You are on page 1of 22

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225166928

Rock Mass Characterization and Rock Property Variability Considerations for


Tunnel and Cavern Design

Article  in  Rock Mechanics and Rock Engineering · July 2011


DOI: 10.1007/s00603-011-0138-5

CITATIONS READS

84 1,479

1 author:

M. Cai
Laurentian University
140 PUBLICATIONS   4,717 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A study of pillars instability in deep underground mines (NSERC Discovery Grants Program) View project

Time-dependent deformation behavior of brittle rock View project

All content following this page was uploaded by M. Cai on 05 June 2014.

The user has requested enhancement of the downloaded file.


Rock Mech Rock Eng (2011) 44:379–399
DOI 10.1007/s00603-011-0138-5

ORIGINAL PAPER

Rock Mass Characterization and Rock Property Variability


Considerations for Tunnel and Cavern Design
M. Cai

Received: 7 February 2011 / Accepted: 5 March 2011 / Published online: 23 March 2011
Ó Springer-Verlag 2011

Abstract Geotechnical design input parameters, such as and it can assist us to better understand how uncertainty
in situ stress field, rock mass strength parameters and arises and how the rock support system design decision
deformation modulus, are never known precisely. There may be affected by it.
are always uncertainties involved in these parameters,
some are intrinsic and others are due to lack of knowl- Keywords Variability  Uncertainty  Tunnel  Cavern 
edge or understanding of these parameters. To quantify Rock support  Rock mass strength  Deformation
the effects of these uncertainties on tunnel and cavern modulus  GSI system  Probabilistic design
design, it is necessary to utilize probabilistic analysis
methods. In the present study, a quantitative, probabilistic
approach to use the Geological Strength Index (GSI) 1 Introduction
system for rock mass characterization is presented. It
employs the block volume and a joint condition factor as The design of rock support systems is important in tunnel
quantitative characterization factors to determine the GSI and cavern design. Rock support system design practices
values. The approach is built on the linkage between for tunnels and large-scale underground caverns had been
descriptive geological terms and measurable field reviewed by many researchers such as Cording et al.
parameters such as joint spacing and joint roughness, (1971), Barton et al. (1977), Hoek and Brown (1980),
which are random variables. Using GSI values obtained Bieniawski (1984), Hönish (1995), Hönish and Nagel
from field mapping data, and in combination with the (1988), Cai et al. (2000), and Kaiser et al. (2000), etc.
intact rock strength properties, the probability density These datasets serve as empirical design references to
distributions of rock mass strength parameters and elastic which a contemporary design can be conducted or
moduli of the jointed rock mass can be calculated using compared.
Monte Carlo method. Furthermore, probabilistic analysis At the preliminary design stage, rock support systems
of tunnel and cavern stability based on the variable input for tunnels and caverns can be selected based on the
parameters is conducted employing the point estimate Q system (Barton et al. 1977), the RMR (Rock Mass
method. One example is given to illustrate how to con- Rating) system (Bieniawski 1984), or other empirical sys-
sider the variability of in situ stress and the rock mass tems. The empirical design approach is most effective if
properties in tunnel and cavern design. The method pre- the current tunnel/cavern size, shape, spacing between
sents an approach for systematic assessment of uncer- excavations, rock mass conditions are similar to precedent
tainty in rock mass characterization in rock engineering, case histories. The empirical design methods provide a
means to compare quantitatively different tunnel align-
ments or cavern layouts when only limited rock mass data
are available. They also provide a means to communicate
M. Cai (&)
between designers and contractors and to develop con-
School of Engineering, Laurentian University,
Sudbury, Ontario, Canada struction cost parameters, either for the comparative pur-
e-mail: mcai@laurentian.ca poses or for the development of a construction cost budget.

123
380 M. Cai

As a design base, rock mass classification is required when simplification; human omissions and errors. For example,
using the empirical design approach. the joint spacing varies significantly over space (spatial
However, some limitations in applying rock mass clas- variability), and the uniaxial compressive strength of the
sification systems for rock support design should be men- intact rock cannot be correctly predicted (randomness).
tioned. The RMR system is based on the case histories of Joint geometry distribution is three-dimensional but the
tunnels with relatively narrow spans. As reported by Pells measurement of joint geometry parameters is often made in
(2002), the Q system predicted significantly less support one-dimensional spaces (e.g., boreholes and scanlines) and
than actually adopted in several caverns he studied in in two-dimensional spaces (e.g., outcrops and tunnel
Australia. A recent discussion on the use and misuse of excavation walls), thus greatly limits our ability to know
rock mass classification systems can be found in Palmstrøm the real joint geometry properly. Hence, the suitable
and Broch (2006). In addition to the points that have been approach is to cope with uncertainties, assess and manage
made clearly by those researchers, it is observed that reli- the risk associated with uncertainties, i.e., to incorporate
ance on rock mass classification-based empirical methods uncertainties into the design and decision-making process.
as the sole basis for support design during bid document Mazzoccola et al. (1997) presented a method of apply-
preparation and detailed design can be inappropriate. Such ing information theory in rock mass characterization. The
systems do not promote the analysis and proper apprecia- probability distribution function (PDF) of RMR was
tion of the actual rock mass conditions such as can be computed as an example. This type of approach opens the
exposed in exploratory adits nor do they promote the door for considering variability of RMR in tunnel and
design approach of deductive reasoning based on field cavern design. As mentioned before, numerical tools need
observations. Most importantly, the variability of rock to be used along with the empirical design approach and in
mass properties cannot be properly considered in design recent years, the generalized Hoek–Brown strength crite-
using the empirical methods. In designing tunnels and rion (Hoek et al. 2002) is widely used and is now available
caverns of large spans, tunnels and caverns in highly in various numerical analysis tools, in association with the
stressed and difficult grounds, where no precedent experi- Geological Strength Index (GSI) system (Hoek et al. 1995).
ences exist, the numerical design approach is deemed The GSI system is a rock mass classification system that is
necessary. The empirical design approach must be sub- directly linked to engineering parameters such as Mohr–
stantiated with the design approach utilizing numerical Coulomb, Hoek–Brown strength parameters or rock mass
methods. modulus. Unfortunately, it is observed that very often the
Furthermore, in tunnel and cavern design, there are engineering condition of a rock mass in many applications
various scenarios or mechanisms, through which the was solely quantified with a single GSI value, and a
excavation may fail, as well as the complex rock mass question of great importance is how much uncertainty in
properties that need to be identified. Having identified the the analysis result is associated with this value.
limiting states of greatest concern, the next step is to In the present paper, we propose a probabilistic design
identify and attempt to quantify those geotechnical uncer- approach which utilizes the GSI system to characterize the
tainties upon which engineering performance depends. The rock mass and obtain input parameters for design using
engineers have to advance from the vagueness and quali- numerical methods. Using the quantitative GSI chart pro-
tative concept of uncertainty toward the more specific posed by Cai et al. (2004b), we can achieve the objective to
concept of probability. This is important to underground reflect rock mass property uncertainties in the numerical
construction engineering in particular because rock engi- analysis of tunnel/cavern stability. We will demonstrate,
neering practice faces many types of uncertainties. through an example, how to thoroughly study the likely
In rock engineering practice, data, because of huge cost failure mode, plastic zones and other instability problems
involved for their acquisition, are often incomplete and associated with rock support system design with the help of
hence contain uncertainties. Uncertainties are inherent and site mapping, laboratory testing and numerical modeling
unavoidable in the rock mass classification process. before recommending the final support system. Further-
Uncertainties are divided into several types: (a) uncertain- more, we will show how to properly consider and propa-
ties attributed to inherent randomness, natural variation, gate uncertainties in rock mass properties using the Monte
etc.; (b) uncertainties attributed to lack of data, knowledge Carlo method and consider the impact of input parameter
about events and processes; (c) uncertainties due to our variability on the output parameter using the point estimate
inability to understand decision objectives. Common method (Rosenblueth 1981), and reflect the quantitative
sources of uncertainties in rock engineering include the risk assessment results in tunnel and cavern design. In this
spatial and temporal variability of the rock mass properties; fashion, the design engineers are able to address the chal-
random and systematic errors in data mapping, logging, lenging problem of achieving a balance between the safety
testing, and monitoring; analytical and numerical model and economy in the design of rock support systems.

123
Rock Mass Characterization 381

2 Rock Mass Characterization Using the GSI System


where mi is a Hoek–Brown constant for the intact rock, D is
a factor that depends on the degree of disturbance to which
2.1 GSI System
the rock mass has been subjected by blast damage and
stress relaxation.
The rock mass strength and deformation modulus are
required as inputs to analyze the rock mass behavior by
2.3 Deformation Parameter
numerical models. The determination of the global
mechanical properties of a jointed rock mass remains one The mean deformation modulus, for rc \ 100 MPa, is
of the most difficult tasks in the field of rock mechanics.
related to the GSI value as (Hoek et al. 2002)
Because there are so many parameters that affect the  rffiffiffiffiffiffiffiffi
strength and deformability of an arbitrary rock mass, it is D rc ðGSI10
E ¼ 1 10 40 Þ ðGPaÞ ð5Þ
generally impossible to develop a universal law that can be 2 100
used in any practical way to predict the strength of the rock
For rc [ 100 MPa, the deformation modulus is
mass. Traditional methods to determine these parameters
estimated from
include in situ block shear tests for strength parameters and
E ¼ 10ð Þ ðGPaÞ
GSI10
plate-loading tests for deformation modulus. These tests 40 ð6Þ
can only be performed when the exploration adits are
where D = 0 has been assumed. When GSI is greater than
excavated and the cost of conducting these tests is high.
80, Eqs. 5 and 6 tend to overestimate the deformation
A few attempts have been made to develop methods to
modulus. In this circumstance, the equation proposed by
characterize the jointed rock mass to estimate the strength
Bieniawski (1978) can be used, i.e.,
and deformability indirectly. The GSI system, developed
by Hoek et al. (1995), is one of them. It uses properties of E ¼ 2GSI  100 ðGPaÞ ð7Þ
intact rock and jointing to determine/estimate the rock
in which the RMR value (RMR76) has been assumed to be
mass strength and deformability. GSI values can be esti-
equal to the GSI value (Hoek et al. 1995), for GSI [ 25.
mated based on the geological description of the rock mass
The inclusion of rc in Eq. 5 shows indirectly the influ-
and the GSI system consolidates various versions of the
ence of the modulus of the intact rock (Ei) on the defor-
Hoek–Brown criterion into a single simplified and gen-
mation modulus of the rock mass, because there is a good
eralized criterion that covers all of the rock mass types
correlation between Ei and rc (Deere 1968). A more recent
normally encountered in rock engineering. Using the GSI
update on the deformation of modulus of jointed rock mass
system, one can obtain a complete set of mechanical
is (Hoek and Diederichs 2006)
properties (Hoek–Brown strength parameters mb and s, or  
the equivalent Mohr–Coulomb strength parameters c and / 1  D=2
E ¼ Ei ð8Þ
, as well as elastic modulus E) for design purpose. 1 þ eðð75þ25DGSIÞ=11Þ

2.4 Determination of GSI Based on Block Volume


2.2 Rock Mass Strength Parameters
and Joint Surface Condition Factor
The generalized Hoek–Brown criterion (Hoek et al. 2002)
The GSI system has been developed and evolved over
for jointed rock masses is
many years based on the practical experience and field
 a
r3 observations. GSI is estimated based on the geological
r1 ¼ r3 þ rc mb þ s ð1Þ descriptions of the rock mass involving two factors, rock
rc
structure or block size and joint or block surface condi-
where mb, s, a are constants for the rock mass, and rc is the tions. Although careful consideration has been given to the
uniaxial compressive strength of the intact rock. Once the precise wording for each category and to the relative
GSI value is known, the Hoek–Brown strength parameters weights assigned to each combination of structural and
mb, s, a are given as surface conditions, the use of the original GSI chart
  involves some subjectivity. Hence, long-term experience
GSI  100
mb ¼ mi exp ð2Þ and sound judgment are required to successfully apply the
28  14D
  GSI system.
GSI  100 To facilitate the use of the system, Cai et al. (2004b)
s ¼ exp ð3Þ
9  3D presented a quantitative approach that employed the block
1  volume Vb and a joint surface condition factor Jc as
a ¼ 0:5 þ eGSI=15  e20=3 ð4Þ quantitative characterization factors. The quantitative
6

123
382 M. Cai

approach was validated using field test data and applied to where si and ci are the joint spacing and the angle between
the estimation of the rock mass properties at two cavern joint sets, respectively (Fig. 2). The assumptions included
sites in Japan. The quantified GSI chart is presented in in the block volume calculation are that three joint sets are
Fig. 1. present and joints are persistent.
If the joints are not persistent, i.e., with rock bridges, the
2.5 Block Volume rock mass strength is higher and the global rock stability is
enhanced. Strictly speaking, the joint persistence should be
Block size, which is determined from the joint spacing, defined using the area obtained from field survey. Since
joint orientation, number of joint sets and joint persistence, this is difficult, the joint persistence can only be approxi-
is an extremely important indicator of rock mass quality. mated by the measurement of the trace length on the rock
Block size is a volumetric expression of joint density, surface exposures. A joint persistence factor pi is defined as
which can be calculated as (
li
s1 s2 s3 pi ¼ L li \L ð10Þ
Vb ¼ ð9Þ 1 li  L
sin c1 sin c2 sin c3

Fig. 1 Quantification of GSI


chart (Cai et al. 2004b)

123
Rock Mass Characterization 383

condition factor to calculate the GSI value and hence the


Hoek–Brown strength parameters and deformation modu-
lus of the jointed rock mass. In other words, the Hoek–
Brown strength parameters and deformation modulus can
be directly expressed as a function of Vb and Jc. The
closed-form solution to obtain the GSI value from depen-
dent variables such as joint spacing, orientation, persis-
tence, surface condition factor, etc., makes it suitable for
probabilistic analysis using the Monte Carlo method.
To apply the GSI system for rock mass characterization,
two groups of parameters need to be determined. One is the
intact rock parameter, which includes rc and mi. Another
group is the joint parameter, which is further divided into
Fig. 2 Block volume delineated by joint sets the joint geometry and strength parameters. All these
parameters can be considered as random variables. In the
where li is the accumulated joint length of set i in the next section, we will discuss the typical probability dis-
sampling plane, respectively, and L is the characteristic tributions of these parameters and present methods to
length of the rock mass under consideration. The derive the probability distributions of output variables that
equivalent block volume is given as (Cai et al. 2004b) depend on these input variables.
s1 s2 s3
Vb ¼ p ffiffiffiffiffiffiffiffiffiffiffiffiffi
3 p p p  sin c sin c sin c
ð11Þ
1 2 3 1 2 3
3 Consideration of Variability of Geotechnical
The validity of the concept of equivalent block volume
Parameters in Design
for non-persistent joints has been verified by the block
system simulation using 3DEC (Kim et al. 2007).
In the present study, we focus only on the variabilities of in
situ stress field and the mechanical properties of jointed
2.6 Joint Condition Factor
rock masses, which are mainly due to the inherent ran-
domness or natural variation. The probability distributions
In the GSI system, the joint surface condition is defined by
of all input parameters are discussed.
the roughness, weathering and infilling condition. The
combination of these factors defines the strength of a joint
3.1 Variability in rc and mi
or block surface. This joint condition factor, Jc, is defined as
JW  JS To use the Hoek–Brown strength criterion, rc and mi for
Jc ¼ ð12Þ
JA the intact rock are needed. These two parameters can be
where JW and JS are the large-scale waviness and small- determined from a set of triaxial tests on carefully prepared
scale smoothness (Barton and Bandis 1990; Palmstrøm core samples. Recently, a simple but yet practical method
1995) and JA is the joint alteration factor (Barton et al. to estimate mi values of brittle rocks was proposed by Cai
1974). The ratings for JW, JS, and JA can be found in Cai (2010). Compared with the traditional approach which
et al. (2004b). requires triaxial test data, the proposed method needs only
data from uniaxial compression test with either volumetric
2.7 Calculating GSI from Vb and Jc strain measurement or acoustic emission (AE) monitoring
to detect the crack initiation stresses.
In numerical model implementation, it is sometimes trou- Statistically, a Weibull distribution has been used by
blesome to refer to the GSI chart (Fig. 1) for the determi- some researchers (e.g., Tang and Kaiser 1998) to describe
nation of the GSI values. Based on the proposed the rock strength distribution due to heterogeneity in the
quantitative chart, and using surface fitting techniques, Cai small-scale (e.g., grain size); in the large-scale, a normal
and Kaiser (2006) proposed the following equation for the distribution with the mean and the coefficient of variation
calculation of GSI from Jc and Vb (COV) can be used to describe the probability distribution
of rc and mi. COV is a fairly stable measure of variability.
26:5 þ 8:79 ln Jc þ 0:9 ln Vb
GSI ¼ ; ð13Þ Uncertainty for rc varies from low to high, and a COV of
1 þ 0:0151 ln Jc  0:0253 ln Vb
25% can be encountered. The uncertainty for mi is usually
where Jc is a dimensionless factor and Vb is in cm3. The small, with COV about 5–10% for most rock types (Hoek
user needs only to supply the block volume and joint and Marinos 2000a, b). This is because mi in the

123
384 M. Cai

Hoek–Brown strength criterion describes the frictional distribution approximates well the joint spacing distribution,
strength component. Test results have shown that the var- as shown in Fig. 3b. In this study, we recommend to use the
iability of the friction angle is small (Koyama et al. 1997), lognormal distribution for joint spacing. It is seen that the
with a COV of the same order as mi (5–10%). block volume will have a lognormal distribution if the joint
spacings of all three joint sets follow lognormal distributions.
3.2 Variability in Joint Spacing
3.3 Variability in Joint Orientation
Joint spacing is the single most important parameter that
defines the block size of a rock mass, which in turn Joint orientation, which is usually defined by dip direction
determines the rock mass quality. Joint spacing is defined and dip angle, affects the block shape and size (see Eq. 9).
as the perpendicular distance between adjacent joints and it Joint orientations are stochastic but quite often are clus-
determines not only the size but also the shape of blocks tered in preferred orientations to form joint sets. The joint
constituting the rock mass. Priest and Hudson (1976) stated orientation variability is thus governed by the degree of
that statistically, joint spacing follows a negative expo- orientation clustering within each set.
nential distribution. This is supported by our logged joint Joint orientation data can be collected from field map-
spacing histogram shown in Fig. 3a. The data were col- ping of tunnel walls and faces or oriented cores from
lected at a mine site in Sudbury, Ontario, Canada, from a boreholes. Tools such as DIPS (from Rocsicence Inc.) can
borehole drilled from the surface. The frequency of a given be used to identify joint clustering. The von Mises–Fisher
joint spacing is defined by f ðsÞ ¼ keks , where k & 1/
s is distribution is often used to describe joint orientation dis-
the mean joint frequency and s is the mean joint spacing. tribution. The Fisher distribution is a symmetric distribu-
Note that the bin size used for the histogram analysis was tion about the mean orientation and the probability density
0.25 m. The negative exponential distribution had been function is given by
used by Priest and Hudson (1976) to derive the relationship k
between RQD and k, i.e., f ðaÞ ¼ expðk cos aÞ sin a; ð15Þ
expðkÞ
0:1k
RQD ¼ 100e ð0:1k þ 1Þ ð14Þ
where a is the angle between the fracture normal vector and
in which a threshold of 0.1 m is assumed. the vector of its mean orientation and k is a parameter.
However, some researchers consider that the joint spacing Figure 4 presents one example showing the Fisher joint
distribution is logarithmic. If the interaction of jointing orientation distribution obtained from a borehole logging
corresponds to a multiplicatory process, lognormal distri- using oriented cores. Because the borehole only intersects
bution may result (Dershowitz and Einstein 1988). We find joint sets that are not sub-parallel to the drill hole, only two
that the type of distribution seems to be affected by the major joint sets are identified in this case. To identify all
minimum bin size used in the histogram analysis. When the major joint sets and their spacings at one size using bore-
same data set presented in Fig. 3a was analysed, using a hole logging technique, it is, therefore, important to have
minimum bin size of 0.005 m, we found that a lognormal boreholes orientated in different directions.

(a) (b)
Fig. 3 Joint spacing distribution: a bin size 0.25 m leading to a negative exponential distribution; b bin size 0.005 m leading to a lognormal
distribution

123
Rock Mass Characterization 385

Fig. 5 Joint length distribution (Song and Lee 2001)


Fig. 4 Joint orientation distribution determined from a borehole
logging with oriented cores
measure of the joint strength against shearing. The wavi-
3.4 Variability in Joint Size ness is measured by the undulation expressed as a per-
centage. According to Barton and Bandis (1990), both the
Joint sizes and trace lengths vary over a wide range. Sev- large-scale (waviness) and the small-scale (unevenness)
eral distributions, such as negative exponential, lognormal, roughness can be estimated by the amplitude of the
hyperbolic, Gamma-1 distributions, have been proposed to asperities. The joint alteration factor, which covers the joint
describe joint trace length distribution. To some extent, the infilling and weathering, has the most impact on the joint
reason that so many distribution types were used to condition factor as it can reduce Jc by more than one order
describe joint length distribution may be attributed to the of magnitude. In reality, because the small-scale asperities
fact that one observes joint surface traces and not actual have a base length of some centimeters and the amplitudes
sizes (Dershowitz and Einstein 1988). In the field, this are in the order of hundreds of millimeters that are difficult
quantity can only be measured from the exposure of the to measure, a descriptive rating system is often used. The
rock so that the real length or size of a joint can only be a same strategy can be applied to the rating of the joint
guess. The variability in joint size thus includes the alteration.
inherent natural randomness and our inability to measure or In practice, when the joint small-scale smoothness is, on
model this property, rendering it one of the most difficult average, ‘‘rough,’’ there are possibilities that some portions
parameters in joint system modeling. of the joint are ‘‘very rough’’ while other portions are
Priest and Hudson (1981) used the negative exponential ‘‘slightly rough.’’ This uncertainty can result from both
distribution to describe joint length distribution. Joint spatial variability of the joint surface condition and human
lengths presented in Fig. 5, surveyed at an underground errors in field mapping. A decision, therefore, has to be
storage cavern for liquefied petroleum gas in Korea, show a made about the rating range and the distribution type. For
lognormal distribution (Song and Lee 2001). Again, we simplicity, we have used the normal distribution to repre-
think that the issue of using either negative exponential or sent the joint roughness and alteration variability, with a
lognormal distribution is similar to the same issue of bin mean rating and a standard deviation determined by the
size and sampling details discussed in the previous section range divided by 6. When the mean values are near the
on joint spacing. Depending on the problem scale and bin extreme values in the rating, a truncated normal distribu-
size used, both negative exponential distribution and log- tion can be used. Figure 6 presents one example of the
normal distribution are appropriate to describe the joint truncated normal distribution of the JA parameter. The
length distribution. The joint length or persistence infor- average and standard deviation of JA are 1 and 0.1,
mation is used to calculate the equivalent block volume respectively. Because the minimum and maximum values
size in the GSI system (Eqs. 10–11). of JA are 0.75 and 12 (Cai et al. 2004b), respectively, the
normal distribution is truncated at 0.75.
3.5 Variability in Joint Surface Condition
3.6 Variability in In Situ Stress
The joint surface condition is controlled by joint large-
scale waviness, small-scale smoothness, and joint alter- The stability of tunnels and caverns is controlled by not
ation. The joint condition factor defined in Eq. 12 is a only the rock mass strength but also the in situ stress

123
386 M. Cai

joint spacing, assessing the joint surface condition. All


these uncertainties are attributed to the inherent nature of
the rock mass and/or the errors made by site engineers or
laboratory technicians.
A probabilistic approach to tunnel and cavern design
analysis allows the observed natural variability of many
parameters in the rock mass characterization process to be
taken into consideration. Within the probabilistic frame-
work, the derived parameters such as the GSI value and the
depth of the yielding zone are regarded as random variables
following a particular probability distribution. The proba-
bility distribution of the output variables can be derived
from those of the input variables. As discussed above, the
Fig. 6 Truncated normal distribution of JA
rock mass strength and deformation parameters can be
determined from equations linking mb, s, and E directly to
orientations and magnitudes. Hence, knowledge of the in GSI, and GSI is linked directly to Vb and Jc, which are
situ stress field (both stress orientations and magnitudes) is again linked to joint spacing, orientation, length, rough-
essential in underground excavation design. In general, ness, etc. This is suited for the calculation of probability
uncertainties in the in situ stress measurement results are distributions of the output parameters from the known
high. Stress magnitude and orientation distributions, as probability distributions of the input parameters using the
determined from the overcoring or hydraulic fracturing Monte Carlo method. Once the strength and deformation
method, often show a large variability. modulus distributions have been determined, the proba-
Whereas budget and time permits, a number of in situ bility distributions of the yielding zones around the exca-
stress measurements can be carried out and histogram plots vations can be evaluated and proper rock support system
of principal stress frequencies and contour plots of prin- design conducted. To achieve this goal, stress analysis
cipal stress orientation can be made. One example from using a finite element method (FEM) or a finite difference
Brady and Brown (2004) is shown in Fig. 7. A normal method (FDM) program has to be performed considering
distribution can be assumed for both the magnitudes and the variability of all input parameters, including the vari-
orientations of the in situ stress components. In this ability of in situ stress. Because extensive numerical
example, the standard deviations of r01 and r03 are about 3 analysis is needed to calculate the probability distributions
and 3.3 MPa, or the COV are about 6 and 16.6%, respec- of output parameters, the point estimate method (Rosen-
tively. The principal stress orientations are found to be in a blueth 1981) best suits this purpose. These two sampling
range of ±5° to 10°. For one realization of the magnitudes, methods are briefly introduced in the following discussion.
it is required that the principal stresses are mutually
orthogonal. 3.7.1 The Monte Carlo Method
The reliability of the in situ stress field increases grad-
ually as more field measurement data are available and The Monte Carlo simulation method, which is based on
back-analyses are performed. Just like the uncertainties in randomized input, has been widely used in many engi-
rock mass strength and deformability, the uncertainties neering disciplines including geomechanics. When the
about the in situ stress can be attributed to the actual spatial distribution of each variable is known and the output var-
variation of the stress field at various locations, our iable can be explicitly related to the input variables, the
inability to know the parameters precisely, and measure- Monte Carlo method can be used. For example, Eqs. 9–13
ment error. Further discussion about the in situ stress var- can be used to calculate the distribution of GSI based on
iability can be found in Martin et al. (2003). the distributions of joint spacing, persistence, and surface
condition factors, using the Excel add-in program @RISK
3.7 Methods to Derive Probability Distributions (2001). Once the GSI distribution is obtained, Eqs. 2–8 can
of Output Variables be used to calculate the probability distribution of the
Hoek–Brown strength parameters (mb, s) and deformation
The uncertainties in rock mass mechanical properties and modulus (E).
in situ stress field require a probabilistic approach to design The disadvantage of the Monte Carlo method is that it
tunnels and caverns in rocks. As discussed above, the requires a large number of simulations to be executed. It
uncertainties can occur in determining the in situ stress also requires that the distributions of the input variables
field, defining the strength of intact rocks, mapping the should either be known or that they can be assumed, i.e.,

123
Rock Mass Characterization 387

Fig. 7 Distribution of principal


in situ stress components:
a magnitude; b orientation
(Brady and Brown 2004)

for the calculation to be automatic, explicit functions are strength increase (Kaiser et al. 2000). Because of the lack
needed. In addition, the large samples required in the of research in the covariance of the geotechnical parame-
analysis render it less appealing for certain applications ters interested to the tunnel and cavern design engineers,
such as stress analysis using FEM or FDM techniques. we assume, for the sake of simplicity, that all variables are
independent in the illustration example presented in the
3.7.2 The Point Estimate Method next section.

The point estimate method was proposed by Rosenblueth


(1981) as a simple way to approximate the mean and 4 Application Example
standard deviation of a design parameter depending on
several other input parameters. The method is easy to use In the following discussion, we use field data from an
and requires little knowledge of probability theory. It is in- underground powerhouse to demonstrate the probabilistic
line with the practice of approximating complicated func- design approach using the GSI system.
tions by a series of simpler functions.
If there are n random variables, 2n solutions are needed 4.1 Kazunogawa Underground Powerhouse
to find the mean and standard deviation of the targeted
quantity. When the skewness is zero or negligible, the Kazunogawa power station (Koyama et al. 1997), located
procedure chooses 2n points selected so that the value of in Yamanashi Prefecture, Japan, at about 500 m depth and
each variable is one standard deviation above or below its has a generating capacity of 1,600 MW. The cavern
mean. The point estimate method is mathematically well dimensions are width 34 m, height 54 m, and length
founded and robust as well as accurate for a range of 210 m. The cavern excavation was started in 1994 and the
practical applications. We will use it for the stability last bench was excavated in 1996.
analysis of tunnels and caverns using numerical models Extensive field mapping and in situ testing were carried
which consider the variability of design input parameters. out in the exploration tunnels. The rock mass consists of
In the present study, the probabilistic parameters are sandstone and composite rock of sandstone and mudstone,
assumed to be independent for simplicity. Some of the described as two groups (CH and CM) of rock mass types
variables encountered in the above discussion may be based on the Denken system (Tanaka 1964). 75 uniaxial
dependent and the covariance between random variables compressive tests of intact rocks were conducted. Three
plays an important role in the analysis using probabilistic orthogonal joint sets were observed from field mapping.
approach. For example, the Hoek–Brown strength param- The joint spacing of the major joint set is in the range of
eters mb and s could be dependent. mb and s denote the 1–20 cm. The average joint spacings of the other two joint
frictional and cohesion strength components, respectively. sets are 25 and 50 cm, respectively. Joints are fresh, have
As shown in the strength models of brittle rocks, the pre- small undulation and are rough. The joint surface assess-
and post-peak rock strength mobilization process is asso- ment is supported by joint profiles obtained using a laser
ciated with the cohesion strength loss and frictional scanner in laboratory tests. The rock grouping (CH and CM)

123
388 M. Cai

and the block sizes are basically controlled by the joint


frequency of the major joint set. From joint mapping data,
it is seen that the average joint spacing is about 10 cm for
CH and about 2.5 cm for CM.

4.2 Rock Mass Properties and Their Variability

In this section, the GSI system is firstly used to estimate the


mechanical properties of the rock mass at the Kazunogawa
cavern site. As it will be seen later, the quantitative
approach of the GSI system allows us to consider the
variability of the rock mass strength and deformation
parameters. The estimated values, both mean and standard Fig. 8 Block volume Vb distribution simulated using @RISK
deviation, are then compared with the in situ test data to
demonstrate the applicability of the GSI system to jointed alteration JA are 2, 2, 1, respectively, and the coefficients of
rock masses. Based on the estimated values, numerical variation for all three factors were assumed to be 8%.
simulation using an FEM program is carried out to study Truncated normal distributions were assumed for Jw, Js,
the influence of the uncertainties in the input parameters on and JA. The truncation was based on the minimum and
the yielding zone distribution and the deformation behavior maximum ratings for each parameter (Cai et al. 2004b). For
of the cavern. To provide a comprehensive understanding example, Jw should not be less than 1 and not greater than
of the uncertainties in two different categories, the effects 3. Jw was thus described by a normal distribution with a
of in situ stress and rock mass properties are studied sep- mean of 2 and a standard deviation of 0.16, truncated at 1
arately to reveal the relative importance of the variability and 3. Jc thus calculated using Eq. 12 also followed a
of these two important factors in underground excavation normal distribution as shown in Fig. 9.
design. Equation 13 was used to calculate the GSI distribution
As discussed above, most of the input variables based on Vb and Jc. Although Vb followed a lognormal
encountered in geomechanics, such as uniaxial compres- distribution, the calculated GSI values followed a normal
sive strengths, cohesions, friction angles, and Young’s distribution, as shown in Fig. 10. The average GSI was
moduli, usually follow a normal distribution. Other types 59.9 with a standard deviation of 2.1. The standard devi-
of distributions such as lognormal, Weibull, and Poisson ation of GSI was greatly affected by the deviation of joint
are applicable to some variables. For example, the joint spacing. If the standard deviations for the lognormal
spacing distribution can be described using a lognormal joint spacing are 4, 10, 20 cm, respectively, for the three joint
probability function. It should be noted that when only sets, the standard deviation of GSI will increase to 2.5.
limited data are available, the data set may only be best Next, the Hoek–Brown strength parameters (mb and
represented by a truncated distribution function. In addi- s) and the equivalent Mohr–Coulomb strength parameters
tion, when a variable depends on several input variables (c and /), rock mass tensile strength (rtm), and the defor-
whose distribution functions differ, the resulting distribu- mation modulus (E) are calculated based on GSI, rc and mi.
tion of the output variable should be determined by a data The uniaxial compressive test results for the intact rock
fitting process, also available in @RISK. showed a large variability in the dataset. The average of rc
In the following illustration example, we consider the
rock mass classification for the CH rock mass and assume
that the joint spacing follows a lognormal distribution. The
average joint spacings for the three joint sets are 10, 25,
50 cm and the standard deviations are 3, 7.5, 15 cm,
respectively. Using the Monte Carlo simulation technique
available in @RISK, the PDF of the block volume Vb for
the rock mass is calculated using 5,000 iterations and the
result is presented in Fig. 8. It is seen that the block volume
also follows a lognormal distribution. In the calculation, a
persistent joint condition was assumed. Field data also
support this assumption.
It was determined that the average values for joint large-
scale waviness Jw, small-scale smoothness Js, and joint Fig. 9 Jc distribution simulated using @RISK

123
Rock Mass Characterization 389

distributions illustrated in Figs. 12 and 13. The average


values of c and / are 2.23 MPa and 54.5°, respectively. For
comparison, the average values of c and / obtained from
12 in situ block shear tests are 1.5 MPa and 58°, respec-
tively. Figure 14 presents the strength envelope for the
rock mass. The shaded areas of the strength envelopes are
obtained assuming combinations of c? and /?, c- and
/-, respectively. Here, c? and c- mean the values of
average c plus or minus one standard deviation of c,
respectively. In general, the average envelopes from the
GSI system are very close to those obtained from in situ
shear tests.
Fig. 10 GSI distribution simulated using @RISK The probability density distributions for the Hoek–
Brown strength parameters mb and s are presented in
was 108 MPa with a standard deviation of 42 MPa. The Figs. 15 and 16, respectively. mb values follow a normal
test data range was from 36.5 to 253 MPa. A truncated distribution. Although s values are best described by a
normal distribution, truncated by a minimum value of lognormal distribution, they can also be reasonably
36.5 MPa and a maximum value of 253 MPa, was assumed approximated by a normal distribution as shown in Fig. 16.
for rc. The distribution of the probability density function The distribution for the tensile strength of the rock mass
of rc is presented in Fig. 11. The truncated function shows shown in Fig. 17 can be represented by a truncated normal
an asymmetrical distribution, which is best fitted by a distribution (the truncation is made at rtm = 0 MPa).
Weibull distribution. The truncation may affect the distri-
bution of dependent variables such as c and /. mi was
assumed to obey a normal distribution with an average of
19 and a standard deviation of 2.375. The distributions of
mb, s, c, /, rtm, and E were calculated based on the dis-
tribution of rc, mi, and GSI, where D = 0 had been
assumed in all calculations.
The probability density distributions for c and / are
plotted in Figs. 12 and 13, respectively. Both the best-fit
probability density functions of c and / follow Weibull
distributions. This is attributed to the truncated normal
distribution of rc. If rc follows a normal distribution
without truncation, then, the resulting c and / can be
described by normal distributions. Although Weibull dis-
tributions best fit the output distributions, it is seen that
c and / could be adequately described by the normal Fig. 12 c distribution

Fig. 11 rc distribution Fig. 13 / distribution

123
390 M. Cai

Fig. 17 rtm distribution

Fig. 14 Comparison of shear strengths of rock masses from the GSI


system and field test data

Fig. 18 E distribution

deviation of the deformation modulus determined from 29


in situ plate-load tests are 12.9 and 2.84 GPa, respectively.
Fig. 15 mb distribution
This suggests that the GSI system overestimates the mod-
ulus for CH rock mass by about 28%. Despite this, the GSI
system predicts the variability of the deformation modulus
well when compared with the distribution of plate-loading
test data. The average and standard deviation of the
strength and deformation parameters are summarized in
Table 1. Normal distributions have been used for all
parameters.

4.3 Consideration of In Situ Stress and Rock Mass


Property Variability in Design

It is seen from the results presented above that the


mechanical properties of jointed rock masses exhibit vari-
Fig. 16 s distribution ability. These properties are not best represented by aver-
age values, but have respective distributions about the
The probability density distribution of the deformation means, even under the perfect conditions. The design will
modulus E is presented in Fig. 18. As expected, E follows a make more sense if these input property distributions are
normal distribution with an average of 16.5 GPa and a properly considered in the design process. In the sub-
standard deviation of 2.7 GPa. The average and standard sequent discussion, we perform FEM stress analyses for the

123
Rock Mass Characterization 391

Table 1 Summary of the rock mass properties obtained from the GSI Table 2 Simulation cases for in situ stress uncertainty (Series-1)
system
Parameters
Average Standard deviation
Case # r01 (MPa) r03 (MPa) h (°)
mb 4.54 0.67
1 10.71 9.61 11.50
s 0.0119 0.00278
2 14.49 9.61 11.50
c (MPa) 2.23 0.56
3 10.71 13.00 11.50
/ (°) 54.4 2.56
4 14.49 13.00 11.50
E (GPa) 16.5 2.72
5 10.71 9.61 16.50
rtm (MPa) 0.28 0.118
6 14.49 9.61 16.50
7 10.71 13.00 16.50
Kazunogawa cavern to demonstrate the influence of the 8 14.49 13.00 16.50
material property variability on design.
The stress analysis is performed using Phase2 (2008). contracted Rocscience Inc. to add point estimate method
The in situ stress field and its variability play an equally capability in Phase2 (Valley et al. 2010), and the new
important role in affecting the outcome of the stress anal- version will be available in future release. In the following
ysis as the material properties. At the cavern site, two in discussion, the model runs are conducted manually.
situ stress measurements were carried out using the over- To limit the stress analysis cases and separate the
coring technique (Koyama et al. 1997). From one mea- influence of in situ stress from that of mechanical proper-
surement, the principal stresses in the plane perpendicular ties, we first conducted simulations considering the vari-
to the cavern axis were determined as 12.6 and 11.2 MPa, ability of the in situ stress only (Series-1). The average
respectively. The maximum principal stress inclined about mechanical properties were used in this case, i.e., rock
14° with respect to the vertical direction. From another mass strength and deformation modulus were treated as
measurement, the principal stresses in the plane perpen- deterministic variables. The total solutions were reduced to
dicular to the cavern axis were obtained as 14.2 and 23 = 8. The input parameters for all simulation cases are
9.6 MPa, respectively. The maximum principal stress listed in Table 2. Case-3 and Case-7 show that the mini-
inclined about 7.6° with respect to the vertical direction. mum principal stress is greater than the maximum principal
Obviously, there are large differences in both the in situ stress, which is physically unsound. It was decided that
stress magnitudes and orientations. In general, the confi- these two cases be rejected from the analysis list. The
dence on the in situ stress measurement results is low at the average strength and deformation modulus of the rock
initial design stage. The confidence can be gradually mass are listed in Table 1. The residual strength parameters
improved when field monitoring data such as displacement were considered as deterministic to reduce the simulation
measurements are available at the construction stage. At cases. The residual cohesion, friction angle, and dilation
the Kazunogawa site, back-analysis performed using the angle for all cases were assumed to be 0.5 MPa and 50°,
displacement monitoring data confirmed that the in situ 5°, respectively. The residual c and / values are deter-
stress field was close to the first measurement result mined based on the field block shear test results.
(Tasaka et al. 2000). In this paper, we decide to consider Next, the in situ stress field (1st measurement data) was
the variability of the in situ stress field at the initial design treated at deterministic and the rock mass strength and
stage. Based on the data presented in Fig. 7, it is assumed deformation parameters were considered as variables
that the COV for the principal in situ stress components (r01 which contain uncertainties (Series-2). The equivalent
and r03 ) is 15% and the stress orientation (h) variation range Mohr–Coulomb strength parameters c and /, the tensile
is ±7.5°. strength of the rock mass rtm, and the deformation modulus
Because the computation effort is huge using the Monte E were used in the simulation. The total solution case
Carlo method, it was decided to obtain the distributions of numbers were 24 = 16. The input parameters for all sim-
yield zone and displacement by the point estimate method. ulation cases are listed in Table 3. Again, the residual
The variables to be considered include c, /, rtm, E as well cohesion, friction angle, and dilation angle for all cases
as the principal stresses (r01 ,r03 , h). The mean and standard were considered as deterministic, assumed to be 0.5 MPa
deviation of the yielding zones and roof or wall displace- and 50°, 5°, respectively.
ments can then be calculated from 27 = 128 solutions. This The FEM mesh for the cavern excavation is presented in
still requires a significant amount of model runs. To Fig. 19. To reduce computation time, the arch excavation
facilitate such a design analysis using FEM tools, the was simulated in one excavation step and the lower cavern
Centre for Excellence in Mining Innovation (CEMI) has excavation was simulated in six bench steps instead of 17

123
392 M. Cai

Table 3 Simulation cases for rock mass mechanical property


uncertainty (Series-2)
Parameters
Case # c (MPa) / (°) E (GPa) rtm (MPa)

1 1.67 51.8 13.78 0.162


2 2.79 51.8 13.78 0.162
3 1.67 57.0 13.78 0.162
4 2.79 57.0 13.78 0.162
5 1.67 51.8 19.22 0.162
6 2.79 51.8 19.22 0.162
7 1.67 57.0 19.22 0.162
8 2.79 57.0 19.22 0.162
9 1.67 51.8 13.78 0.398
10 2.79 51.8 13.78 0.398
Fig. 19 FEM mesh used for the simulation (without rock support
11 1.67 57.0 13.78 0.398
system)
12 2.79 57.0 13.78 0.398
13 1.67 51.8 19.22 0.398
14 2.79 51.8 19.22 0.398
15 1.67 57.0 19.22 0.398
the sidewalls and 3.1 m in the roof. COV’s of the total
16 2.79 57.0 19.22 0.398
displacements are about 7–8%, with standard deviations in
the range of 2.3–3.3 mm.
The probability density distributions of yielding zones
benches (about 3 m each) actually executed during the and total displacements in the roof and on the sidewalls for
cavern construction. The rock support system was not Series-2 analysis are presented in Figs. 23 and 24,
considered in the stress analysis initially. respectively. The in situ stress field was considered as
Figure 20 presents one example showing the yielding deterministic and the rock mass mechanical property var-
zone distribution and the total displacement contour around iability was considered in this case. COV’s of the yielding
the cavern when the last bench excavation is completed. depth are 20% in the roof, 15% on the left sidewall, and
Both tensile and shear yielding of the rock masses are 21% on the right sidewall. The standard deviations of the
observed from the model simulation result. The yielding yielding zone depth are 0.68, 1.44, and 1.95 m for the roof,
depths are larger on the sidewalls than in the roof. The left sidewall, and right sidewall, respectively. It is seen that
same calculations were carried out for each case listed in due to the rock mass property variability, the predicted
Tables 2 and 3. The maximum yielding depths on the maximum and minimum depths of yielding zone can differ
sidewalls and in the roof were recorded for each simula- in a range of 11.9 m on the sidewalls and 4.1 m in the roof.
tion. The total displacements on the cavern wall were Compared with the variability due to in situ stress uncer-
also recorded (one in the roof and two on the sidewalls). tainty, the actual variation of rock mass mechanical prop-
Statistical analysis was then performed to reveal the erties has a more pronounced effect on the excavation
probability density distributions of yielding depth and response behavior.
displacement. In the rock support system design for the Kazunogawa
The probability density distributions of yielding zone cavern, the pre-stressed anchors designed for the cavern
and total displacement in the roof and on the sidewalls (at were 15 m long for the average ground (CH) and 20 m for
three sampling points indicated in the figures) for Series-1 the poor ground (CM) (Koyama et al. 1997). The deter-
analysis are presented in Figs. 21 and 22, respectively. The ministic design approach uses the average material
rock mass mechanical properties are considered as deter- properties. For example, based on the deterministic design
ministic and only the in situ stress variability is considered approach, the average yielding depths for the CH rock
in this case. COV’s of the yielding depth are about 6% on mass are 3.4, 9.8, and 9.5 m for the cavern roof, left
the sidewalls and 15% in the roof. The standard deviations sidewall, and right sidewall, respectively. If 15 m long
of the yielding zone depth are 0.51, 0.65, and 0.59 m for anchors are selected for the average quality ground, the
the roof, left sidewall, and right sidewall, respectively. It is deterministic approach indicates that the anchor coverage
seen that due to the uncertainty of the in situ stress at the is sufficient even if 3 m long grouting length (anchorage
initial design stage, the predicted maximum and minimum length in non-yielded rock mass) is considered. The fac-
depths of yielding zone can differ in a range of 3.9 m on tors of safety, defined as the pre-stressed anchor length

123
Rock Mass Characterization 393

Fig. 20 Yielding zone and total


displacement distribution
(shown as contour) in the rock
mass (Case-11, Series-2)

Fig. 21 Probability
distributions of yielding zones
in the roof and on the sidewalls
for analysis Series-1

divided by the average yielding zone depth, are 4.37, One design consideration in the support of large-scale
1.53, and 1.58 for the roof, left sidewall, and right side- caverns is that the anchors should be sufficient to cover the
wall, respectively. yielding zone depth around the opening, with additional

123
394 M. Cai

Fig. 22 Probability
distributions of total
displacement in the roof and on
the sidewalls for analysis
Series-1

properties are used, this design consideration can be sat-


isfied. However, when the rock mass property variability is
considered, the probability that the 15 m anchor may be
shorter than the yielding zone depth on the right sidewall is
0.19%, and the probability that the anchor’s 3 m anchorage
portion may fall in the yielding zone is 8.8%. Certain levels
of risk show up in the design. What does the probability
shown here mean? We know that probability is neither a
property of the world nor a property of the mind, but a
property of the model used to make sense of a physical
world, a property based on which we make decisions under
uncertainty. When a probability is translated into a verbal
description of uncertainty (Lichtenstein and Newman
1967), 1% is equivalent to ‘‘virtually impossible,’’ 10%
corresponds to ‘‘very unlikely,’’ and 50% means ‘‘fair
chance.’’ Accordingly, we can say that it is virtually
impossible that the lengths of the 15 m long anchors on the
right sidewall will be shorter than the yielding zone depth;
however, it is very unlikely that the anchorage portion of
the anchors will fall in the yielding zone.
Fig. 23 Probability distributions of yielding zones in the roof and on
the sidewalls for analysis Series-2. The anchor lengths adopted in the The anchor length in the roof is long (15 m) as com-
cavern constructs were 15 m for average ground and 20 m for below pared to the yielding zone depth, as shown in Fig. 23. The
average ground probability that the 15 m long anchor may be shorter than
the yielding zone depth in the roof and the probability that
anchorage length (e.g., 3 m for pre-stressed cables) beyond the anchor’s 3 m anchorage length may fall in the yielding
the yielding zone into the undisturbed rocks (Koyama et al. zone are all zero. In practice, the anchor length needs only
1997). As shown in Fig. 25, when the average rock mass to cover this yielding zone plus a reasonable additional

123
Rock Mass Characterization 395

Fig. 24 Probability
distributions of total
displacement in the roof and on
the sidewalls for analysis
Series-2

deviations in the range of 5.3–8.4 mm. The actual cavern


deformations in any given vertical or horizontal cross-
sections showed monitoring values with some variation
ranges (Koyama et al. 1997). For example, the actually
measured surface deformations at the middle elevation of
the sidewalls indicated in Fig. 24 were from 30 to 56 mm
on the penstock side and 15–40 mm on the tailrace side.
This variation can be due to the spatial rock mass property
variation. As can be seen from Fig. 24, the total displace-
ment on the right sidewall can vary in a range from 15 to
70 mm, depending on the rock mass condition. Hence,
when a field measurement is obtained, one should pay
Fig. 25 Probability distributions of the yielding zone depth on the careful attention to interpreting the measurement results
right sidewalls for analysis Series-2 (without rock support system) and the associated acceptance or rejection of the prediction
models. Quite often, design engineers are used to the
anchorage length unless there are structural-controlled deterministic design approach. When the field measure-
deep unstable blocks demanding longer cables. Cai et al. ment data do not agree with their model predictions, some
(2000) identified that while reinforcement lengths for of them feel ‘‘disappointed,’’ while others rush to a quick
cavern sidewalls are generally acceptable, unnecessarily conclusion stating that the input parameters might be
long rockbolts or anchors have been applied to many wrong. A probabilistic design approach can avoid such
cavern roofs around the world. Significant economic ben- unsound engineering judgments and distraction in the
efits could be achieved from shortening the reinforcement routine design work.
in the roof and further studies to confirm this conclusion The analysis presented above can be considered as the
are recommended. first step in rock support system design. After the rock
The probabilistic design analysis can assist in the support system has been chosen, it should be included in
interpretation of field monitoring results. COV’s of the the numerical model for design verification and refinement.
total displacements are about 20%, with standard When the selected rock support system is included in the

123
396 M. Cai

FEM analysis, it is anticipated that the distribution of the anchor’s 3 m anchorage length may fall in the yielding
yielding zone depths will change, i.e., become smaller zone is 0.5%. Now we confirm that the 15 m long anchor is
compared with the analysis cases where the rock support able to cover all the yielding zones around the cavern and
system is not included. The point estimate method outlined the probability that an anchorage portion of the cablebolt
above can be readily applied to the stress analysis including may fall into the yielding zone is sufficiently small that the
the rock support system. In this fashion, the risk level can original design of the rock anchor length is appropriate.
be recalculated. In tunnel and cavern design, the yielding of rock mass
To illustrate the approach, the actual rock support sys- indicates that the stress magnitudes have reached the rock
tem adopted for the CH rock mass (Koyama et al. 1997) mass strength, and the designers are provided with several
was included in the model for a refined analysis. The options to deal with the potential construction hazard
rockbolt and anchor lengths were 5 and 15 m, respectively. associated with rock mass yielding. If the rock mass’s post-
Support pressures were 0.213 MPa for the roof and peak strength does not drop, it can carry large loads even if
0.183 MPa for the side walls. The shotcrete thicknesses rock mass yielding occurs. On the other hand, if the rock
were 32 and 24 cm for the roof and sidewalls, respectively. mass’s post-peak strength drops significantly to a low
Figure 26 presents the yielding zone and the total dis- residual level after yielding occurs, the rock mass’s
placement contour around the cavern for Case-11 listed in integrity and load carrying capacity are reduced. Hence, it
Table 3. When the rock support system is considered in the is better to place the anchoring portion beyond the yielding
model, both the yielding zone and the total displacements zone. By ensuring that the reinforcement length is not
are reduced. shorter than the yielding depth is one of the commonly
The probability density function of the yielding zone applied design criteria, although it is conservative in gen-
distribution on the right sidewall is presented in Fig. 27. eral. For large-scale caverns and tunnels in weak ground,
The probability that the 15 m anchor may be shorter than the yielding may cause large displacement and hence
the yielding zone depth on the right sidewall is 0.0077%, another displacement or strain based design criterion exists.
and the probability that the anchor’s 3 m anchorage length A third design criterion is to check the loads in rockbolts
may fall in the yielding zone is 1.3%. In the same fashion, and anchors. At any given excavation stage, the loads
we can examine the probabilities for the left sidewall. It is should be below the design capacity of the reinforcements.
found that, on the left sidewall, the probability that the Sakurai (1981) proposed the concept of critical strain
15 m anchor may be shorter than the yielding zone depth and used it as a direct strain control technique for tunnel
on the left sidewall is 0.0003%, and the probability that the and cavern construction. The critical strain ranges

Fig. 26 Yielding zone and total


displacement distribution in the
rock mass (Case-11, Series-2,
with rock support system)

123
Rock Mass Characterization 397

σ
σcm Peak strength

0.8∼1.0σ AE damage
cm

Residual strength

0.4∼0.6σ AE initiation
cm

Accumulated AE

εc εf εa ε
Fig. 27 Probability distributions of the yielding zone depth on the Fig. 28 Definition of critical strain and failure strain in association
right sidewalls for analysis Series-2 (with rock support system) with crack initiation [the figure is based on the generalized crack
initiation and crack damage stress thresholds from Cai et al. (2004a)]
approximately from 0.1 to 1.0% for good rocks and from
1.0 to 5.0% for weak rocks and soils. Based on the concept
of the critical strain of rocks and rock masses by Sakurai the point with no further stress drop is of significant
(1981), we obtain the average critical strain ec for the CH importance in design calculations. Strain can increase
rock mass using the following equation without further stress decrease beyond this point. Hence,
the allowable strain ea can be related to the critical strain as
rcm
ec ¼ ð16Þ ec
E ea ¼ ð18Þ
1  Ra
where rcm and E are the average uniaxial compressive
strength and initial elastic modulus of the rock mass, where Ra is a parameter representing the residual strength
respectively. Based on the parameters listed in Table 1, the (see Fig. 28). Assuming Ra = 0.6, 0.65, 0.7, the allowable
average critical strain for CH rock mass is 0.084%. The strain thus calculated from the critical strain is 0.21, 0.24,
failure strain ef is related to the critical strain as 0.28%, respectively. Hoek (2001) suggests that if the tun-
nel strain is \1%, then, little or no construction instability
ec
ef ¼ ð17Þ problems will occur. The strain at this cavern site is far
1  Rf
\1%, suggesting that there will be no major instability
where Rf is a parameter representing the failure strength, problems at the cavern construction site (as was demon-
which lies in between 0.05 and 0.8 according to Sakurai strated by the actual successful completion of the cavern
(1981). construction).
The determination of Rf needs further discussion as the In some cases, deformation monitoring is conducted to
wide range of variation of this parameter suggested by detect wall deformation. The failure strains can be con-
Sakurai is not applicable to a specific case such as Kazu- verted into allowable wall displacements assuming an
nogawa cavern site. Based on the generalized crack initi- opening with an equivalent radius of (34 ? 54)/4 m. The
ation and propagation thresholds (Cai et al. 2004a, b), we allowable right wall displacements are 0.0464, 0.0531,
know that when the stress level is above 0.4–0.6 rcm, the 0.0619 m for Ra = 0.6, 0.65, 0.7, respectively, and the
stress–strain curve deviates from the linear relationship due probabilities that the wall displacement might be greater
to damage (or AE) initiation (Fig. 28). When the stress than the allowable displacements are 11, 1.7, and 0.0525%,
level is above 0.8–1.0 rcm, the damage coalescence and respectively (Fig. 29). In the same figure, the allowable
propagation generate additional strain to the rock mass. strain of 0.5% (which corresponds to Ra = 0.83) and the
Hence, the Rf parameter is at least 0.1–0.3 for most hard corresponding allowable displacement are shown. The
rock masses. Depending on the damage coalescence near average wall displacement is 0.0369 m, which is always
the peak strength, the Rf parameter can further increase to smaller than the allowable displacements for Ra [ 0.5. The
0.4. influence of Ra on the probability of failure in terms of
Note that some of the rock masses around the excavation strain is large and careful field monitoring programs are
will inevitably reach the peak strength. For most strain- needed to better define this parameter. In general, a large
softening rocks, the residual strength level governs the Ra value is expected when a rock mass is supported
amount of plastic deformation in the total deformation due because the residual strengths of well-supported rock
to excavation (Cai et al. 2007). The strain corresponding to masses are high.

123
398 M. Cai

Hoek–Brown strength parameters and deformation moduli


of jointed rock masses. The uncertainty in the rock mass
characterization utilizing the GSI system stems from the
variabilities in the strength of the intact rock, the joint
spacing, the joint surface condition.
This paper illustrates an approach for tunnel and cavern
design utilizing quantitative risk assessment with the use of
the GSI system to determine the required rock support and
to quantify uncertainty in the analysis. The Hoek–Brown or
the equivalent Mohr–Coulomb strength parameters and the
deformation modulus were calculated using the Monte
Carlo method in @RISK and FEM stress analyses were
performed applying the point estimate method. The method
Fig. 29 Probability distributions of the total displacement on the
right sidewalls for analysis Series-2 (with rock support system) was applied to characterizing the jointed rock masses at the
Kazunogawa underground powerhouse. Based on the dis-
tributions obtained for the rock mass properties, the yield-
In rock support design for large-scale caverns, another ing zone depth and displacement distributions were
design criterion often used is the measure of the rockbolt/ obtained. Risk analysis indicates that for the average quality
anchor loads. Using the same probabilistic design approach rock mass, the possibilities that the reinforcement length
for the examination of anchor lengths and displacements, (15 m) may not cover sufficiently all yielding zones on the
we can check the anchor load distribution due to rock mass sidewalls, are extremely small (‘‘virtually impossible’’ in
property variability and calculate the probabilities that the verbal probabilistic terms). The possibilities that the
anchor load might be larger than the design loads and anchorage portions of the anchors may fall into the yielding
yielding loads. Due to the length limitation of the paper, zone are very small as well and this uncertainty can be dealt
the results are omitted here. with using the observational construction method.
The probabilistic design approach identifies the risk The probabilistic method presents a not-so-black-and-
associated with design. The information obtained will white, systematic assessment of uncertainty in rock mass
assist us in better executing the plan using the observa- characterization in rock engineering design. It provides an
tional construction method. It is obvious that cost saving in approach to evaluate the effects of variability in each of the
terms of reducing the support quantities is achievable if a input parameters on the output ones. It also allows us to
certain level of risk is acceptable. By applying the method better understand how uncertainty arises and how the rock
outlined here, further cost-benefit analysis can be con- support system design decision may be affected by it, thus
ducted to reach a rational design of underground works. As revealing the key actions needed to reduce uncertainties in
shown in the illustration example, the benefit and the cost the most cost-effective way. Although it provides a tool for
of including risk can be weighed in the decision-making stress analysis, rock support design, and statistical esti-
process. mation of rock mass properties, its successful application
relies heavily on the professional judgment of experts, as is
always the case in rock mechanics and rock engineering.
5 Conclusion
Acknowledgments The method and analysis presented in this paper
were carried out to illustrate the probabilistic design approach and did
It is difficult to develop ‘‘precise’’ values to describe the
not necessarily reflect the design approach adopted by Tokyo Electric
rock mass properties. Instead, distributed data are often Power Services Co. Ltd. (TEPSCO) and Tokyo Electric Power
encountered. In the design approach using the deterministic Company (TEPCO). The author would like to thank TEPSCO and
method, engineers are often faced with difficulty when TEPCO for previous support on the application of the GSI system for
rock mass characterization.
selecting representative values from widely scattered data
for design. Application of the probabilistic design approach
to tunnel and cavern design makes it possible to consider References
variability in geotechnical parameters of jointed rock
masses. The rock mass property variability is both intrinsic Barton NR, Bandis SC (1990) Review of predictive capability of
and subjective and has a considerable influence on the JRC-JCS model in engineering practice. In: Rock Joints, Proc.
Int. Symp. on Rock Joints, Balkema, Rotterdam, pp 603–610
decision-making in the design works.
Barton NR, Lien R, Lunde J (1974) Engineering classification of rock
The GSI system is now widely used to derive engi- masses for the design of tunnel support. Rock Mech
neering design parameters such as the Mohr–Coulomb or 6(4):189–239

123
Rock Mass Characterization 399

Barton NR, Lien R, Lunde J (1977) Estimation of support require- Hönish K, Nagel KH (1988) Practical use of rock mass classification
ments for underground excavations. In: 16th US Symp. Rock for cavern & tunnel support. Int J Rock Mech Min Sci 35:4–5
Mech, pp 164–177 (Paper No. 131)
Bieniawski ZT (1978) Determining rock mass deformability— Kaiser PK, Diederichs MS, Martin CD, Sharp J, Steiner W (2000)
experience from case histories. Int J Rock Mech Min Sci Underground Works in Hard Rock Tunnelling and Mining. In:
Geomech Abstr 15(5):237–247 Keynote lecture at GeoEng2000, Technomic Publishing Co.,
Bieniawski ZT (1984) Rock mechanics design in mining and Melbourne, pp 841–926
tunneling. A.A. Balkema, Rotterdam, p 272 Kim BH, Cai M, Kaiser PK, Yang HS (2007) Estimation of block
Brady BHG, Brown ET (2004) Rock Mechanics for Underground sizes for rock masses with non-persistent joints. Rock Mech
Mining. Springer, Berlin Rock Eng 40(2):169–192
Cai M (2010) Practical estimates of tensile strength and Hoek–Brown Koyama T, Nanbu S, Komatsuzaki Y (1997) Large-scale cavern at a
strength parameter mi of brittle rocks. Rock Mech Rock Eng depth of 500 m. Tunn Undergr 28(1):37–45 (in Japanese)
43(2):167–184 Lichtenstein S, Newman JR (1967) Empirical scaling of common
Cai M, Kaiser PK (2006) Visualization of rock mass classification verbal phrases associated with numerical probabilities. Psychon
systems. Geotech Geol Eng 24(4):1089–1102 Sci 9(10):563–564
Cai M, Kaiser PK, Uno H, Tasaka Y (2000) Comparative study of Martin CD, Kaiser PK, Christiansson R (2003) Stress, instability and
rock support system design practice for large-scale underground design of underground excavations. Int J Rock Mech Min Sci
excavations. In: Proc. 4th North American Rock Mech. Sympo- 40(7–8):1027–1047
sium, Balkema, Rotterdam, pp 1027–1034 Mazzoccola DF, Millar DL, Hudson JA (1997) Information, uncer-
Cai M, Kaiser PK, Tasaka Y, Maejima T, Morioka H, Minami M tainty and decision making in site investigation for rock
(2004a) Generalized crack initiation and crack damage stress engineering. Geotech Geol Eng 15(2):145–180
thresholds of brittle rock masses near underground excavations. Palisade Corporation (2001) @RISK. Palisade Corporation, v.4
Int J Rock Mech Min Sci 41(5):833–847 Palmstrøm A (1995) RMi—a rock mass characterization system for
Cai M, Kaiser PK, Uno H, Tasaka Y, Minami M (2004b) Estimation rock engineering purposes. Ph. D. thesis
of rock mass strength and deformation modulus of jointed hard Palmstrøm A, Broch E (2006) Use and misuse of rock mass
rock masses using the GSI system. Int J Rock Mech Min Sci classification systems with particular reference to the Q-system.
41(1):3–19 Tunn Undergr Space Technol 21(6):575–593
Cai M, Kaiser PK, Tasaka Y, Minami M (2007) Determination of Pells PJN (2002) Developments in the design of tunnels and caverns
residual strength parameters of jointed rock masses using the in the Triassic rocks of the Sydney region. Int J Rock Mech Min
GSI system. Int J Rock Mech Min Sci 44(2):247–265 Sci 39(5):569–587
Cording EJ, Hendron AJ Jr, Deere DU (1971) Rock engineering for Priest SD, Hudson JA (1976) Discontinuity spacings in rock. Int J
underground caverns. In: Proc. ASCE Symp. on Underground Rock Mech Min Sci Geomech Abstr 13(5):135–148
Rock Chambers, pp 567–600 Priest SD, Hudson JA (1981) Estimation of discontinuity spacing and
Deere DU (1968) Geological consideration. In: Stagg KG, Zie- trace length using scanline surveys. Int J Rock Mech Min Sci
nkiewicz OC (eds) Rock mechanics in engineering practice. John Geomech Abstr 18(3):183–197
Wiley & Sons, New York, pp 1–20 Rocscience Inc. (2008) Phase2. Rocscience Inc., v.7
Dershowitz WS, Einstein HH (1988) Characterizing rock joint Rosenblueth E (1981) Two-point estimates in probabilities. J Appl
geometry with joint system models. Rock Mech Rock Eng Math Model 5(5):329–335
21(1):21–51 Sakurai S (1981) Direct strain evaluation technique in construction of
Hoek E (2001) Big tunnels in bad rock, 2000 Terzaghi lecture. underground opening. In: 22th US Symp. Rock Mech, MIT,
J Geotech Geoenviron Eng ASCE 127(9):726–740 Cambridge, pp 278–282
Hoek E, Brown ET (1980) Underground excavations in rock. Song JJ, Lee CI (2001) Estimation of joint length distribution using
Institution of Mining and Metallurgy, London, p 527 window sampling. Int J Rock Mech Min Sci 38(4):519–528
Hoek E, Diederichs MS (2006) Empirical estimation of rock mass Tanaka H (1964) Introduction of geology for civil engineers.
modulus. Int J Rock Mech Min Sci 43(2):203–215 Sankaidou, Tokyo
Hoek E, Marinos P (2000a) Predicting tunnel squeezing problems in Tang CA, Kaiser PK (1998) Numerical simulation of cumulative
weak heterogeneous rock masses. Tunn Tunn 32(11):45–51 damage and seismic energy release in unstable failure of brittle
Hoek E, Marinos P (2000b) Predicting tunnel squeezing problems in rock. Part I: Fundamentals. Int J Rock Mech Min Sci Geomech
weak heterogeneous rock masses. Tunn Tunn 32(12):34–36 Abstr 35(2):113–121
Hoek E, Kaiser PK, Bawden WF (1995) Support of underground Tasaka Y, Uno H, Omori T, Kudoh K (2000) A joint and rock failure
excavations in hard rock. A.A. Balkema, Rotterdam, p 215 strain-softening model and its application to the excavation
Hoek E, Carranza-Torres C, Corkum B (2002) Hoek–Brown failure simulation of large-scale underground caverns. J Jpn Soc Civil
criterion—2002 edition. In: Proc. 5th North American Rock Eng. III-51(652):73–90 (in Japanese)
Mech. Symposium, Toronto, pp 267–273 Valley B, Kaiser PK, Duff D (2010) Consideration of uncertainty in
Hönish K (1995) Conclusions from 100 constructed power caverns modelling the behaviour of underground excavations. In: 5th
for future planning: keynote lecture. In: Eurock’ 93, Balkema, International Seminar on Deep and High Stress Mining, Santi-
Rotterdam, pp 1013–1027 ago, pp 423–435

123

View publication stats

You might also like