flows 2. Evaluation and
    Application, Water Resour. Res., Vol. 14, No.  5,  pp.  953-959,  October
    1978.

27. Gutjahr, A.L. and L.W. Gelhar.  Stochastic  Models of  Subsurface Flow:
    Infinite Versus Finite Domain and Stationarity.   Water Resour.  Res., Vol.
    17, No. 2, pp. 337-350, April 1981.

28. Gutjahr, A.L.., E.J. Bonano and R.M. Cranwell.   Treatment  of Parameter
    Uncertainties in Modeling Contaminant Transport  in Geologic Media:
    Stochastic Models vs. Deterministic Models  with Statistical Parameter
    Sampling.  Proceedings of International Association of Hydraulic  Research
    Symposium on  the Stochastic Approach to Subsurface Flow,  Montvillargenne,
    France, June  1985.

29. Gradshteyn, I.S. and  I.M. Ryzhik.  Table of Integrals, Series and
    Products, Academic Press, 1980.

30. Hinze, J.O.   Turbulence, McGraw-Hill, 1975.

31. Hoeksema, R.J. and P.K. Kitanidis.  An  Application of the  GeostatistxcaL
    Approach to the Inverse Problem in Two-Dimensional Groundwater Modeling.
    Water Resour. Res., Vol. 20 (7), pp. 1003-1020,'1984.

32. Hoeksema, R.J. and P.K. Kitanidis.  Analysis  of the spatial structure of
    properties of selected aquifiers, Water Resour.  Res.,  21,  563-572, 1985.

33. Hufschaiied, P. Estimation of  three-dimensional statistically anisotropic
    hyraulic conductivity field by means of single well pumping tests combined
    with flowmeter measurements.  Proceedings of  International Association of
    Hydraulic Research Symposium  on the Stochastic Approach  to Subsurface
    Flow, Montvillargenne, France, June 1985.

34. Kitanidis, P.K., and E.G. Vomvoris.  A  Geostatistical Approach to the
    Inverse Problem in Groundwater Modeling (Steady State) and One-Dimensional
    Simulations.  Water Resour. Res., Vol.  14(3),  pp.  677-690, 1983.

35. Klotz, D., K.-P. Seiler, H. Moser and F. Neumaier. Dispersivity  and
    velocity relationship from laboratory and field experiments.  Journal of
    Hydrology, 45, 169-184, 1980.
                                        71

-------
36. LeBlanc, D.R.,  Ed., Movement and  fate of  solutes  in  a plume of sewage
    contaminated groundwater, Cape Cod, Massachusetts: U.S.  Geological Survey
    Toxic Waste Groundwater Contamination Program  Papers presented at the
    Toxic Waste Technical Meeting, Tucson, Arizona, March 20-22,  1984,
    U.S.G.S. Open File Report 34-475,  1984.

37. Lumley, J.L. and H.A. Panofsky.   The Structure of  Atmospheric Turbulence,
    John Wiley and Sons, Inc., New York, 1964.

38. Matheron, G. and G. de Marsily.   Is transport  in  porous  media always
    diffusive?  A counterexample.  Water Resour. Res.,  16,  pp.  901-917,  1980.

39. McLaughlin, D.B.  A Distributed State Space Approach for Evaluating  the
    Accuracy of Groundwater Predictions, Ph.D. Thesis, Dept. of Civil
    Engineering, Princeton University, Princeton,  1985.

40. Mizell, S.A., A.L. Gutjahr and L.W. Gelhar.  Stochastic  Analysis of
    Spatial Variability in Two-Diraensional Steady  Groundwater Flow Assuming
    Stationary and Nonstationary Heads.  Water Resour. Res., Vol. 18, No. 4,
    pp. 1053-1067,  August 1982.

41. Moltyaner, G.L. Stochastic versus  deterministic:   A  case study.
    Proceedings of Intern. Assoc. of  Hydr. Res. Symposium on the Stochastic
    Approach to Subsurface Flow, Montvillargenne,  France,  June  1985.

42. Naff, R.L.  A Continuum Approach  to the Study  and  Determination of Field
    Longitudinal Dispersion Coefficients.  Ph.D. thesis, New Mexico Institute
    of Mining and Technology, Socorro, New Mexico,  pp. 176,  1973.

43. Netter, J. and W. Wasserman.  Applied linear statistical models, R.  D.
    Irwin, Inc., Hornewood, Illinois,  842 p.,  1974.

44. Schwartz, F.W.  >4acroscopic Dispersion in  Porous  Media:   The Controlling
    Factors.  Water Resour. Res., Vol. 13, No. 4,  pp.  743-752,  August 1977.

45. Smith, L., and R.A. Freeze.  Stochastic Analysis  of  Steady  State
    Groundwater Flow in a Bounded Domain. 1.   One-Dimensional Simulations.
    Water Resour. Res., Vol.  15, No.  3, pp. 521-528,  June 1979a.

46. Smith, L., and R.A. Freeze.  Stochastic Analysis  of  Steady  State
    Groundwater Flow in a Bounded Domain. 2.   Two-Dimensional Simulations.
    Water Resour. Res., Vol.  15, No.  6, pp. 1543-1559, Dec.  1979b.

47. Smith, L. and F.W. Schwartz.  Mass Transport 3.  Role of  Hydraulic
    Conductivity Data in Prediction.   Water Resour.  Res., Vol.  17, No. 5, pp.
    1463-1479, October 1981.
                                       72

-------
48. Sudicky, E.A.  Spatial Variability of Hydraulic  Conductivity  at  the Borden
    Tracer Test Site.  Proceedings of Intern.  Assoc. of  Hydr.  Res.  Symposium
    on the Stochastic Approach to Subsurface Flow, Montvillargenne,  France,
    June, 1985a.

49. Sudicky, E.A.  A Natural-Gradient Experiment on  Solute  Transport in a Sand
    Aquifer: Spatial Variability of Hydraulic  Conductivity  and  its  Role in the
    Dispersion Process.  Water Resour. Res., (Submitted,  1985b).

50. Taylor, G.I.  Diffusion by Continuous Movements.  Proc.  of  London Math.
    Soc., Ser. 2, Vol. XX, pp. 196-212,  1921.

51. Vanmarke, E.  Random Fields: Analysis and  Synthesis.  The MIT Press,
    Cambridge, 1983.

52. Vomvoris, E.  Groundwater Parameter  Estimation:  A Geostatistical
    Approach.  M.S.  thesis, Univ. of  Iowa,  p.  188, July  1982.

53. Vomvoris, E.G. and L.W. Gelhar.   Stochastic  Analysis  of Concentration
    Variability in Transport in Aquifers.   Proceedings of the Intern. Assoc.
    of Hydr. Res. Symposium on the Stochastic  Approach to Subsurface Flow,
    Montvillargenne, France, June 1985.
                                        73

-------
                                   APPENDIX A

           THE MATHEMATICS OF THE  APPROXIMATE  ANALYTICAL EVALUATIONS
    This appendix describes  the mathematical  solutions used for the
approximate analytical results presented  in Table 5.1.  The following symbols
will be used repeatedly to provide  non-dimensional quantities:

                                       2     2     2^2
    ul = klV U2 = V2' U3 = *3*3'  U = Ul   + "2  + U3
         aL       aT       Ul         ^3         ^3
    .2     2  2  . -2
    u  = p. u.  + U

    ,,2    2 2   ,,2    ,-,2      2  2  2  .  -2     "2      222.   f 2
    V  =eu  +U  ,  V  =p1eu  +U   ,   V   =p1eu  +
-------
Changing Che variables  with  Che  aid of Eq. (A.I)



                                                2




    III -  /" --^O   22*   2*4  (1 - P? I?'2 duldu2du3            (A'3)
          -»(l+u)    p.u.+eu           u



Introducing the variable  u = u,/e,  Eq. (A. 3) can be written as:
                    2               1               2  2 2
                n +vV      2 2 2 +  2-4       Pl
                (1 +  V  )     P  u e  + e V
Taking the limit  e +  0,


    ..2     2^2    ..2
    V  ->• u?  + u,   = U
    (. - Pl*       >2  ,  i



      22-4      22.   2-4
    PL U  +V  +P1U  +KU


and hence, 1^ takes  the  form




                        "
                                                                         (A.4)
                     -oo    (1 + U)   P i  U  + 1C U



Integrating over  u,  after changing variables x = Plu,





  /  ~~n     w  du  = Z"2 /   2 .  2-4dx = p~ 2 i ~n =  p~ ~f^       (A>b)
-co   p,U+, u3 = U sin ij>,




    I,,  = — —    f  f   	5—z-   —=—z	j	^— U  dU d$        (A.8)

          e pl  0 0    (I + U )    U (p2 cos % + sin",),)
                                          75

-------
Evaluating separately, for each variable,  [Gradshteyn  &  Ryzhik,  1980,
3.642(3)]

     o   U+u2)3      4             '                                  (A.9)
       2ir        .
            o
     0    p» cos 4> + sin §       P2

which result after substitution  in Eq.  (A. 8)  and  Eq.  (A. 2),
           2

             .2
                3
°f -2 2 1  J_ JL                                              (A.LO)
Variance; vertical  isotropy - !„„.


    T22 = E I22



    I°2 = I22 («L=  a, - 0) -    /" —4-3   %  duldu2du3             (A. 11)
                             -=>   C 1  T  u  ;    u

Changing to spherical coordinates  (u, p,  9),  with p2  = 1 ,


    U» = U COScj)
    u2 = u
    u  = u
                       2   2                     2
                  + sin             0  (1 - R x
                                                                        (A. 12
                                                     sn   « do         (A. 13)
     22   0   (1 + uV      0   0    (p^cos^d,  + sin


The  three integrals  can  be  evaluated  separately,


     ;" - ?L— du = Ig-                                              (A- I4a)
     0  (1 + u2)3        16



       /  7rCos2e  d9 -  «                                                  (A.Ub)
     0
                                         76

-------
    =  -4-  + -4 +  £ *n JLJLj£%                               
       2Pl2    R2    2   R2  R    R - 2R2


    The combination of Eqs. (A.14a), (A.14b) and  (A.14c) gives  the  result
presented in Table 5.1.  Note that when A1=a2=A3=4» eq.  (A.14c)  becomes


     f  sin3* d* = ^                                                    (A.15)
    0              J

and, then,
                           Y

For T33 the only change occurs at Eq.  (14b), which  becomes

      »   T
     / sin 9 de = ir                                                     (A. 16b)
    0
Thus, T22 = T33 as claimed in Table  5.1.



Covariance; isotropic case - T. .(5, 0,0)


From eq. (5.12),
                                                                        (A. 17)
The spectrum Scc(lc)  is even with  respect  to  k. = (kj.kj.kj)  for the case
considered here.  The exponential  can  be  expressed  as,
    ei-'-S-= cos(k.£)
and, since the sin(.)  is odd  it will  cancel  out.   Therefore,  the Tn(£) is
given from,


                                             u2

    Tu(l> - f-   /"cosCu-i')      1  2.4  (1  -  ^-)2Sff(u)  du              (A.19)
             E-oa            U  + £ U       U



Introducing u = VE*  and  takin8  the  limit e * °'


             C1      €2       £3            "424
                                        77

-------
          2    ,
    (1 -^ e >  * 1           sff^> -  S£f(0,u2,u3>                   (A.20)
         u

The uȣ argument in the cosine does not allow for analytical  integration
unless only one of the Si's, or, equivalent ly, uj/s, is  retained.   Hence,
for £= (5,0,0), Eqs. (A.19), (A.20) and (5.32)  for SfE(k),
           = E    / cos(euc)  224  -  ~2~3 du du2du3              U<21)
                 -B>          u + K U   (1+U)
Integrating over u first,
                                     'KU ^                             (A.22)
      / cos(euc) -5	o-T du = -^T e
                 u  + K V      KIT

The remaining integral becomes,
                        CO     2

                   eK -co         (I + U )

Converting to polar coordinates, u2 = U cos$, u3 = U sin


    T,,(5,0,0) -E-J  /*d4  /V8U     "      dU                      (A.24)
     11            EK 0      0        (1 + U  )

The second integral can be evaluated analytically in terms  of  the  Exponential
integral [Gradshteyn and Ryzhik, 3.353(4)]
                      2
    TU (5,0,0) =E — £[ 1-S-B  eBEi(-B)]                        (A.25)

Note that the term in the brackets equals 1 when 3=0, which  verifies  the
variance relationship.

For S « 1,

             B) = C + InB - B + •••                                   (A.26)

         eB = 1 + g + ...

where C = 0.577215, Euler's constant  [Gradshteyn & Ryzhik,  8.214  (1)]  and
the term in the brackets becomes

         [ 1 - B - B2 (C + Infi) +  ...]                                  (A.27)

For B » 1.

         eBEi(-B) = - + -^ - \ +  ....                                  (A-28)
                    B   B    6
                                        78

-------
[Gradshteyn & Ryzhik, 8.215] and  the  term  in the brackets becomes


         i!  i                                                          ('


Covariance; isotropic case - T..(0,5,0)

In this case  Eqs. (A.17)  through  (A.20)  are the same.  With £ = (0,5,0),
Eq. (A.21)  becomes
                        09                        2
    Tu(0,5,0) = E -   /  cos(u2?) ——^y£	—1~3  du du2 du3     (A.30)

From Eq. (A.22) the  integral over u,  for 3 = 0, ts known.  Changing to polar
coordinates,*                               . i
                          «>   2ir
    T  ,(0,E,0) = E i-    f   f    cos(Ucos^) 0 du d                     (A-31)
     11            CK   0J  0     (L  + U2)3

The integral  over ,)>  is  given in  Gradshceyn & Ryzhik, [3.715(18)],

       2n
       J  cos(cos,bUc) d^, = 2 TT  J  (U;;)                                    (A. 32)
     0

and, finally,

     j"j (U?) 	!L_ du  = I ? ZK (?)                                  (A.33)


from [Gradshteyn & Ryzhik 6.565(4), 8.486(16)] where Jo(.) and K2(.)  are  a
Bessel function of first  kind, zero order and a modified Bessel function  of
second kind,  second  order,  respectively.  The coefficient T22(0,5,0)  is given
from,
                   2
                                     (?)                                 (A.34)
                  Y

Note that  e2IC(i-)  =  2,  at  5 = 0,  verifying the variance result.
                                                                          t *•
For the case £ =  (0,0,g)  the only change is in eq. (A.32) which becomes,'"-

       2*
      /  cos(U? sin,),)  44  = 2n J  (^U)                                    (A.35)
    0

[Gradshteyn &  Ryzhik,  3.715(9)], therefore,

    T11(0,g,0) =  Tu(0,0,?)                                             (A.36)


                                         79

-------
Covariance; isotropic case  -  !„({;,0,0)



    As in  the variance  case,  the coefficient T22 can be evaluated  when  e  = 0.

The same arguments for  the  exponential, Eq. (A.18), hold.  Therefore,  for

e = 0 and  a combination of  Eqs.  (A.ll), (A.17) and (A.18);



                                            2
                 «               1        u2

    T.-Cj) = E   / cosU £') 	Y~3    ~2~  di^du^u-j                 (A.37)

              -w            (1 + u )      u





Changing to spherical coordinates and for g_ = (g,0,0)

                                           2



                 00   "              !       u       32
    T00(£) = E   f  / /  cos  (u^cosij)) 	=—=• sin ^cos 9 de d du       (A.38)

     22        0 0  0                 U+uV



The integral over u  equals,



                      u2             -lal            2
     f cos(u£COSrji)	5—y du = yV- e I  I [ 1 +  a  - a  ]                 (A. 39)

    o             (i +  u r                    '  '


where a =  ^cose. In addition,  the integral over 9 equals IT,  from

Eq. (A.lAb).  Hence



    T.,(6,0,0) = E yj-   |"V|a|[  1 + lal - a2] sin3* d,j,                  (A.40)
     Li            10 Q             II



Changing variables to COS(J»  =  x,  and using the following equalities,



     /  bx.,,    2N  .       b,2     2 ,    ,1   2 >
     f  e  (l-x)dx   =e (-^	y) -  (r	r)

    0                         b     b          b



     t1  bx,.    2N   .      b,2     6  .  6 -. ,  ,1    f> *                  ,. AM
     J  e  (1 -  x )x dx = e (-=•	r- + -r) + (-J	r>                  QA.41)

    0                         b     b    b      b    b


     ,l  bx,.    2.  2.      b,2     10 .  24   24.   f-2   24
     J  e  (l-x)xdx = e (—r	r- + —7-	r)  + V—T  + —?)

    0                         bbbbbb^^


the expression for T22(5,0,0,)  becomes,


                   2




                  Y          5        SCC

                           2    3

Expanding, e ^ = 1 - 5  + -jjy - -iy + ...,  shows  that the term  in  the brackets

becomes for 5=0,
                                         80

-------
               [ TI  1                                                   (A'43)
verifyiag Che variance  result.

Covar lance; isotropic case - T22(0,c,0)


The general Eq. (A. 37)  is  the  same.   The conversion to spherical coordinates
is introduced in  the following way,

             U2 = UCOSij)
             Uj = usin 0, approaches  the  value
1/3 ^, verifying  thus the asymptotic result for the variance.
                                        81

-------
Covariance; isotropic  case  -  T33(0,0,£)

The general Eq.  (A.37)  is the same.   If  the following conversion  to  spherical
coordinates is introduced,

    u3 = u cos
    ut = u sinifcose                                                      (A.49)
    u2 = u sinsin9

the expression for T22  (0,0,5)  becomes,

                    «   TI  2ir                U2       32
    T  (0,0,£) = E  f   f  f  cos (u^coscjj)     ~ « sin ijisin 0 d9d(((du      (A.50)
     22           000                 (1+u )
Since

     2ir   -        2ir   2
    f  sin e de  = / cos e de   = IT                                       (A.51)
   0             0

It can be easily seen  that

    T22(0,0,0 = T22(g,0,0)                                             (A.52)


Covariance; isotropic  case  -  T33(5,0,0)

The general equation is:'
                                               2
                     oa                      U-
    T^^(t.O.O) =4   f   cos  (u,c)  	T-7—2  du1du2du3              (A.53)
                    — o»
                                   (1 + U'
     u2 = u sincose                                                (A. 54)
Changing  to  spherical  coordinates,

          u1  = u  cose
          u2  = u  sin^cose
          u3  = u  sinij)Sin0

Eq.  (A. 53) becomes  identical to Eq. (A. 50).  Hence,

     T33U,0,0) = T22(5,0,0)                                              (A.55)

Covariance;  isotropic  case - T33(0,{;,0)

The  general  equation  is,
                                             2
                                           U_
T  (0,5,0) -    /  cos(u  C) - —r^ -
                                         82
                                               du du du                  (A.56).

-------
Changing to spherical coordinates as  in  Eq.  (A. 44),  it  can be seen that




         T33(0,c,0) = T33(?J0,0)                                        (A.57)




Covariance; isotropic case - T33(0,0,£)




The general equation is,




                                            2
    T33(0,0>5) =    / cos(u3c) - T^~ 2  du!du2du3                 (A<58)








Changing the spherical coordinates  as  in  Eq.  (A. 49),





                 r.  =•  TT  27T                  2
                                                  cos  •{•sin^ded^du       (A. 59)
                                         83

-------
                                     APPENDIX  B

                          THE INTEGRATION OVER WAVE-NUMBER

      The spectrum of the concentration can  be written as  follows,
    Eq. (5.10) and (5.11),


                   k?4J     S  (k)                  kk        kk
     •«<» - * 7^3  -^ I W-VV  ^> V ^'l       »•»


           1  4  123
 where E = -=-r- -r- - 5 — , and summation over  j,£,ra,n in the branches is assumed.

           T      *
      Note that introduction of spherical coordinates  makes  the terms inside
 the brackets independent of k.  Hence, when one  evaluates the covariance,
              00
     R  <£> - T exp(lk.e) S  (k) dk                                       (B.2)
      cc •*•    '       ~~ •*•   cc —   —
             — oo

 transforming to spherical coordinates and evaluating  the  integral
 over wave number k, results in the following:
              n 2-       »                22
     R  <£> - / / l(B,u) /exp(ikcosfj) - **  •>      o   9  •>  dkd8d"        (B'3)
      cc      00        0            (l+k-A^)J     B'+a-k-
 where A, B depend on the angular coordinates  from  the used  transformation
 and 1(9 ,u) contains the remaining terms in  Eq.  (B.2).

      Note that due to symmetry of Scc(k_) , i.e.,

                Scc(k1 ,k2,k3) = Scc(-k1,-k2 ,-k3)

 expanding the exponent, exp(iD) = cosD + isinD,  in the integral
 Eq. (B.2) cancels the imaginary part.

      The spherical coordinate transformation  is  chosen so that k«£_ = k 5 cosg ,
 i.e., depends only in two parameters of integration k,g.   Appendix C describes
 the details of the used transformation which  is  similar  to  the one used by
 Bakr et al. (1978) to obtain analytical results  for the  head field covariance.
                 "f*—
 Denote,

      B = 
-------
The integrals can be nondiraensionalized using the parameters


      "1= ^ ' "2- TJ ' (p3= 1J«   e = "I; '  K = ^   ' Ut = Mi    (B-5)

 Then converting to spherical coordinates and denoting


       E2= e2[P2B2-hc(p2C2+D2)]2                                         (B.6a)


       A*= -T (aij"juj]2+ -T l'yPjuj]2+ l-sj'j"/                   (B'6b)
            P!              P2

       [a  ]. = Transformation matrix between the hydraulic conductivity
          -1    principal axes and the axes aligned to the mean flow direction

                          C,     = 5/i                                (B.6c)
                                                                        (B.6d)
             232
            Y        IT
the integrals become

                    2 2 v 2lF
      R  (£1X1111) = oc£0 f f PF. .(B>(i);xi'l')sinBM djjdu G.G.                (B.7a)
       cc           f 3 0 0   1J                      1  J

                                22
               M =  fcos(Ru)   A "   .   , / ,  .  du                     (B.7b)
                   0        (l+AZu  )J  pfB^VE-
Eq. (B.7) is the same as Eq.  (5.33c)  if one denotes  the  three-dimensional
integral as TIJ(.).  The coefficient  Fij(.) in  the  integrand  is  given
from:
                     22222                                        (B'8)
                L =   B^+C +0
The integral over u in Eq.  (B.7b)  can  be estimated  analytically.

     From partial fraction  expansion,  for A2*0,  E2#pjA2B2,
                                       85

-------
     k2A2
223  p2222
(l+kA)     B+Ek
                               k2-
                                                                 (k2-Pl)3
                                                                       (B.9a)
                                                                       (B'9b)
                               - P
                                                               - P
11           3
     (prp2)3
                                 '  A
                                          I3
                                                       »  A
with the aid of the following integration formulas [Gradshteyn and Ryzhik,

1980, 3.737]
                dx =
                     16
                                     3   e
                                               -labl
          dx .
                      4 b
                                    abe
                                          -ab
    0  (bN-x')J         2|b|


Eq. (7b) becomes:
     Case 1:  A2* 0, E2* A2B2Pl2 ,8*0.
         E A
                 A   (    + 3
                      A
                                               exp  •
                                                                       (B-U)
                                        86

-------
If B = 0, Eq. (B.6) becomes,
     M-      fe <    + 3     +3) SXP  I"    ]                           (B>12)
          E       A


If E2 = p 2A2B2 ,  Eq. (B.6) becomes


           »                     2
     M =  /cos (kR) — J-r-      k    4  dk
         0            E-A*  (-L. * kV
         E
Finally, if A = 0,

               M = 0                                                    (B.14)

The remaining eq. (B.7a)  Ls solved numerically  using  a  Romberg integration
scheme for the integration over 3 and  the  trapezoidal rule  for the resulting
one-dimensional integral.
                                        87

-------
                                  APPENDIX C

                      TRANSFORMATION TO POLAR COORDINATES
    The coordinate system (x°, y°, z°) in Figure C-l represents  the
original coordinate system to which (x1,x2,x3) or  (k1,k2,k3)  are aligned.   The
objective of the transformation to the polar coordinates  is  to express  the dot
product jt .§£ in the exponential in Eq. (5.12) in  terms of k and 0,  i.e.,  two
parameters only.  The transformation will be performed in three  steps.


Step 1   Create a Cartesian system of axes, (gi,e.l)i where  e belongs  to the

         plane (xo,yo) and |j is on vector |.

The vectors fps.Jt can be expressed in terras of the angles Xi'|»  (Figure  C-l).
                                T
|j = [ cosy, sinxcosiji, sinxsm^] , where the superscript  "T"  stands  for

transpose and gj has unit length.
E = [e , ev, 0], since it belongs in (xQ,yo).  Invoking its  orthogonality  with

respect to £, E  • 5 = 0, or,
    e
                            n
                   , cosy, Oj
                                       1
                                 (L - sin
             f
Finally, for j, from orthogonality

    t  • e = 0    and      t  • 1  = 0

and the following expression can be obtained:
                                                I/0
         sin
                2    2
         1 - sin xcos
                          (1 - sin
•*)
Note, as a verification,  that when  5  ||  x  ,  i.e.,  x  =  °>
                 1  0  0
                 0  1  0
                 0  0  1
    [I.S.I] •

For notation purposes, denote,
                 x    y    z

                 "    8   8
                 •y    v    V
                 'x   'y  'z
                                        88

-------
Step 2  Express £,  in  terms  of polar coordinates (k,g,u)  in  the  system (|,e,J)


                                          ip      i   i   i ip
    it  =  [k cosg, k singcosu,  k singsinu]  = [k.,k2>k_]



where, for the  integration purposes, B:0-nr, u:0->-2Ti, k:0+M.

Step 3  Express  [kj1,k2',k3']T in terms of [kj.k2.k3].  This operation is
        given  from:
kl
k2
k3

=

ax ay az
Sx ey Sz
V V V
'x 'y 'z



V
k2'
k3'
It can be easily  verified that jt • £ = kgcosg.
 T       T
JL  I - *'  I
0

0
                                               kgcosf]
                                         89

-------
Figure C.I  Schematic graph of the coordinate transformation
            geometry
                            90

-------
                                  APPENDIX D

              MEAN CONCENTRATIONS FITTED AT THE BORDEN SITE DATA


    For Che longitudinal section the fitted curve had the  following  form,


         c(x) - K • exp [ - (X " X)  ]                                  (D.I)

                             AV

Through a manual trial and error adjustment of the parameters,  the values
selected for Figure 7.5 are:

         K = 319 ag/ji
         AL = 0.34 m

         x = 17.83 m

For the vertical sections the fitted curves had the  following  form:


                           (z - Z)2
         c(z) = X  exp [ -
                               (38.3)
where 38.3 m = x, the longitudinal coordinate of  the  center  of  mass  of  the
plume, as given in Sudicky (1985).  Through a similar manual trial and  error
adjustment of the parameters, the values selected  for Figures 7.6(a)  through
7.6(1) are shown in Table D.I.
                                   TABLE D.I
            PARAMETERS USED TO FIT MEAN CONCENTRATION  AT  THE  WELLS
x (m)
35
38.3
40.3
42.3
44.3
46.22
48.3
y (m)
12
16
17.94
20
22
24
25.95
Kv 
-------