WATER RESOURCES RESEARCH, VOL. 45, W00B15, doi:10.

1029/2007WR006799, 2009

Aleatoric and epistemic uncertainty in groundwater flow

and transport simulation
James L. Ross,1 Metin M. Ozbek,2 and George F. Pinder1
Received 7 December 2007; revised 2 March 2009; accepted 23 March 2009; published 16 May 2009.
[1] The characterization of aleatory hydrogeological parameter uncertainty has
traditionally been accomplished using probability theory. However, when consideration is
given to epistemic as well as aleatory uncertainty, probability theory is not necessarily
appropriate. This is especially the case where expert opinion is regarded as a suitable
source of information. When experts opine upon the uncertainty of a parameter value, both
aleatoric and epistemic uncertainties are introduced and must be modeled appropriately.
A novel approach to expert-provided parameter uncertainty characterization can be
defined that bridges an historical gap between probability theory and fuzzy set theory.
Herein, a random set, a generalization of a random variable is employed to formalize
expert knowledge, and fuzzy sets are used to propagate this uncertainty to model estimates
of contaminant transport. The resultant random set-based concentration estimates are
shown to be more general than the corresponding random variable estimates. In some
cases, the random set-based results are shown as upper and lower probabilities that bound
the corresponding random variable’s cumulative distribution function.
Citation: Ross, J. L., M. M. Ozbek, and G. F. Pinder (2009), Aleatoric and epistemic uncertainty in groundwater flow and transport
simulation, Water Resour. Res., 45, W00B15, doi:10.1029/2007WR006799.

1. Introduction
[2] Uncertainty in groundwater flow and transport mod-
eling comes in two forms: aleatory and epistemic. Such
distinctions in uncertainty are most often identified in risk
assessment and reliability engineering [Helton et al., 2000a,
2000b, 2004; Hofer et al., 2002; Helton and Oberkampf,
2004; Oberkampf et al., 2004]; and only recently have these
distinctions been identified in hydrogeological applications
[Srinivasan et al., 2007]. Aleatory uncertainty, also called
stochastic or variable uncertainty, refers to uncertainty that
cannot be reduced by more exhaustive measurements or a
better model. Epistemic uncertainty, or subjective uncer-
tainty, on the other hand, refers to uncertainty that can be
reduced.
[3] Despite these apparent distinctions in uncertainty,
probability theory alone has traditionally been used to
characterize both forms of uncertainty in engineering appli-
cations [Apostolakis, 1990; Helton et al., 2004]. While it is
commonly accepted that probability theory is ideal for the
characterization of aleatory uncertainty [Ganoulis, 1996],
the facility with which probability theory effectively cap-
tures epistemic uncertainty has been called into question
[O’Hagan and Oakley, 2004], especially given the intro-
duction of a number of alternative methods of epistemic
uncertainty characterization [Choquet, 1954; Zadeh, 1965,
1978; Shafer, 1976].
1
Center for Groundwater Remediation Design, School of Engineering,
University of Vermont, Burlington, Vermont, USA.
2
Environ International Corporation, Princeton, New Jersey, USA.
Copyright 2009 by the American Geophysical Union.
0043-1397/09/2007WR006799
[4] One such method, fuzzy set theory, has failed to gain
acceptance in engineering, much less hydrogeological,
applications. A possible reason for this is the necessary
paradigm shift one must make in order to apply fuzzy set
theory to uncertainty characterization.
[5] Random set theory [Zadeh, 1965] provides an intui-
tive means for both epistemic and aleatory uncertainty
characterization. Whereas probability theory’s basic tool
for uncertainty characterization is the probability density
function (PDF), discrete random set theory is predicated
upon the assignment of probabilities to intervals, rather than
points, as with discrete PDFs. As such, a random set is a
generalization of a random variable, since intervals are more
imprecise than point values. The use of random set theory is
more appropriate for representation of subjective knowledge
because it does not rely upon means, variances and prob-
abilistic models, which are inconsistent with the nature of
human thought and discourse.
[6] Random set theory [Helton and Oberkampf, 2004;
Joslyn and Kreinovich, 2005], however, is a general
approach to subjective knowledge characterization, and
in addition, random sets can be transformed into fuzzy
sets [Joslyn and Booker, 2004; Joslyn and Ferson, 2004]
with little difficulty. As will be shown, the transformation
from random sets to fuzzy sets facilitates the efficient
solution of groundwater flow and transport model equations
characterized by uncertainty.
[7] Though a few applications of fuzzy set theory to
expert knowledge characterization in hydrogeologic appli-
cations have been published [Bardossy et al., 1989, 1990a,
1990b, 1990c; Bagtzoglou et al., 1996; Dou et al., 1995,
1997a, 1997b, 1999; Fang and Chen, 1997; Demmico and
Klir, 2004; Guan and Aral, 2004; Ozbek and Pinder, 2006;
Ross et al., 2006, 2007, 2008], fuzzy sets are not a

W00B15 1 of 10
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION
Figure 1. Inversely estimated hydraulic conductivity field; different numbers identify fields with
distinct hydraulic conductivity values. Gray ovals represent contaminant source locations. Dots denoted
as A and B are pumping wells.
mainstream method of uncertainty characterization in
hydrogeology. Consequently, an explanation of the distinc-
tions between probability and fuzzy set theories and alea-
toric and epistemic uncertainties in the characterization of
hydrogeological uncertainty is needed.
[8] The purpose of this paper is to frame uncertain
hydraulic conductivity information in terms of aleatory
and epistemic uncertainty, to show how it is related to
random set theory, and to demonstrate how it can be used
in groundwater flow and transport modeling. To this end, we
present the random set characterization of hydraulic conduc-
tivity using both uncertainty types and the corresponding
simultaneous use of random set, probability and fuzzy set
theories. We also describe how to propagate both types of
uncertainty in model estimates of concentration. In doing
so, we provide a reasonable method of uncertainty char-
acterization that is compatible with both probability and
fuzzy set theories.
2. Problem Statement
2.1. Site Information
[9] In groundwater modeling problems where hydraulic
conductivity measurements are few, a hydraulic conductivity
field is often assumed to be composed of a few large
subdomains of equal hydraulic conductivity, like the simpli-
fied representation of the Woburn, Massachusetts site pre-
sented in Figure 1, which we will use as our illustrative
example problem. A small number of hydraulic conductivity
measurements are available in each of these hydraulic
conductivity subdomains. Total correlation is assumed with-
in boundaries and zero correlation is assumed between the
units. Constant head conditions are specified on the left
(22.9 m) and right (38.1 m) boundaries, and no-flow
boundary conditions are specified along the top and bottom
of the domain. Contaminant sources (gray ovals in Figure 1)
are located in formations four and five at concentrations of
2000 ppb and 1500 ppb, respectively. Finally, two pumping
wells are placed in formation two. Wells A and B pump at
8.19
104 m3/s and 4.91
104 m3/s, respectively.
2 of 10
W00B15
[10] The Princeton Transport Code (PTC), a three-space
dimensional finite element simulator was employed to
model groundwater flow and transport, with a mesh density
value of 15.2 m everywhere but at the well locations, where
the density is increased to 3.0 m. Dispersivity, storativity
and porosity values are defined as 0.3 m, 0.0001, and 0.2,
respectively, throughout the domain. These are the default
PTC values for the parameters, and were deemed adequate
as the purpose of the study was the novel characterization of
hydraulic conductivity uncertainty and model estimates of
concentration. Though the low dispersivity value in con-
junction with the mesh density suggests a high Peclet
number, and, as a result the possibility of significant
numerical errors in the finite element transport model,
automatic upstream weighting of the convection term
adjusts for small dispersivities.
2.2. Traditional Approach (Confidence Intervals)
[11] For a given subdomain, a representative random
variable hydraulic conductivity value can be constructed
from the mean and variance of that subdomain’s measure-
ment data set. Because these measurements are themselves
inherently uncertain owing to measurement and inverse
model uncertainty, both the mean and variance of the
measurement data set may not be representative of their
true values. An appropriate expert familiar with the hydro-
geology of the area may be asked to provide some measure
of the uncertainty in the form of a 95% confidence interval.
Given a mean value, the assumption of hydraulic conduc-
tivity’s lognormality and this 95% confidence interval a
PDF defining the hydraulic conductivity random variable
can be constructed. Such an approach to both epistemic and
aleatoric uncertainty characterization is predicated strictly
upon probability theory. The two sources of uncertainty,
natural randomness and expert knowledge, are not distin-
guishable when both are built into a single probability
distribution.
[12] The mean hydraulic conductivity values associated
with the domain in Figure 1 are provided in Table 1, along
with the expert provided confidence intervals for the five
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION
Table 1. Hydraulic Conductivity Random Variable Properties for
the Domain in Figure 1
Unit Number Mean LnK (m/s) Confidence Interval Variance
1 25.6 [27.4, 23.8] 0.83
2 18.8 [18.8, 18.8] 0
3 19.2 [20.9, 17.4] 0.79
4 15.0 [17.2, 12.8] 1.26
5 11.9 [11.9, 11.9] 0 attempt to capture both the measurement error as well as the
formations and the resulting calculated variances. In this
case, the expert possessed an awareness of the devices used
to measure hydraulic conductivity at the various locations as
well as an implicit understanding of the subdomain-wise
homogeneity throughout the site. On the basis of this
background knowledge, the intervals for three of the for-
mations were specified to approximate 2 orders of magni-
tude variation in hydraulic conductivity. For simplicity, the
remaining two formations were assigned zero variation.
[13] Consider some of the limitations of the pure proba-
bilistic approach. Aside from blurring the two uncertainty
sources (aleatory and epistemic) into a single probability
distribution, the probabilistic form of the model estimates of
concentration is significantly dependent upon the certainty
with which the expert can define 95% confidence, a rather
abstract notion, and the appropriateness of the lognormality
assumption. In fact, the longstanding assumption of lognor-
mality for hydraulic conductivity may not be correct in all
cases [Ricciardi et al., 1998; Mathon et al., 2009] The true
random variable may actually be best defined using an
alternative probability function.
[ 14 ] Using rank-ordered Latin hypercube sampling
[Zhang and Pinder, 2003] model estimates of concentration
were determined using the flow and transport simulator. The
uncertainty in these concentration estimates is sensitive to
relatively small variations in estimated hydraulic conduc-
tivity intervals. Note that 2 orders of magnitude change in a
confidence interval is considered small relative to the range
of hydraulic conductivity values one can encounter in the
field, which, according to Domenico and Schwartz [1990]
can range over 11 orders of magnitude from clay to gravel.
[15] The results of the three cases of varying hydraulic
conductivity uncertainty presented in Table 2 are plotted in
Figure 2 as estimates of discrete concentration random
variables. The cumulative distribution functions plotted
are constructed from the realizations of concentration esti-
mates that result from the application of Latin hypercube
sampling. The steeper distributions (case 2, black squares)
result from the smaller hydraulic conductivity confidence
intervals in Table 3. The longer, wider distributions (case 3,
hollow circles) are the random variables resulting from a
less certain expert, who provided wider confidence inter-
vals. Moderate uncertainty (case 1, black circles) produces
cumulative distributions situated between the two extreme
cases. Thus, the opining expert’s certainty regarding model
input parameters can produce larger changes in the model
output random variables.
3. Random Set Uncertainty
[16] An approximation of a hydraulic conductivity sub-
domain’s uncertain hydraulic conductivity is provided by
3 of 10
W00B15
the discrete CDF calculated from the available measure-
ments in that subdomain. However, as stated above, these
individual measurements are themselves uncertain. Intui-
tively, then, this measurement error warrants characteriza-
tion before the subdomain’s hydraulic conductivity random
variable can be defined.
[17] Rather than restrict an expert to a single interval in an
stochasticity of hydraulic conductivity throughout a partic-
ular hydraulic conductivity zone, as is traditionally accom-
plished by confidence intervals, it is more intuitive to permit
the expert to opine upon the uncertainty in the individual
measurement values that helped determine the random
variable in the purely probabilistic approach, above.
[18] It has been demonstrated that an appropriate expert
can simply opine upon the uncertainty of hydrogeological
measurements by assigning an interval in which the true
value is expected to lie [Joslyn and Kreinovich, 2005]
using ‘‘what is known about the underlying quantity’’
[Ferson et al., 2002]. Ferson et al. [2002] aptly note that
though this is the simplest approach, it is also the most
difficult to defend to others. They also provide alternative
methods for defining random set structures. For example,
knowing the measuring device (i.e., pump test, slug test),
an expert can simply define such an interval by stating that
the true value lies within x orders of magnitude of the
measured value [Ferson et al., 2002; Mathon et al., 2009].
Thus, in any one hydraulic conductivity subdomain with t
equiprobable measurements, a collection of t equiprobable
intervals is defined, representing t uncertain measurements.
Where the measurements are used to construct a discrete
CDF, these expert-provided intervals essentially bracket the
unknown true random variable hydraulic conductivity. A
collection of these intervals [Helton and Oberkampf, 2004;
Joslyn and Kreinovich, 2005], forms a random set, which is
isomorphic to a Dempster-Shafer body of evidence [Joslyn
and Booker, 2004].
[19] In this framework, a random set and associated
probability function are composed of a set of focal elements
{F, m}, where F is the set of focals (intervals) and m is the
basic mass assignment function that assigns nonzero prob-
abilities to the focals. A random set can be transformed into
lower and upper probability bounds, thereby bracketing the
unknown true random variable. These bounds are called
belief, bel(IG), and plausibility, pl(IG), respectively, for some
arbitrary focal element IG 2 G, a set of focals. Because the
random set definition requires less precision from the opining
expert, the resulting upper and lower bounding curves
(Figure 3) do not impose false precision in the parameter
uncertainty characterization and avoid any inaccuracies that
result from forcing an expert to provide information that
Table 2. Locations and Concentration Statistics for the Nodes
Whose Locations Are Plotted in Figure 1a
Variance of Variance of Variance of
Node Easting Northing Mean Case 1 Case 2 Case 3
400 1107 635 238 1499 1104 2248
234 1590 1030 645 829 234 3609
162 1450 682 998 1113 176 8713
190 1420 634 722 4144 979 13689
a
Locations and concentration are in feet.
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION W00B15

Figure 2. Random variables representing the concentrations at nodes 400 (top left), 234 (top right),
162 (bottom left), and 190 (bottom right). Intuitively, the variance changes throughout space and as the
expert-provided hydraulic conductivity confidence intervals are narrow (case 2) relative to the case of
interest (case 1) and relatively wide (case 3). The data for these random variables are given in Table 2.
The confidence intervals for the three cases are provided in Table 3.

would define a single random variable. Moreover, no prob- associated with a probability mass assignment m(IFj). The
ability model need be selected or assumed, which is desirable propagation of the random set-based hydraulic conductivity
in light of the above-mentioned possible inaccuracies in the values through to concentration values necessitates the use
lognormality assumption for hydraulic conductivity. of a tool to extend the flow and transport model such that it
[20] Formally, a hydraulic conductivity random set (F, mF), can operate upon these focal elements. Such an extension
is defined on the Cartesian product K = K1
. . .
Kn1, where would permit the calculation of concentration focal
Kj denotes the domain of the hydraulic conductivity value elements that can be aggregated into a concentration
and n1 is the number of uncertainty hydraulic conductivity random set at each location throughout the spatial domain.
zones (in the trial case, n1 = 3). In this formal definition, mF
is a function mapping elements, IF 2 F, of F to the interval
[0, 1],
Table 3. Expert-Provided Hydraulic Conductivity Confidence
mðIF Þ :! ½0; 1 IF 2 F; Intervals for the Base Case, a High-Certainty Case, and a Low-
P Certainty Casea
mF ðIF Þ ¼ 1:
IF :IF 2F Confidence Confidence Confidence
Unit Interval of Interval of Interval of
Since, in our trial example, the conductivity zones are Number Case 1 Case 2 Case 3
assumed uncorrelated, the n1-dimensional random sets are 1 [27.4, 23.8] [26.4, 24.8] [28.5, 22.6]
marginalized to random sets (Fj, mFj), j = 1,. . .,n1, each 2 [18.8, 18.8] [18.8, 18.8] [18.8, 18.8]
defined solely upon the individual domains Kj. In other 3 [20.9, 17.4] [19.9, 18.4] [21.9, 14.4]
words, one can specify (F, m F ) by means of n1 4 [17.2, 12.8] [16.6, 13.4] [18.5, 11.4]
5 [11.9, 11.9] [11.9, 11.9] [11.9, 11.9]
stochastically independent random sets.
a
[21] Each of these random sets is composed of a finite Conductivity is in m/s. Base case is case 1, high-certainty case is case 2,
number of intervals IFj, IFj 2 Fj, or focal elements, each and low-certainty case is case 3.

4 of 10
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION W00B15

Figure 3. Cumulative random set hydraulic conductivity values for zone 1 (top left), zone 2 (top right),
and zone 5 (bottom), determined by the ±2 orders of magnitude uncertainty on available measurements.

[22] Consider the extension of the transport model y =
f(x), where x = (x1,. . .,xn1) is the vector of uncertain
conductivities in n1 zones and y = (y1,. . .,yn2) is the vector
of uncertain concentration values at n2 nodes. Assuming,
without loss of generality, that n1 = 1 and K1 = W =
{k1,. . .,kL}, the domain of L possible hydraulic conductivity
values, where focal elements are subsets of W. The random
set extension principle [Dubois and Prade, 1991] defines
the random set concentration (G, m) at node i as
G ¼ f f ðIF ÞjIF 2 F g;
and
mðIG Þ ¼ fmðIF ÞjIG ¼ f ðIF Þg;
where IG represents a focal set of concentrations that is an
image of a hydraulic conductivity focal element IF through
the transport model f. This concentration focal element is
defined by
IG ¼ f ðIF Þ ¼ f f ðk Þjk 2 IF g:
A significant drawback to the random set extension
principle is its computational intensity, since a random set
5 of 10
uses the power set of its domain rather than the domain itself.
Therefore, an approximation of the random set hydraulic
conductivities, such that the extended transport model can be
carried out over W rather than its power set, is needed.
[23] In general, simplifying or approximating a random
set means approximating it by another random set, in which
the number of the focals containing relevant information is
reduced [Bauer, 1996]. An approach to approximating a
random set, presented by Dubois and Prade [1990], is
adopted in this paper. This approximation uses the
following steps: (1) formation of sets of focals of the
original random set that are the focals of the approximating
random set and (2) allocation of basic probability masses to
the sets from the first step using a process that is optimal in
the sense that the resulting hydraulic conductivity focal
elements of the approximating random set are the smallest
in size, effectively minimizing the amount of imprecision
introduced by the approximation.
[24] Approximating a random set by the above method
results in nested focal elements, which comprise a fuzzy set
A [Klir and Yuan, 1995], a special type of random set and
defined by a membership function (discussed further
below). This fact permits the solution of the transport
model over the domain W, rather than its power set. For the
exact representation of the approximating conductivity
fuzzy set on W, the membership function is calculated by
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION W00B15

Figure 4. Possibilistic approximations for the random set provided by the expert for the hydraulic
conductivity values associated with zone 1 (top left), zone 2 (top right), and zone 5 (bottom). The
hydraulic conductivity values for zones 3 and 4 are considered certain and precise.

the one-point coverage function [Goodman and Nguyen,
1985] for random sets:
X
for all k 2 W : mA ðk Þ ¼ mðCi Þ;
Ci :k2Ci
where Ci are the focal elements of the approximating
hydraulic conductivity random set and mA denotes the
membership function of the hydraulic conductivity fuzzy set
A. This fuzzy set representation of the uncertain hydraulic
conductivity values allows for the use of the special case of
the extension principle, described above, for fuzzy sets
[Dubois and Prade, 1991], which states that in order to
calculate the possibility value of an uncertain concentration
value one must consider membership values of hydraulic
conductivities used to calculate that concentration:
for all y2 f ðWÞ : mf ð AÞ ð yÞ ¼ supfmA ðk Þj y ¼ f ðk Þg;
where mf(A) represents the membership function of the fuzzy
concentration at a given node.
[25] Where fuzzy sets are used to approximate random
sets, the application of the extension principle [Klir and
Yuan, 1995] to the model equations is relatively straightfor-
ward and has precedent in hydrogeological applications
[Dou et al., 1995, 1997a, 1997b; Prasad and Mathur,
2007]. Nevertheless, in our example case the vertex method
[Ross, 2004], an approximation to the extension principle, is
applied to reduce computational effort. The vertex method
results in, for each a-cut (an interval created by the
horizontal cut of a fuzzy set at a given a, or membership
value), eight concentration values at each location due to the
eight possible permutations of the three lower and three
upper hydraulic conductivity bounds (from three uncertain
hydraulic conductivity values) of the a-cut. The minimum
of these eight values is taken as the lower bound of the
concentration a-cut, and the maximum is assigned as the
upper bound.
[26] If the function being extended (i.e., the transport
model) is nonlinear and monotonic with respect to its
variables (in our case, three hydraulic conductivity values),
then there is a possibility that the function will take smaller
or larger values for a combination of three hydraulic
conductivity values that are not necessarily a permutation
of the respective a-cut bounds, but rather a permutation of
values sampled anywhere within these bounds. If one
computes with the 8 permutations of the bounds of the
a-cuts of the three fuzzy hydraulic conductivity fuzzy sets
and calculates eight concentration values, one assumes that
the concentration at that node cannot get any smaller than
the minimum of the eight values and cannot get any larger
than the maximum of the eight values. In other words, one

6 of 10
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION W00B15

Figure 5. The cumulative belief and plausibility curves are relatively narrow and do not entirely bound
the corresponding cumulative distribution functions for the three cases at nodes 400 (top left) and 234
(top right), but are rather wide and do bound the cumulative distribution functions for nodes 162 (bottom
left) and 190 (bottom right).

admits that there is no need to investigate the entire
hydraulic conductivity a-cuts in order to determine the
a-cut bounds of the concentration at any node.
[27] If the alpha cuts of the fuzzy hydraulic conductivities
have one or more extreme points in the interior, then the
vertex method approach can be taken as approximations to
the true global extreme values that determine the bounds of
the concentration at the given alpha level. This will result in
fuzzy nodal concentration values with narrower support
implying higher specificity in the information content than
there actually exists [Klir and Yuan, 1995]. However, it is
the authors’ opinion that the nonlinearity of the relationship
between nodal concentrations and hydraulic conductivity is
mostly monotonic and the extreme concentrations will be at
the vertices (a-cut bounds), rather than anywhere between
the vertices. Thus, for the contaminant transport problem
considered here, it is more efficient to use the vertex method
than to invoke a global optimization tool that implements
the extension principle and, as such, searches the entire
a-cut. Where monotonicity cannot be justified, the authors
recommend applying the extension principle.
[28] Figure 4 provides examples of a fuzzy set defining
an uncertain hydraulic conductivity value. The interpreta-
tion of these fuzzy sets in Figure 4 that is most pertinent to
this topic is that of a possibility distribution [Zadeh, 1978].
7 of 10
Where model inputs are defined as fuzzy sets, model
estimates of concentration are interpreted as possibility
distributions. A possibility distribution defines for each
value along the horizontal axis the degree to which that
value is possible, given available evidence. In Figure 4 (top
left), for instance, the hydraulic conductivity value 3.5

109 m/s is most possible. This is similar to probability
theory, whereupon inspection of the peak of a probability
density function would reveal the most probable value.
[29] Whereas, through the transformation to fuzzy sets,
random sets offer a relatively facile strategy for computation
with uncertainty, the most significant advantage to the
random set approach is its potential to characterize both
aleatory and epistemic uncertainty. The foundation for
random sets lies in probability theory, which, as mentioned
above, is ideally suited for aleatory uncertainty character-
ization. On the other hand, random sets are less specific, or
less precise, than random variables [Joslyn and Booker,
2004], because focal elements, upon which random set are
based, are a source of imprecision in the uncertainty
quantification process (focal elements associated with
conventional probabilities, random variables, are points
and therefore more precise than general random sets). This
imprecision is a form of epistemic uncertainty. An expert
who provides a body of evidence to characterize the
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION
uncertainty regarding some hydraulic conductivity measure-
ment actually admits to the existence of these two forms of
uncertainty. The natural randomness (aleatory uncertainty)
of hydraulic conductivity is captured by the stochastic
nature of random sets (the basic mass assignments). Owing
to an expert’s inability to precisely define this natural
randomness, a random set merely provides bounds on the
exact random variable (because the basic mass function is
defined over intervals of the conductivity domain rather
than the domain itself), thereby imprecisely defining the
random variable (epistemic uncertainty).
[30] The starting point for the application of random set-
based uncertainty characterization is similar to the probabi-
listic approach presented above. Given the measurements in
each zone in Figure 1, the expert, armed with knowledge of
the measurement technique and aquifer characteristics,
provides the aforementioned intervals on the each measured
value by specifying that the true hydraulic conductivity
value lies within ±2 orders of magnitude of the measure-
ment. As mentioned above, where the measurements in a
particular zone are used to construct an approximate CDF,
these expert-provided intervals become upper and lower
bounds on the zone’s true random variable hydraulic con-
ductivity (plausibility and belief, respectively). Figure 3
shows these upper and lower bounds for zones one, two
and five. As in the Monte Carlo approach, the hydraulic
conductivity values for zones three and four are considered
certain and precise. The fuzzy set approximations of these
random sets, whose construction is outlined above, are
shown in Figure 4.
[31] Though the set of intervals and associated probabil-
ities provided by the expert are a natural extension of the
confidence interval in the Monte Carlo approach above,
they comprise a greater amount of information. As such, the
bodies of evidence provided by the expert capture both the
uncertainty surrounding the mean hydraulic conductivity
values and the imprecision with which the expert can truly
characterize this uncertainty.
[32] The vertex method [Ross, 2004], an approximation
to the extension principle, was applied to the finite element
approximation equations of the groundwater flow and
transport model in order to propagate the possibilistic
uncertainty through to the concentrations values. As a
result, uncertain concentration estimates are described by
possibility distributions. The possibilistic concentration
values can be transformed into upper (plausibility) and
lower (belief) bounds on the unknown random variable. The
resulting bounds for the nodes of interest are plotted in
Figure 5 with the corresponding random variables from all
three cases of uncertainty in cumulative distribution
function form (dashed lines) from Figure 2. If the intervals
used to construct the random sets are certain to contain the
value of the measured variable, the true (and unknown)
probability distribution, which the confidence intervals aim
to characterize, defining a random variable lies between the
plausibility and belief curves, especially where these bounds
are widely separated [Ferson et al., 2002]. Thus, the burden
is upon the expert, who specified the magnitude of
measurement error that creates the random sets, to ensure
that the measurement intervals are wide enough to bound
the true hydraulic conductivity measurement yet narrow
8 of 10
W00B15
enough so as to be meaningful. Since the error of various
measurement techniques are commonly acknowledged and,
at times, quantified [Mathon et al., 2009] we perceive this
task to be reasonable.
4. Discussion of Results
[33] The outcome of the random set approach to uncer-
tainty characterization is significantly distinct from that of
the traditional purely probabilistic approach. Intuitively,
the strictly probabilistic approach produces random vari-
able concentration values, whereas the random set-based
approach results in upper and lower probability bounds. As
noted above, assuming that uncertainty in both approaches
was characterized by the same expert or different experts with
the same understanding of pertinent data, the upper and lower
probabilities (plausibility and belief) should bracket the
corresponding probability distribution (produced by the
stochastic approach) at a given location. Thus, the range of
concentration values that result from the random set approach
is greater than that which would result from the strict Monte
Carlo method, owing to the imprecision inherent in the
random set approach. However, the degree of precision
presented in the random variable approach is, as argued
herein, difficult to justify.
[34] Consider the same nodal locations whose concentra-
tion random variables are plotted in Figure 2. The upper and
lower probabilities for these same nodal locations are
plotted in Figure 5, along with the concentration random
variables for all three cases of uncertainty presented in
Table 1. Note that, at some nodes, the upper and lower
probabilities entirely bound the corresponding random var-
iables, whereas, other nodes do not entirely bound the
corresponding cumulative distribution function. Such a
discrepancy is likely due to the fact that the expert employed
to define the confidence intervals was not the same as the
expert who provided the information used to construct the
random set intervals.
[35] Uncertainty (variance) associated with concentration
values changes throughout the spatial domain, as is evident
by the different slopes in the cumulative distribution func-
tions in Figure 2. Nevertheless, high variances may not
simply be associated with high degrees of uncertainty, but
rather with means of greater magnitude. Random set-based
probability bounds, however, are independent of the mag-
nitude of the concentration values and provide a true means
of uncertainty identification. Wider bound separation signi-
fies more uncertainty in concentration estimates, and, as a
result, locations where more data are warranted. For in-
stance, the estimate in Figure 5 (bottom right) is more
uncertain than that in Figure 5 (top right) and, as such, is
in need of additional data. In this particular instance, a
combination of node 234’s distance from the contaminant
source and the expert’s precision in providing error bounds
on hydraulic conductivity measurements from zone 4 rela-
tive to the measurements in zone 1 contributed to the lower
separation between the probability bounds for the concen-
tration random set at node 234 (Figure 5, top right) relative
to that for node 190 (Figure 5, bottom right). Likewise, the
expert believed the measurements from zone 2 to be slightly
more reliable than those in zone 4, and thus, the uncertainty
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION
surrounding the concentration at node 400 is lower than it is
for Nodes 190 and 162.
5. Conclusion
[36] Because thorough hydrogeological investigations
cost significant amounts of money and time, an efficient
means of data acquisition and interpretation is valuable. One
such means is expert knowledge extraction. However, the
consideration of expert knowledge introduces epistemic
uncertainty in addition to the existing stochasticity in
hydrogeological parameters, such as hydraulic conductivity.
Thus, appropriate characterizations of uncertainty should
delineate aleatory uncertainty from epistemic uncertainty, as
has been done in risk assessment and reliability engineering
[Helton et al., 2000a, 2000b, 2004; Hofer et al., 2002;
Helton and Oberkampf, 2004; Oberkampf et al., 2004].
[37] In this paper, we have sided with Ganoulis [1996]
who also argued that probability theory alone cannot
accomplish this. The hazard associated with applying
traditional probability theory is that the opining expert may
provide inaccurate and artificially precise characterizations
of the random variable that best captures the naturally
stochastic nature of hydraulic conductivity. Fuzzy set theory,
on the other hand, has failed to find mainstream acceptance
perhaps as a result of its departure from probability theory.
[38] Random sets were introduced in this paper as an
alternative and possibly more appropriate means for the
characterization of both aleatory (the random variable) and
epistemic (expert-characterized measurement error) uncer-
tainty. This approach to uncertainty characterization pro-
vides a methodology for bounding an unknown random
variable and properly capturing the imprecise nature of
expert knowledge. In the provided example, expert knowl-
edge was used to characterize the reducible uncertainty of
individual hydraulic conductivity measurements. Random
sets were collected from these individual measurement
intervals and propagated through a groundwater flow and
transport model using fuzzy set methodologies.
[39] Aside from avoiding the imposition of false preci-
sion, which is an unfortunate side effect of defining confi-
dence intervals, it is important to note that uncertainty
characterization via random sets eliminates the need for
any probability model definition or assumption. Moreover,
the approximation of the random sets by fuzzy sets and
model execution with the fuzzy extension principle accom-
plishes what fuzzy set-based hydrogeological research has,
as yet, failed to embrace – the combination of probability
theory and fuzzy sets for the characterization of parameter
uncertainty. If fuzzy set theory is to find a stronger foothold
in engineering applications, researchers must endeavor to
embrace hybrid frameworks that unite fuzzy sets with more
traditional mathematical tools such as probability, as is
illustrated by this paper.
[40] While the representation of model concentration
estimates as upper and lower probabilities (plausibility
and belief) provides a transparent comparison between the
random variable and random set approaches defined above,
the utility of data in such a form may not be immediately
obvious. What does one do with an imprecise notion of a
stochastic estimate concentration (Figure 5)? In fact, the
representation of concentration estimates as possibility dis-
tributions (like the possibilistic hydraulic conductivity values
9 of 10
W00B15
in Figure 4), which contain the same information, and can be
transformed into, probability bounds, is quite intuitive and
readily interpretable. Inspection of a possibility distribution
reveals not only the most possible concentration value, but
also a range of concentrations that are also possible to lesser
and varying degrees. In fact, algorithms have been developed
to complement and refine possibilistic model estimates with
new information [Fruhwirth-Schnatter, 1993; Pan and Klir,
1996; Yang, 1997; Ross et al., 2007, 2008].
[41] A benefit of separately characterizing aleatory and
epistemic uncertainties is the possibility of identifying
where and what type of additional information is most
beneficial. The value of additional information is correlated
with the reduction in reducible (epistemic) uncertainty
realized by the consideration of the new information; this
is easy to identify using belief and plausibility curves. The
appropriate measure is the magnitude of epistemic uncer-
tainty as indicated by the distance between the belief and
plausibility curves. As noted above, the concentration
estimate at node 190 is less precise than that at node 234
(Figure 5). Thus additional data are most valuable at node
234, where the reducible uncertainty, and likewise the dis-
tance between belief and plausibility curves, is greatest. Klir
[2006] provides a set of measures to quantify the amount of
information as well as uncertainty in random sets.
[42] Though the application presented above focuses
upon the characterization of uncertainty in hydraulic con-
ductivity measurements, other forms of uncertainty such as
boundary conditions also can be considered. In the case of
boundary conditions, which originate predominantly from
expert insight, fuzzy sets can be used directly as a charac-
terization methodology, bypassing the need for random sets.
The propagation of these forms of uncertainty through a
groundwater flow and transport model is executed as
presented above.
[43] Acknowledgments. This material is based upon work supported
by the Strategic Environmental Research and Development Program
(SERDP). Any opinions, findings, and conclusions or recommendations
expressed in this material are those of the authors and do not necessarily
reflect the views of SERDP.
References
Apostolakis, G. (1990), The concept of probability in safety assessments of
technological systems, Science, 250(4986), 1359 – 1364, doi:10.1126/
science.2255906.
Bagtzoglou, A. C., A. Nedungadi, and B. Sagar (1996), A fuzzy rule-based
model for flow simulation in heterogeneous media, in Computational
Methods in Water Resources XI, pp. 629 – 637, Comput. Mech. Publ.,
Boston.
Bardossy, A., I. Bogardi, and W. E. Kelly (1989), Geostatistics utilizing
imprecise (fuzzy) information, Fuzzy Sets Syst., 31, 311 – 328,
doi:10.1016/0165-0114(89)90203-0.
Bardossy, A., I. Bogardi, and L. Duckstein (1990a), Fuzzy regression in
hydrology, Water Resour. Res., 26(7), 1497 – 1508.
Bardossy, A., I. Bogardi, and W. E. Kelly (1990b), Kriging with imprecise
(fuzzy) variograms I: Theory, Math. Geol., 22(1), 63 – 79, doi:10.1007/
BF00890297.
Bardossy, A., I. Bogardi, and W. E. Kelly (1990c), Kriging with imprecise
(fuzzy) variograms II: Application, Math. Geol., 22(1), 81 – 94,
doi:10.1007/BF00890298.
Bauer, M. (1996), Approximation algorithms and decision making in the
Dempster-Shafer theory of evidence: An empirical study, Int. J. Approx-
imate Reasoning, 17, 217 – 237, doi:10.1016/S0888-613X(97)00013-3.
Choquet, G. (1954), Theory of capacities, Ann. Inst. Fourier, 5, 131 – 295.
Demmico, R. V., and G. J. Klir (2004), Fuzzy Logic in Geology, 347 pp.,
Elsevier, Amsterdam.
W00B15 ROSS ET AL.: UNCERTAINTY IN GROUNDWATER SIMULATION
Domenico, P. A., and F. W. Schwartz (1990), Physical and Chemical
Hydrogeology, 824 pp., John Wiley, New York.
Dou, C., W. Woldt, I. Bogardi, and M. Dahab (1995), Steady state ground-
water flow simulation with imprecise parameters, Water Resour. Res.,
31(11), 2709 – 2719, doi:10.1029/95WR02310.
Dou, C., W. Woldt, I. Bogardi, and M. Daheb (1997a), Numerical solute
transport simulation using fuzzy sets approach, J. Contam. Hydrol.,
27(1 – 2), 107 – 126, doi:10.1016/S0169-7722(96)00047-2.
Dou, C., W. Woldt, M. Daheb, and I. Bogardi (1997b), Transient ground-
water flow simulation using a fuzzy set approach, Ground Water, 35(2),
205 – 215, doi:10.1111/j.1745-6584.1997.tb00076.x.
Dou, C., W. Woldt, and I. Bogardi (1999), Fuzzy rule-based approach to
describe solute transport in the unsaturated zone, J. Hydrol., 220(1 – 2),
74 – 85, doi:10.1016/S0022-1694(99)00065-7.
Dubois, D., and H. Prade (1990), Consonant approximations of belief
functions, Int. J. Approximate Reasoning, 4, 419 – 449, doi:10.1016/
0888-613X(90)90015-T.
Dubois, D., and H. Prade (1991), Random sets and fuzzy interval analysis,
Fuzzy Sets Syst., 42, 87 – 101, doi:10.1016/0165-0114(91)90091-4.
Fang, J. H., and H. C. Chen (1997), Fuzzy modeling and the prediction of
porosity and permeability from the compositional and textural attributes
of sandstone, J. Pet. Geol., 20(2), 185 – 204, doi:10.1111/j.1747-
5457.1997.tb00772.x.
Ferson, S., V. Kreinovich, L. Ginzburg, D. S. Myers, and K. Sentz (2002),
Constructing probability boxes and Dempster-Shafer structures, Tech.
Rep. SAND2002 – 4015, Sandia Natl. Lab., Albuquerque, N. M.
Fruhwirth-Schnatter, S. (1993), On fuzzy bayesian inference, Fuzzy Sets
Syst., 60, 41 – 58, doi:10.1016/0165-0114(93)90288-S.
Ganoulis, J. (1996), Modeling hydrologic phenomena, Rev. Sci. Eau, 9(4),
421 – 434.
Goodman, I. R., and H. T. Nguyen (1985), Uncertainty Models for
Knowledge-Based Systems: A Unified Approach to the Measurement
of Uncertainty, 643 pp., Elsevier, Amsterdam.
Guan, J., and M. M. Aral (2004), Optimal design of groundwater remedia-
tion systems using fuzzy set theory, Water Resour. Res., 40, W01518,
doi:10.1029/2003WR002121.
Helton, J. C., and W. L. Oberkampf (2004), Alternative representations
of epistemic uncertainty (Guest Editorial), Reliab. Eng. Syst. Safety,
85(1 – 3), 1 – 10, doi:10.1016/j.ress.2004.03.001.
Helton, J. C., F. J. Davis, and J. D. Johnson (2000a), Characterization of
stochastic uncertainty in the 1996 performance assessment for the Waste
Isolation Pilot Plant, Reliab. Eng. Syst. Safety, 69(1 – 3), 167 – 189,
doi:10.1016/S0951-8320(00)00031-4.
Helton, J. C., M.-A. Martell, and M. S. Tierney (2000b), Characterization of
subjective uncertainty in the 1996 performance assessment for the Waste
Isolation Pilot Plant, Reliab. Eng. Syst. Safety, 69(1 – 3), 191 – 204,
doi:10.1016/S0951-8320(00)00032-6.
Helton, J. C., J. D. Johnson, and W. L. Oberkampf (2004), An exploration
of alternative approaches to the representation of uncertainty in model
predictions, Reliab. Eng. Syst. Safety, 85(1 – 3), 39 – 71, doi:10.1016/
j.ress.2004.03.025.
Hofer, E., M. Kloos, B. Krzykacz-Hausmann, J. Peschke, and M. Woltereck
(2002), An approximate epistemic uncertainty analysis approach in the
presence of epistemic and aleatory uncertainties, Reliab. Eng. Syst.
Safety, 77(3), 229 – 238, doi:10.1016/S0951-8320(02)00056-X.
Joslyn, C., and J. M. Booker (2004), Generalized information theory for
engineering modeling and simulation, in Engineering Design and
Reliability Handbook, edited by E. Nikolaidis et al., pp. 9-1 – 9-40,
CRC, New York.
Joslyn, C., and S. Ferson (2004), Approximate representations of random
intervals for hybrid uncertainty quantification in engineering modeling,
in Sensitivity Analysis of Model Output (SAM2004), edited by K. M.
10 of 10
W00B15
Hanson and F. M. Hemez, pp. 453 – 469, Los Alamos Natl. Lab., Los
Alamos, N. M.
Joslyn, C., and V. Kreinovich (2005), Convergence properties of an interval
probabilistic approach to system reliability estimation, Int. J. Gen. Syst.,
34(4), 465 – 482, doi:10.1080/03081070500033880.
Klir, G. J. (2006), Uncertainty and Information: Foundations of General-
ized Information Theory, 499 pp., John Wiley, Hoboken, N. J.
Klir, G. J., and B. Yuan (1995), Fuzzy Sets and Fuzzy Logic: Theory and
Applications, 574 pp., Prentice-Hall, Upper Saddle River, N. J.
Mathon, B., M. Ozbek, and G. F. Pinder (2009), Dempster-Shafer theory
applied to uncertainty surrounding permeability, Math. Geosci., in press.
Oberkampf, W. L., J. C. Helton, C. A. Joslyn, S. F. Wojtkiewicz, and
S. Ferson (2004), Challenge problems: Uncertainty in system response
given uncertain parameters, Reliab. Eng. Syst. Safety, 85(1 – 3), 11 – 19,
doi:10.1016/j.ress.2004.03.002.
O’Hagan, A., and J. E. Oakley (2004), Probability is perfect, but we can’t
elicit it properly, Reliab. Eng. Syst. Safety, 85(1 – 3), 239 – 248,
doi:10.1016/j.ress.2004.03.014.
Ozbek, M. M., and G. F. Pinder (2006), Non-probabilistic uncertainty in
subsurface hydrology and its applications: An overview, Water Air Soil
Pollut., 6, 35 – 46, doi:10.1007/s11267-005-9011-4.
Pan, Y., and G. J. Klir (1996), Bayesian inference of fuzzy probabilities, Int.
J. Gen. Syst., 26(1 – 2), 73 – 90.
Prasad, R. K., and S. Mathur (2007), Groundwater flow and contaminant
transport simulation with imprecise parameters, J. Irrig. Drain. Eng.,
133(1), 61 – 70.
Ricciardi, K., G. F. Pinder, and G. P. Karatzas (1998), A new probability
density function for hydraulic conductivity in optimal design, Eos Trans.
AGU, 79(45), Fall Meet. Suppl., F280.
Ross, J., M. Ozbek, and G. F. Pinder (2006), Fuzzy kalman filtering of
hydraulic conductivity, in Computational Methods in Water Resources
XVI [CD ROM], Comput. Mech. Publ., Boston.
Ross, J., M. M. Ozbek, and G. F. Pinder (2007), Hydraulic conductivity
estimation via fuzzy analysis of grain size data, Math. Geol., 39(8),
765 – 780, doi:10.1007/s11004-007-9123-7.
Ross, J., M. Ozbek, and G. F. Pinder (2008), Kalman filter updating of
possibilistic hydraulic conductivity, J. Hydrol., 354(1 – 4), 149 – 159,
doi:10.1016/j.jhydrol.2008.03.005.
Ross, T. (2004), Fuzzy Logic with Engineering Applications, 654 pp.,
John Wiley, West Sussex, U. K.
Shafer, G. (1976), A Mathematical Theory of Evidence, 297 pp., Princeton
Univ. Press, Princeton, N. J.
Srinivasan, G., D. M. Tartakovsky, B. A. Robinson, and A. B. Aceves
(2007), Quantification of uncertainty in geochemical reactions, Water
Resour. Res., 43, W12415, doi:10.1029/2007WR006003.
Yang, C. C. (1997), Fuzzy bayesian inference, IEEE Trans. Syst. Man
Cybern., 3, 2707 – 2712.
Zadeh, L. (1965), Fuzzy sets, Infect. Control, 8, 338 – 353, doi:10.1016/
S0019-9958(65)90241-X.
Zadeh, L. (1978), Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets
Syst., 1, 3 – 28, doi:10.1016/0165-0114(78)90029-5.
Zhang, Y., and G. Pinder (2003), Latin hypercube sample selection strategy
for correlated random hydraulic conductivity fields, Water Resour. Res.,
39(8), 1226, doi:10.1029/2002WR001822.


M. M. Ozbek, Environ International Corporation, 214 Carnegie Center,
Princeton, NJ 08540, USA.
G. F. Pinder and J. L. Ross, Center for Groundwater Remediation
Design, School of Engineering, University of Vermont, Burlington, VT
05405, USA. (jlross@uvm.edu)