You are on page 1of 13

Dealing with epistemic uncertainty in fractured rock masses

A. Bedi
Imperial College, London, United Kingdom
J.P. Harrison
University of Toronto, Canada


Abstract: This paper presents key concepts regarding the use of Limit State Design
(LSD) in rock engineering.
The last two decades has seen significant development in LSD principles for
application in geotechnical engineering, with the result that they can now be routinely
applied to designs involving soil. However, robust application of LSD seems to be not
possible for designs involving fractured rock. We suggest that this is due to the
uncertainty model that forms the basis of LSD, and hence Eurocode 7, being inappro-
priate for fractured rock.
Eurocode 7 (EC7) requires that uncertainty in geotechnical properties is stochastic
in nature, the so-called aleatory model. However, we show that this model is not
always applicable to rock mass parameters, and that uncertainty is often due to
insufficient knowledge or subjectivity so-called epistemic uncertainty.
Based on an understanding of these two uncertainty models, this paper discusses
the shortcomings of EC7 to deal with epistemic uncertainty in the context of fractured
rock masses, and emphasises the importance of firstly recognising the differences
between aleatory variability and epistemic uncertainty, and secondly appreciating the
implications of such uncertainty for rock engineering design to limit state concepts.
Finally, we give pragmatic suggestions for how the epistemic uncertainty of rock
mechanics properties may be incorporated in the current LSD paradigm.

Theme: Maintenance, LCC and Risk Management.

Keywords: Eurocode 7, Limit State Design, aleatory, epistemic, uncertainty.
Rock Engineering and Technology for Sustainable Underground Construction
Eurock 2012 the 2012 ISRM International Symposium, 28-30 May 2012, Stockholm, Sweden.
BeFo and ISRM, 2012
Eurock 2012 Page 2
1 INTRODUCTION
Limit State Design (LSD) principles and codes are commonplace in structural
engineering designs, and in order that geotechnical designs might coherently interface
with these (e.g. in the design of foundations or deep basement structures) there is
concerted effort underway around the world to develop and implement LSD
techniques for geotechnical engineering.
Within Europe, EN-1997-1 (commonly referred to as Eurocode 7, or EC7)
embodies LSD principles for geotechnical engineering designs (CEN, 2004; Becker &
Moore, 2007). Figure 1a & b schematically shows the complete contents of EC7 and
those parts applicable to rock engineering design, respectively. Currently EC7 is well
developed for LSD where the geological material is soil, but anecdotal evidence
suggests that it is difficult to apply to rock engineering designs. As such, ad hoc and
empirical approaches continue to be used (Feng & Hudson, 2010). However, as rock
engineering structures often incorporate elements that are designed using limit state
principles (for example, the steel and concrete support elements used in tunnels and
other underground openings), the requirement for geotechnical design to coherently
interface with structural design means it is critically important that LSD be extended
to embrace rock engineering.
In this paper, we show that the putative difficulty in applying LSD principles to
rock engineering is the means by which the uncertainty associated with the
geotechnical input parameters required to characterise fractured rock is handled. Some
of this uncertainty derives from inherent random variability related to natural
fluctuations of a quantity within the physical system under consideration (Dubois &
Guyonnet, 2011), and is termed aleatory (derived from the Latin alea, which
means the rolling of dice). Other uncertainty relates to insufficient knowledge, and is
both subjective in nature and influenced by preconceptions of what is considered
realistic for the system in question. This is termed epistemic (derived from the Greek
episteme, meaning knowledge) (Kiureghian & Ditlevsen, 2009). Uzielli (2008)
affirms that in geotechnical engineering the terms variability and uncertainty are often,
incorrectly, used interchangeably, and goes on to define these terms as:
Variability an observable manifestation of heterogeneity of one or more
physical parameters and/or processes.
Uncertainty pertains to the modellers state of knowledge and strategy, and
reflects the decision to recognise and address the observed variability in a qualitative
or quantitative manner.
Thus, aleatory uncertainty also referred to as irreducible, objective, or stochastic
uncertainty describes the inherent variability in a physical system or environment

a) Contents of Eurocode 7 b) Contents of Eurocode 7 applicable to rock
engineering
Figure 1. Contents of Eurocode 7 Part 1(after Harrison, 2010)
Pile
foundations
(13%)
Hydraulic
failure
(4%)
Overall
stability
(4%)
Embankments
(3%)
Annexes A-J
(24%)
General
(6%)
Basis of
geotechnical design
(12%)
Geotechnical
data
(7%)
Supervision of
construction,
monitoring and
maintenance
(4%)
Fill, dewatering, ground improvement
and reinforcement (4%)
Spread
foundations
(6%)
Anchorages
(4%)
Retaining
structures
(9%)
Overall
stability
(5%)
Annexes A-J
(34%)
General
(8%)
Basis of
geotechnical design
(17%)
Geotechnical
data
(10%)
Supervision of
construction,
monitoring and
maintenance
(6%)
Fill, dewatering, ground
improvement and reinforcement
(6%)
Spread
foundations (8%)
Anchorages
(6%)
Eurock 2012 Page 3
(Oberkampf et al., 2004) and can be modelled using stochastic models and handled
using probabilistic methods. Here, we refer to this as aleatory variability. Epistemic
uncertainty, however, is abstract in nature, has properties that are neither amenable to
parameterisation nor repeated experiment and, as we will show in the following
discussion presented in this report, thus cannot be modelled stochastically.
The current formulation of LSD for geotechnical engineering assumes that
uncertainty is entirely aleatory, and can be modelled using a known probability
density distribution. However, we give examples to show that much uncertainty in
rock mechanics (and hence rock engineering) is epistemic. As there is currently a lack
of understanding on either how to address epistemic uncertainty or whether it can be
considered within an aleatory framework, it is questionable whether the LSD
principles within EC7 can be applied to rock engineering. In this paper, we firstly
explore the nature of uncertainty in rock mechanics and rock engineering, and then
following a brief explanation of the fundamental concepts of LSD outline the
difficulties of applying LSD to rock mechanics and rock engineering, before
concluding with pragmatic recommendations for the adoption of LSD in rock
engineering.
2 UNCERTAINTY AND VARIABILITY IN ROCK ENGINEERING
In comparison to man-made materials such as steel and concrete, geomaterials possess
marked variability in material parameters. Figure 2 uses coefficient of variation (i.e.
ratio of standard deviation to the mean value) to demonstrate this for various material
characteristics. Although coefficient of variation has been used here as a measure of
variability, the underlying nature of such uncertainty, i.e. aleatory or epistemic, must
be realised before one can embark on any analysis of it.
In the context of rock mechanics, it is the inherent randomness of natural, geologi-
cal processes that gives rise to aleatory variability. This is displayed in the variation,
within a nominally uniform material, of properties such as uniaxial compressive
strength. Such variability is usually characterised by stochastic models and handled
using probability theory. On the other hand, epistemic uncertainty may result from the
lack of field or laboratory investigation data, information about events and processes,
or understanding of physical laws governing the real world (Baecher & Christian,
2003). Such lack of knowledge means that the assumption that a stochastic model and
associated probabilistic analysis is applicable may well be incorrect. The appropria-
teness of stochastic models to define aleatory variability and epistemic uncertainty is
demonstrated with the aid of an example.

Figure 2. Range of coefficient of variation (COV) for some geotechnical and man-made materials
(after Bond & Harris, 2008; Uzielli, 2008)
P
r
o
p
e
r
t
y
Eurock 2012 Page 4
Figure 3a presents a set of data containing thirty samples (e.g. laboratory test results)
and the distribution fitted to it. This figure demonstrates that when sufficient samples
are available, the uncertainty may be modelled as aleatory, i.e. a known stochastic
function in this case, Gaussian is appropriate to characterise the variability. On the
contrary, if presented with insufficient data for example, either of the two subsets
(A or B) shown in Figure 3b, each limited to seven outcomes drawn from the data
set the uncertainty must be regarded as epistemic, and thus the assumption of a
stochastic function to adequately define it cannot be justified.
As much of epistemic uncertainty arises from a lack of knowledge, it may be
reduced but not completely eliminated through additional information such as
further field or laboratory tests (see Figure 4). If sufficient additional information is
obtained, it may become justifiable to characterise the uncertainty as variability, i.e.
apply an aleatory model. Once an acceptable aleatory model has been developed,
additional investigation will not reduce the variability but may increase the precision
of the parameters that describe it (Christian, 2004). This concept is demonstrated
further in the following section through an example of characterisation of the uniaxial
compressive strength of intact rock.
At present, the fundamental and intrinsic difference between aleatory and
epistemic uncertainty appears to be the source of much confusion in geotechnical
analysis. In rock mechanics and rock engineering in particular, although probabilistic
approaches to analysing uncertainty have a long history (e.g. Priest & Brown, 1983;
Zhang & Einstein, 1998), it may be more appropriate to utilise alternative, non-
probabilistic techniques to handle rock mass parameters that exhibit epistemic
uncertainty (Wenner & Harrison, 1996; Harrison & Hudson, 2010). In this way, the
true nature of geotechnical uncertainty would be addressed rather than considering all
uncertainty as aleatory variability (Uzielli, 2008).

Figure 4. Uncertainty type (from Guo & Du, 2007)
2.1 Rock mass parameters: aleatory or epistemic?
Two useful acronyms to describe rock masses are CHILE (Continuous, Homogeneous,
Isotropic, Linear, and Elastic) and DIANE (Discontinuous, Inhomogeneous, Aniso-
tropic, Non-linear Elastic) (Hudson & Harrison, 1997). The first of these is the
Present knowledge
Complete
knowledge
Complete
ignorance
Epistemic uncertainty Aleatory uncertainty
K
n
o
w
l
e
d
g
e

a) Data exhibiting aleatory uncertainty b) Data exhibiting epistemic uncertainty
Figure 3. Appropriateness of stochastic model to define datasets exhibiting aleatory and epistemic
uncertainty.
Value
P
r
o
b
a
b
i
l
i
t
y
F
r
e
q
u
e
n
c
y
Gaussian
distribution
fitted to 30
0
1
2
3
4
5
6
7
0
0.05
0.10
0.15
0.20
0.25
Value
P
r
o
b
a
b
i
l
i
t
y
F
r
e
q
u
e
n
c
y
0
1
2
3
4
5
6
7
0
0.05
0.10
0.15
0.20
0.25
Subset A
n = 7
Subset B
n = 7
Gaussian
distribution
fitted to 30
Eurock 2012 Page 5
simplifying assumption commonly adopted when undertaking design of rock
engineering structures, whereas the second is the physical nature of the material in
which engineering takes place. Undertaking rock engineering in CHILE rock masses
would be straightforward: the material properties could be fully understood and
variability would not exist. However, the heterogeneity in DIANE rock masses needs
to be characterised, the various rock mass parameters assessed as aleatory or epistemic
(as shown in Figure 5), and these parameters then handled using an appropriate
variability or uncertainty model.
The main components of epistemic uncertainty are: measurement or interpretation
errors or inadequate data representation during site characterisation; modelling
uncertainty, as to whether the selected mathematical model is an accurate representa-
tion of reality; and, parameter uncertainty in terms of how model parameters are
estimated and analysed. Additionally, geotechnical engineers often rely on empiri-
cism or expert judgement to determine rock mass parameters, and these may introduce
subjectivity as a form of epistemic uncertainty.

Figure 5. Sources of uncertainty (after Baecher & Christian, 2003)
2.1.1 Assessment of properties
The difficulties in assessing properties that display such epistemic uncertainty are
exemplified in the work of Beer et al. (2002), which describes the results of an online
test of visual assessment of rock profile roughness in terms of joint roughness
coefficient (JRC) (Barton & Choubey, 1977). In this test, individuals involved in
geotechnical engineering were asked to visually assess the JRC values of three surface
profiles obtained from the same granite block; the results are presented in Figure 6.
Through various statistical hypothesis tests, the authors concluded that the
observations could not be defined by a specific stochastic function. This demonstrates
that rock mass parameters derived through expert judgement may be epistemic, rather
than aleatory.
It is important also to recognise that, in this study, the number of participants
and hence estimates was high (in the region of 122-125). In general this will not be
the case. For example, in practice a single or small team of design engineers would
agree on a value or range of values of JRC to be adopted for design. This is likely to
introduce subjectivity into the characterisation process, and, unless an appropriate
model is used to capture the uncertainty, may neither adequately represent the
epistemic uncertainty nor provide appropriate parameter values (Crawford et al.,
2006). Indeed, Ang & Tang (2006) state: Epistemic uncertainty is highly relevant to
geotechnical engineers, who rarely have sufficient test data with which to justify a
statistical approach to parameter selection.
Eurock 2012 Page 6
In this example there is sufficient test data to attempt a statistical analysis. Having
done so, the original authors found that the mean and standard deviation of the data
fluctuated until 50 or so estimations had been made (Figure 7). This supports the
earlier statement (represented by Figure 4) that epistemic uncertainty is reducible,
because the collection of more information or an increase in knowledge helps decrease
the level of uncertainty. Although Milne (1990) suggested that better assessments of
JRC may be obtained by increasing the number of people used to estimate values, the
results shown here indicate that little benefit is accrued by using more than 50 people.
Nevertheless, the geotechnical engineer seldom has the luxury of vast amounts of data
or the availability of a large group of experts to assist with subjective assessment.

Figure 7. Variation in JRC standard deviation with number of results (from Beer et al., 2002)
2.1.2 Type of uncertainty as a function of information
The epistemic uncertainty associated with small data sets is exemplified by the
variability of unconfined compressive strength (UCS) with respect to the number of
strength measurements made (Ruffolo & Shakoor, 2009). Five different rock types
were analysed by Ruffolo and Shakoor, with statistical analyses being undertaken on
subsets of test specimens to determine the minimum number of strength tests required
to render a reliable estimate of the average strength of the entire set of specimens.
Figure 8 presents typical results for one of the rock types tested, and shows
Number of estimations
S
t
a
n
d
a
r
d

d
e
v
i
a
t
i
o
n

Figure 6. Epistemic uncertainty in Joint Roughness Coefficient (from Beer et al., 2002)
50
F
r
e
q
u
e
n
c
y
(
%
)
100
0
Frequency histogram of estimations (ignoring experience level)
for profile B
Profile B
JRC
0-2 4-6 2-4 6-8 10-12 8-10 12-14 14-1616-1818-20

0-2 4-6 2-4 6-8 10-12 8-10 12-14 14-1616-1818-20
0
100
JRC
50
Profile C
Frequency histogram of estimations (ignoring experience level)
for profile C
F
r
e
q
u
e
n
c
y
(
%
)
F
r
e
q
u
e
n
c
y
(
%
)
Frequency histogram of estimations (ignoring experience level)
for profile A
Profile A
0-2 4-6 2-4 6-8 10-12 8-10 12-14 14-1616-1818-20
Eurock 2012 Page 7
uncertainty in UCS reducing with increasing number of specimens. This is to be
expected, and demonstrates that the central limit theorem (Spatz, 1993; Davis, 2002)
applies to these tests. However, if we simply consider the case of very small sample
sizes (e.g. five or fewer specimens), then such statistical considerations are invalid and
thus strength must be considered as epistemic (i.e. similar to the concept presented in
the example of Figure 3). These results demonstrate that, whilst UCS may be
intrinsically aleatory (resulting, for example, from variability within a rock layer),
unless sufficient data exist with which to characterise it, the use of an aleatory model
may be inappropriate. In such a case, uncertainty in UCS should be treated as
epistemic, and handled using an appropriate, non-stochastic, approach.
The work of Ruffolo and Shakoor also showed that strength variability and hence
the number of tests required to make adequate estimates of mean strength varies with
rock type, as shown in Figure 9. In this figure, the degree of anisotropy and
heterogeneity in the rock type (sandstone to schist) increases from left to right. This
suggests that there may be a geological link between variability and number of
samples required to reduce uncertainty, and implies that the minimum number of
strength tests required may not be the same for all rock types. If true, this will have
important ramifications for the codification of testing requirements in order to
characterise rock strength as aleatory.

Figure 9. Minimum number of samples needed to estimate the mean unconfined compressive
strength (after Ruffolo & Shakoor, 2009).
Coefficient of Variation
0
5 10 15 20 25 30
10
20
15
20
25
30
25% deviation
20% deviation
15% deviation
10% deviation
N
o
.

o
f

s
p
e
c
i
m
e
n
s
B
e
r
e
a
s
a
n
d
s
t
o
n
e
I
n
d
i
a
n
a
l
i
m
e
s
t
o
n
e
M
a
r
b
l
eM
i
l
b
a
n
k
g
r
a
n
i
t
e
Wissahickon
schist
INCREASING ANISOTROPY
AND INHOMOGENEITY

Figure 8. Confidence intervals and acceptable strength deviation of Milbank granite (after Ruffolo
& Shakoor, 2009)
Upper 95% confidence interval
Lower 95% confidence interval
Number of measurements in group
10 30 20 40 50 60
U
n
i
a
x
i
a
l

c
o
m
p
r
e
s
s
i
v
e

s
t
r
e
n
g
t
h

[
M
P
a
]
0
350
35
70
105
140
175
210
245
280
315
Eurock 2012 Page 8
3 LIMIT STATE DESIGN
LSD methodologies represent a means of applying probabilistically based methods to
routine design procedures. They have been used successfully in structural engineering,
but their application in geotechnical engineering, especially rock mechanics, has been
considered to be controversial and unproven (Christian, 2004). This statement is made
with reference to the preceding discussion, where the inappropriateness of stochastic
methods for epistemic uncertainty was demonstrated. The following sections present
discussion on the principles embodied in LSD and the difficulties in applying LSD to
rock engineering.
3.1 Aleatory principles of Limit State Design in accordance with EC7
The fundamental assumption embodied in LSD is that both the effect of actions (i.e.
loads) on, and resistance of, the structure are aleatory, and thus can be described by
statistical distributions (Becker & Moore, 2007). In EC7 these distributions are
defined as a function of the representative actions (
rep
F ) and characteristic material
properties (
K
X ), respectively (see Figure 10). LSD requires verification that the effect
of actions is less than the design resistance of the structure.
Uncertainties in the loads and resistance are accounted for through the
introduction of partial factors. Thus, in Figure 10,
F
is the partial factor accounting
for variability associated with the actions, and
M
is the partial factor accounting for
variability associated with the resistances. The magnitude of the partial factors used in
LSD are governed by the reliability of information and the probability and
consequence of failure and, in EC7, for homogeneous soil properties, have been
derived either by calibration with conventional total factors of safety or by semi-
probabilistic studies of resistance (Meyerhof, 1995). It is noteworthy that EC7 does
not present calibrated partial factors for fractured rock mass parameters. It is likely
that this is because the work necessary to develop the partial factors has never been
undertaken, but in the context of this paper it could simply be that the epistemic nature
of such properties precludes their determination.
3.2 Inadequacies of LSD approaches to rock engineering
As demonstrated in the benchmark papers of Becker (1996a and 1996b) and illustrated
in Figure 10, the concept of LSD was initially developed for engineering with man
made materials (e.g. materials employed in structural engineering) in which the

Figure 10. Both effects of actions and material resistance are considered as random variables in
geotechnical LSD (after Becker, 1996b)
E R
Distribution
defining the
effect of actions
Eurock 2012 Page 9
material properties follow known probability density distributions. The partial factors
encapsulated in LSD have an advantage in design: they can be determined and applied
to individual material properties and thus, at least partly, can incorporate not only
variability in the geotechnical properties but also other uncertainties (Meyerhof,
1982). However, as Becker (1996b) states, To date, only the variation in loads and
material strengths have been considered explicitly in reliability based design and LSD
using partial factors. The other sources of uncertainty are less amenable to systematic
treatment either because the information required to characterise them is lacking, or
by their nature, they do not lend themselves readily to statistical analysis. Here, in
mentioning the lack of knowledge, Becker is referring to epistemic uncertainty.
The partial factor method, embodied in EC7 and other LSD codes, is based on the
principle that variability in both load (i.e. Design effect of actions, in Figure 10) and
Design resistance follow a stochastic distribution, to which an appropriate partial
factor may be applied (Figure 10). However, for rock engineering designs, where the
distribution of load and resistance may be derived from input parameters exhibiting
epistemic uncertainty, this assumption may not be valid.
Whilst Figure 10 shows that LSD requires the variables to be defined by statistical
distributions, Becker (1996b) states that epistemic uncertainty cannot be handled using
conventional statistical means. Furthermore, as we have seen, rock mechanics
properties may either be intrinsically epistemic (i.e. we do not know enough about the
processes to be able to demonstrate an aleatory characteristic), or extrinsically
epistemic (there is a lack of information to quantify the aleatory characteristics). Thus,
using the discussion and examples presented above, Beckers statement can be
paraphrased as:
aleatory variability is the basis of the development of LSD, but quantification of
DIANE rock mass parameters, which exhibit epistemic uncertainty, in terms of
characteristic values and partial factors calibrated from stochastic methods is not
valid.
Therefore, we suggest that EC7, in its current form, may not be sufficiently
developed for application to rock engineering designs where the uncertainty is
epistemic. This leads to the question, how might epistemic uncertainty be defined
using alternative models such that it may be incorporated into LSD?
4 ALTERNATIVE MODELS TO DEAL WITH EPISTEMIC UNCERTAINTY
Probabilistic assessment of uncertainty has been in use for a number of years in a wide
range of fields (Oberkampf, 2005), but an essential prerequisite to its application is the
knowledge that the processes being modelled are indeed aleatory, and that the required
distribution functions are either known or can be adequately approximated. Harrison
& Hudson (2010) state: It is not clear that the first of these characteristics is true,
and the second may not be achievable.
Within engineering there is now widespread realisation that aleatory variability
and epistemic uncertainty require different modelling approaches (e.g.; Hoffman &
Hammonds, 1994; Ferson, 1996; Ferson & Ginzburg, 1996; Guyonnet et al., 1999,
Helton et al., 2004; Colyvan, 2008; Dubois & Guyonnet, 2011). In recognition of this,
investigation has commenced in many fields into the development and application of
techniques that properly represent epistemic uncertainty (Oberkampf et al., 2004).
In the context of uncertainty in rock engineering, Wenner & Harrison (1996)
introduced the level of information (LoI) concept and suggested that different
uncertainty models should be employed as knowledge increases (see Figure 11). The
LoI concept is based on the recognition that, for each modelling approach shown in
Figure 11, there is a particular amount of information required. The lowest amount of
information is associated with an uncertain parameter for which there is only a single
Eurock 2012 Page 10
value available; as more information becomes available, so higher modelling
approaches can be applied. Therefore, the available level of information defines an
upper bound for the techniques that can be used, with each technique itself being
defined by the minimum amount of information it requires.
5 APPLYING LSD TO ROCK ENGINEERING
We have shown that LSD principles require the underlying uncertainty in the
processes being modelled to be aleatory, but some uncertainty in rock mechanics is
epistemic. This suggests that there is a fundamental discrepancy between what rock
mechanics is, and what LSD requires it to be, and this, perhaps, is the reason for the
anecdotal evidence suggesting that EC7 is difficult to apply to rock engineering
designs.
In order to overcome this fundamental discrepancy, there are only a few
approaches open to us. Perhaps the trivial approach is to abandon LSD for rock
engineering, and continue with the customary approaches. Whilst appealing, this does
nothing to integrate rock engineering and structural designs. Another approach would
be to develop methods by which epistemic uncertainty may be approximated as
aleatory and thus incorporated in the current LSD paradigm. This may, in the short
term, be the most appropriate. Finally, we could develop a new LSD paradigm that
encompasses both epistemic and aleatory uncertainty. How and whether this might be
achieved is not clear.
Perhaps, in the immediate future, the pragmatic approach is as follows:
i) fully understand which aspects of rock mechanics and rock engineering are
genuinely aleatory;
ii) for those aspects that are extrinsically epistemic (i.e. epistemic simply because we
have never either collected sufficient information or understood the processes well
enough to quantify the aleatory characteristics), we need to either ensure the
required data are collected, or work to quantify the inherent variability and hence
determine appropriate partial factors; and
iii) for those aspects that are intrinsically epistemic, we should eschew LSD principles
and continue with the current load and strength factor approaches.
6 CONCLUSIONS
The discussion presented in this paper has shown that the inherent problem in
applying LSD principles to rock engineering is the issue of handling the uncertainty
associated with the geotechnical input parameters required to characterise fractured
rock masses.

Figure 11. Level of information concept (after Wenner & Harrison, 1996; Guo & Du, 2007)

L
e
v
e
l
o
f
i
n
f
o
r
m
a
t
i
o
n
Present
knowledge
Complete
knowledge
Complete
ignorance
Epistemic
uncertainty
Aleatory
uncertainty
Probabilistic analysis
Evidence theory
Possibility theory (fuzzy
set technique)
Interval analysis
Single value technique
(estimated deterministic value)
Deterministic value
Eurock 2012 Page 11
For rock mass parameters, some of this uncertainty may derive from inherent
random variability. This is known as aleatory variability. However, using two recent
studies on the characterisation of rock mass parameters, we have shown that rock mass
parameters may exhibit epistemic uncertainty. This arises from insufficient
knowledge, and may particularly when expert judgement is required to define the
parameters be subjective in nature. Similarly, epistemic uncertainty may be reduced
(but not completely eliminated) with the acquisition of additional information (such as
further field or laboratory tests), and if sufficient information is obtained it may
become justified to characterise the uncertainty via a stochastic model. Furthermore, if
the uncertainty is intrinsically either epistemic or aleatory, then obtaining further
information will not allow re-categorisation of the type of uncertainty. However,
further information may enable the designer to select a more appropriate uncertainty
model, i.e. start with interval analysis and progress to probabilistic analysis when the
level of information is sufficient to justify an aleatory model. Nonetheless, when the
uncertainty, or level of information pertaining to uncertainty, is epistemic, the use of
stochastic modelling cannot be justified.
The introduction of the limit state code Eurocode 7 (in Europe) for all
geotechnical engineering recognises the need for rock engineering designs to comply
with the LSD paradigm. This paradigm requires that both the effect of actions (i.e.
loads) on, and resistance of, the structure be aleatory in nature. We note that, although
EC7 uses calibrated partial factors to handle the probabilistic concepts, none are
presented for fractured rock mass parameters. We suggest that whilst this may be
because the work necessary to develop them has never been undertaken, in the context
of this paper it could simply be that the epistemic nature of rock mass parameters
precludes their determination.
Finally, recognising that some rock mechanics properties are epistemic and thus
not amenable to treatment via LSD, we have made pragmatic recommendations as to
how EC7 should be applied in these circumstances.
7 REFERENCES
Ang, A.H-S, & Tang, W.H. (2006) Probability concepts in engineering: emphasis on
applications in civil and environmental engineering (2
nd
ed.), London: John Wiley
& Sons Ltd.
Baecher, G.B. & Christian, J.T. (2003) Reliability and statistics in geotechnical
engineering. London: John Wiley & Sons Ltd.
Barton, N.R & Choubey, V. (1977). The shear strength of rock joints in theory and
practice. Rock Mech. 10(1-2), 1-54.
Becker, D. E. 1996a. Eighteenth Canadian geotechnical colloquium: Limit states
design for foundations .1. An overview of the foundation design process. Can.
Geotech. J., 33 (6), 956-983.
Becker, D. E. 1996b. Eighteenth Canadian geotechnical colloquium: Limit states
design for foundations .2. Development for the National Building Code of Canada.
Can. Geotech. J, 33 (6), 984-1007.
Becker, D. E. & Moore I. D. (eds.), 2007. Canadian foundation engineering manual
(4th ed.). Richmond, B.C: Canadian Geotechnical Society,
Beer, A. J., Stead, D. & Coggan, J. S. 2002. Estimation of the Joint Roughness
Coefficient (JRC) by visual comparison. Rock Mech. Rock Eng., 35 (1), 65-74.
Bond, A.J. & Harris, A.J. 2008. Decoding Eurocode 7. London: Taylor & Francis.
CEN (European Committee for Standardisation). 2004. Geotechnical Design: Part 1,
General Rules. EN-1997-1. Brussels, Belgium: CEN.
Christian, J. T. 2004. Geotechnical engineering reliability: How well do we know
what we are doing? J. Geotech. Geoenviron., 130 (10), 985-1003.
Eurock 2012 Page 12
Colyvan, M. 2008. Is probability the only coherent approach to uncertainty? Risk
Anal., 28 (3), 645-652.
Crawford, J., Neretnieks, I. & Moreno, L. 2006. Epistemic and aleatory uncertainty in
recommended, generic rock K-d values used in performance assessment studies.
Mater. Res. Soc. Symp. P., 932, 251-258.
Davis, J.C. 2002. Statistics and data analysis in geology (3rd edition). London: John
Wiley & Sons Ltd.
Dubois, D. & Guyonnet, D. 2011. Risk-informed decision-making in the presence of
epistemic uncertainty. Int. J. Gen. Syst., 40 (2), 145-167.
Feng, X. T. & Hudson, J. A. 2010. Specifying the information required for rock
mechanics modelling and rock engineering design. Int. J. Rock Mech. and Min.
Sci., 47 (2), 179-194.
Ferson, S. 1996. Automated quality assurance checks on model structure in ecological
risk assessments. Hum. Ecol. Risk Assess., 2 (3), 558-569.
Ferson, S. & Ginzburg, L. R. 1996. Different methods are needed to propagate
ignorance and variability. Reliab. Eng. Syst. Safe., 54 (2-3), 133-144.
Guo, J. & Du, X. P. 2007. Sensitivity analysis with mixture of epistemic and aleatory
uncertainties. AIAA Journal, 45, 2337-2349.
Guyonnet, D., Come, B., Perrochet, P. & Parriaux, A. 1999. Comparing two methods
for addressing uncertainty in risk assessments. J. Environ. Eng.-ASCE, 125 (7),
660-666.
Harrison, J.P. 2010. Rock engineering and engineering geology: what about EC7 and
DIANE? [Lecture] Geological Society of London, June 2010.
Harrison, J. P. & Hudson, J. A. 2010. Incorporating Parameter Variability in Rock
Mechanics Analyses: Fuzzy Mathematics Applied to Underground Rock Spalling.
Rock Mech. Rock Eng., 43 (2), 219-224.
Helton, J.C., Johnson, J.D. & Oberkampf, W.L. 2004. An Exploration of Alternative
Approaches to the Representation of Uncertainty in Model Predictions, Reliab.
Eng. Syst. Safe, 85, 3971
Hoffman, F. O. & Hammonds, J. S. 1994. Propagation of Uncertainty in Risk
Assessments - the Need to Distinguish between Uncertainty due to Lack of
Knowledge and Uncertainty due to Variability. Risk Anal., 14 (5), 707-712.
Hudson, J.A. & Harrison, J.P. 1997. Engineering Rock Mechanics: an Introduction to
the Principles, Oxford: Elsevier Science.
Kiureghian, A. D. & Didevsen, O. 2009. Aleatory or epistemic? Does it matter?
Struct. Saf., 31 (2), 105-112.
Meyerhof, G. G. 1982. Limit States Design in Geotechnical Engineering. Struct. Saf.,
1 (1), 67-71.
Meyerhof, G. G. 1995. Development of Geotechnical Limit State Design. Can.
Geotech. J., 32 (1), 128-136.
Milne, D. 1990. Standardized joint descriptions for improved rock classification. In:
Hustrulid, W. A. & Johnson, G. A. (eds.), Proceedings of the 31
st
US Symposium
on Rock Mechanics, Golden, CO. Rotterdam: Balkema. pp. 3541.
Oberkampf, W. L. 2005. Uncertainty Quantification Using Evidence Theory.
[Presentation] Advanced Simulation & Computing Workshop Error Estimation,
Uncertainty Quantification, and Reliability in Numerical Simulations. Available
at http://www.stanford.edu/group/cits/pdf/lectures/oberkampf.pdf
Oberkampf, W.L., Helton, J.C., Joslyn, C.A., Wojtkiewicz, S.F., and Ferson, S., 2004.
Challenge Problems: Uncertainty in System Response Given Uncertain
Parameters, Reliab. Eng. Syst. Safe., 85 (1-3), 11-19.
Priest, S. D. & Brown, E. T. 1983. Probabilistic Stability Analysis of Variable Rock
Slopes. T. I. Min. Metal. A, A1-A12.
Eurock 2012 Page 13
Ruffolo, R. M. & Shakoor, A. 2009. Variability of unconfined compressive strength in
relation to number of test samples. Eng. Geol., 108 (1-2), 16-23.
Spatz, C. 1993. Basic Statistics Tales of Distributions. California: Brooks/Cole
Publishing Company,
Uzielli, M. 2008. Statistical analysis of geotechnical data. In: Huang, A. & Mayne, P.
W. (eds.) Geotechnical and Geophysical Site Characterisation. London: Taylor &
Francis. pp. 173-194
Wenner, D. & Harrison, J. P. 1996 Techniques to cope with uncertain parameters in
geomechanics on different levels of information. Proc. Eurock '96, Torino, Italy,
Vol 1, 327-334.
Zhang, L. & Einstein, H. H. 1998. Estimating the mean trace length of rock
discontinuities. Rock Mech. Rock Eng., 31 (4), 217-235.

You might also like