Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: Reliability-based design (RBD) is being adopted by geotechnical design codes worldwide, and it is
Received 1 August 2018 therefore necessary that rock engineering practice evolves to embrace RBD. This paper examines the
Received in revised form Hoek-Brown (HeB) strength criterion within the RBD framework, and presents three distinct analyses
11 February 2019
using a Bayesian approach. Firstly, a compilation of intact compressive strength test data for six rock
Accepted 18 February 2019
Available online 13 April 2019
types is used to examine uncertainty and variability in the estimated HeB parameters m and sc, and
corresponding predicted axial strength. The results suggest that within- and between-rock type vari-
abilities are so large that these parameters need to be determined from rock testing campaigns, rather
Keywords:
Reliability-based design (RBD)
than reference values being used. The second analysis uses an extensive set of compressive and tensile
Hoek-Brown (HeB) criterion (both direct and indirect) strength data for a granodiorite, together with a new Bayesian regression
Bayesian regression model, to develop joint probability distributions of m and sc suitable for use in RBD. This analysis also
Indirect tensile strength shows how compressive and indirect tensile strength data may be robustly used to fit an HeB criterion.
Characteristic strength criterion The third analysis uses the granodiorite data to investigate the important matter of developing char-
acteristic strength criteria. Using definitions from Eurocode 7, a formal Bayesian interpretation of char-
acteristic strength is proposed and used to analyse strength data to generate a characteristic criterion.
These criteria are presented in terms of characteristic parameters mk and sck, the values of which are
shown to depend on the testing regime used to obtain the strength data. The paper confirms that careful
use of appropriate Bayesian statistical analysis allows the HeB criterion to be brought within the
framework of RBD. It also reveals that testing guidelines such as the International Society for Rock
Mechanics and Rock Engineering (ISRM) suggested methods will require modification in order to support
RBD. Importantly, the need to fully understand the implications of uncertainty in nonlinear strength
criteria is identified.
Ó 2019 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by
Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).
https://doi.org/10.1016/j.jrmge.2019.02.002
1674-7755 Ó 2019 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-
NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627 613
enlightening, because it shows how engineers were grappling with As a precursor to addressing the practical matter of rock mass
a number of issues related to the failure of complex structures strength in the context of RBD, here we concentrate on the
under ill-defined loading conditions. somewhat simpler problem of statistical analysis of intact rock
One paper by Tye (1944), with the memorable title “Factors of strength. For the case of applying the HeB criterion to intact rock
safety e or of habit?”, was seminal in the way it crystallised the strength data, Bozorgzadeh et al. (2018a) identified and dis-
problem. In the paper, Tye (1944) discussed emerging concepts of cussed some of the key shortcomings in current analysis
aeronautical design, and proposed that “The designer shall produce methods. They also strongly noted that, with regard to RBD,
evidence that the probable rate of accidents likely to cause injury to reporting in the literature of strength results is often incomplete.
personnel arising from structural failure is not more than once in ‘x’ Here, and following this work, we first summarise the key as-
thousands of hours of flying”. This statement succinctly encapsulated pects of regression models and then use regression to investigate
how designers should address aeronautical design, and soon led a number of data sets, covering various rock types, in the context
to the recognition that, in order to support it, designers would of RBD. Subsequently, we show how a combination of compres-
need to: sive and tensile strength data requires use of special regression
techniques, and use a previously developed Bayesian regression
(1) develop clear definitions of failure; model (Bozorgzadeh et al., 2018a) to analyse an extensive data
(2) determine acceptable failure rates; and set that contains these strength data. Finally, we discuss how
(3) collect substantial data related to material properties, regression models may be used in the context of modern design
loading and structural behaviour. standards, particularly in Eurocode, to obtain characteristic
values.
It was from the work such as this in the 1940s and 1950s that RBD
was developed, which shows clearly why RBD is based on probabi- 2. Regression models
listic principles and uses probability of a structure failing to perform
as required rather than factors of safety. Nevertheless, it was quickly Bozorgzadeh et al. (2018a) discussed how to formulate fre-
recognised that designers would need to be freed from the compu- quentist and Bayesian regression models for analysis of intact rock
tational burden associated with probabilistic calculation, and so a strength data. This work pays particular attention to the form and
simplified form of calculation based on factors (often referred to as location of variability in rock strength data, as illustrated in Fig. 1. It
partial factors) was developed. The values of these factors are set so is only through recognition of this variability that regression can be
as to ensure that the probabilistic requirements are met, and pre- performed correctly.
scribed in the structural design standards of limit state design (LSD) Customary regression models are based on frequentist princi-
and (North America) load and resistance factor design (LRFD). Un- ples, and as will be shown later, these models are inappropriate
fortunately, the abbreviations RBD, LSD and LRFD proliferate in the when handling combined compressive and tensile data. This,
literature, and although it has been noted that they mean different together with the need to robustly handle limited data, is why we
things to different people (Becker, 1996a, b), here we use RBD when turn to Bayesian models. To facilitate further discussion in this
referring generally to the probabilistic design approach. paper, below we present a concise summary of these Bayesian
Recognising both that RBD is in many ways superior to design regression models, emphasising the role of different variables
using factors of safety (the latter continues to form the basis of most (dependent, independent, and uncertain statistical parameters) in
rock engineering designs) and that many geotechnical designs need these models.
to interface with structural designs, since the 1980s, there has been
increasing recognition that geotechnical design codes and stan-
dards should also adopt RBD (e.g. Phoon, 2017). Now a number of σ1 >0
LSD and LRFD geotechnical codes exist that are engaging RBD
principles (EN-1997-1, 2004; CSA-S6-14, 2014b; AASHTO, 2016).
The literature on the applications of RBD to rock engineering is
σ3 > 0
Probability density
2.1. Regression using compressive strength data nonlinear regression (Bozorgzadeh et al., 2018a). Although it
therefore offers no analytical advantages over nonlinear regression
The simplest regression model is for the situation where the in such cases, the Bayesian approach is valuable in that it allows a
data consist solely of compressive data (both uniaxial and triaxial), straightforward and easy-to-understand probabilistic interpreta-
and this may be concisely written as tion of uncertainties, and also in that it permits the analysis to be
extended to include tensile strength data, as discussed below.
s1ðiÞ ms1 ðiÞ ; us1 wNormal ms1 ðiÞ ; us1 (1a)
merely a particular regression case where only compressive data performed: the left-hand filled circle and vertical line are the re-
for which s3 ¼ 0 are considered in Eqs. (1a) and (1b). sults of a Bayesian analysis using the model of Eqs. (1a), (1b) and
(2a)e(2c), and the right-hand open circle and vertical line are the
2.3. Model updating and computation equivalent of a frequentist analysis made using the Bayesian model
of Eq. (1b). For the Bayesian analysis, the filled circles are posterior
Given the relevant data for each regression model, the prior mean values that may be considered as the best estimate of the
distributions are updated according to Bayes’ rule to give posterior parameter in question and the vertical lines associated with each
distributions of the parameters. These parameters are m, sc (mean point are the 95% posterior credible interval (i.e. the 95% probability
uniaxial compressive strength), st (mean tensile strength), us1 range) for the parameter. For the frequentist analysis, the open
(precision of axial compressive strength), and uT (precision of circles are point estimates with the line indicating the associated
tensile strength). The posterior distribution of each parameter 95% confidence interval. It is noted that a Bayesian credible interval
represents the estimate of that parameter based on both its prior has a straightforward direct probabilistic interpretation: it is simply
distribution and the data. Because we have used vague and weakly the probability that the parameter of interest will have a value that
informative priors, we allow “the data to speak for themselves”. lies within the credible interval itself. This contrasts sharply with
Also, because they are distributions, these estimates include both the frequentist concept of confidence intervals, which strictly
the mean value and the associated uncertainty, and it is this un- should be interpreted in terms of a hypothetical long-run of
certainty that we primarily investigate in the analyses that follow. repeated similar experiments. A discussion on confidence intervals
Finally, it is important to recognise that these estimated parameters and their interpretation in the context of HeB regression, together
are not observable random variables (like rock strength in a single with a conceptual comparison to credible intervals, has been given
compression test) but are statistical parameters that describe the by Bozorgzadeh et al. (2018a), and a general discussion about them
regression model, the values of which are uncertain and are char- may be found in applied statistics textbooks (e.g. Ang and Tang,
acterised using probability in the Bayesian context. More details 2006).
about the Bayesian regression models and the interpretation of We make a number of observations about these analyses. Firstly,
posterior distributions may be found in Bozorgzadeh et al. (2018a). and most significantly in the context of RBD, we note the prepon-
The posterior distributions of the nonlinear regression models derance of small sample sizes: very few sets comprise more than
do not have an analytical form, and thus have to be determined ten results, and only one comprises more than 100 (although one
numerically. The numerical approximation usually uses Markov set contains 95 results). These sizes are in stark contrast to the
chain Monte Carlo (MCMC), and the MCMC for the Bayesian models requirement noted in the Introduction that “designers . collect
in this paper has used the OpenBUGS computer package (Lunn substantial data related to material properties”, and exemplify a
et al., 2009). MCMC is numerically intensive, and requires the significant problem that is faced in applying RBD to rock engi-
computational power of modern desktop computers to make it a neering. The call for more extensive testing of rock strength is not
practicable proposition. new, and the situation seems unlikely to change. This suggests that
another powerful aspect of Bayesian data analysis e namely, the
3. Observed variability of strength parameters objective incorporation of information from other sources e may
be required as standard practice in rock engineering design
As noted in the Introduction, knowledge of the variability of rock (Bozorgzadeh, 2018).
strength parameters is critical for RBD. Although some information For smaller data sets, there are often discrepancies between the
on variability is available in the literature, for example, when tables estimates from the two regression approaches, but there are sig-
of typical parameter values indicate likely ranges, few analyses of nificant overlaps between the 95% Bayesian credible interval and
rock strength report variability (in terms of standard deviation, or the 95% confidence interval. Generally, this is expected due to the
precision as shown in Eq. (1a)) and no rigorous and extensive use of vague and weakly informative prior distributions in the
treatment of the matter appears in the literature except very gen- Bayesian model, as such a model is similar to a frequentist model
eral summary statistics for different rock types (Aladejare and (Bozorgzadeh et al., 2018a). It will be seen that in a few cases, the
Wang, 2017). Some initial efforts have been made in this direc- frequentist estimate is missing from the plots. This indicates that
tion for soils (Phoon and Kulhawy, 1999), showing how the the nonlinear regression model is not able to converge to a solution,
geotechnical engineering community is beginning to tackle the and it suggests that those data sets are not amenable to having the
issue. HeB criterion fitted to them in the frequentist sense.
Here, we make use of 268 of rock strength data compilations Although we have arranged the results in these figures in
from Douglas (2002), each with at least four strength measure- increasing order of sample size, from visual assessment of the re-
ments, to statistically explore the nature of rock strength variability sults for all rock types we have analysed, we conclude that there is
using the regression model of Eqs. (1a) and (1b). These compila- no discernible pattern for variability in terms of sample size,
tions include many rock types, and we have extracted strength data beyond saying that for larger sample sizes (Bozorgzadeh et al.,
for rocks reported as coal, granite, limestone, marble, mudstone 2017), the estimates from both analysis methods are very similar.
and sandstone. As the presence of anisotropy in these materials has Indeed, as shown in Fig. 2a, small variability in m may be found for
not been recorded, we assume that they may be regarded as small (set 6) and large (set 24) data sets, but the same is true for
isotropic and thus the data are analysed on that basis. large variability (sets 5 and 21). We have insufficient knowledge of
For each rock type, we report posterior estimates of the values the particular data sets to say why these differences exist, but the
and variability of the HeB regression parameters m, sc and, results suggest that variability in m must be determined on a case-
importantly, the standard deviation of axial compressive strength. by-case basis.
Fig. 2 shows these results for 25 granite data sets. To aid compar- One aim of this work is to investigate the variability in strength
ison between the figures, they all use the same limits on the ordi- parameters in order to assess whether broadly applicable reference
nate axes. In these figures, the data sets have been arranged in values of variability could be determined for use in lieu of values
ascending order of size (the number of results in each set is shown obtained from rock testing. We conclude from the analyses pre-
at the top of the figure). For each data set, two analyses have been sented here that this will not be possible. Unfortunately, we note
616 N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627
341
sample size:
10
10
10
11
12
16
17
17
38
4
9
100
Bayesian posterior mean &
80 95% credible interval
40
20
0
0 5 10 15 20 25
Data set number
(a) Estimates of m.
341
sample size:
10
10
10
11
12
16
17
17
38
4
9
600 1029 1064 615 689
500
Estimated σ c (MPa)
400
300
200
100
0
0 5 10 15 20 25
Data set number
(b) Estimates of σc.
341
sample size:
10
10
10
11
12
16
17
17
38
4
200
150
100
50
0
0 5 10 15 20 25
Data set number
(c) Estimates of standard deviation of axial strength.
that estimates of parameter variability are currently seldom performed, one examining the ratio between direct and indirect
included in rock strength testing reports. In the context of RBD, this tensile strengths, and the other examining the evidence for a ten-
indicates that customary rock testing and reporting procedures e sile cut-off in the HeB criterion.
such as those set forth in the International Society for Rock Me- The subject of the analysis is an extensive strength dataset of
chanics and Rock Engineering (ISRM) suggested methods e will medium-grained metagraniteegranodiorite occurring at the
need to be modified to ensure that the necessary statistical data are Swedish Nuclear Fuel and Waste Management Company (SKB)
obtained. Forsmark site, Sweden (Elorant, 2004a,b,c; Jacobsson, 2004a,b,c,
2006, 2007; Gorski and Conlon, 2007). These results are from
4. Combined tensile and compressive data strength testing of core arising from boreholes KFM01A and
KFM01D at the Forsmark site, and we analyse them as a single
The principal aim of this section is to investigate the effect that group that, for brevity, is hereafter referred to as the Forsmark data.
different combinations of compressive and tensile strength results These data consist of uniaxial compressive (abbreviated as Cu,
have on the values of uncertainties in m, sc and standard deviation nCu ¼ 21), triaxial compressive (Ct, nCt ¼ 31), direct tensile (Td,
of axial strength. Two secondary investigations have been nTd ¼ 20) and indirect tensile (Ti, nTi ¼ 80) strengths. The Cu, Ct and
N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627 617
Posterior σ c (MPa)
utable to factors such as geological differences, or the strength
measurements being performed by different laboratories. Never-
theless, for the purpose of analysis presented below, which is a
discussion of uncertainties in strength parameter estimation
240
arising from different combinations of strength data type using
Bayesian regression models, we combine the Ti data into one set.
We address the appropriateness of this decision in the following 220
sub-section.
Recognising the testing regimes seen in practice, we have ana-
lysed three combinations of strength data: uniaxial and triaxial
compressive strengths, uniaxial compressive and tensile strengths,
200
10 15 20 25 30
and uniaxial and triaxial compressive strengths in conjunction with
Posterior m
tensile strength. The analysis makes use of the Bayesian regression
(a) Using direct tensile strength data.
model presented in Section 2.2 that combines tensile and
compressive data.
It has been known for some time that direct and indirect tensile
280 Cu & Ct
tests give different strength values, and factors relating the two (i.e.
the ratio Td/Ti) are available in the literature (e.g. Perras and Cu & Ti
Diederichs, 2014; Read and Richards, 2014). In the rock me- Cu & Ct & Ti
t
chanics literature, these are often called “correction factors”, but in 260 tme n
a djus
Posterior σ c (MPa)
the context of RBD, they should be referred to as either “model
factors” (EN 1990, 2002) or “transformation factors” (Phoon and
Kulhawy, 1999). For clarity, we adopt the latter term. Also, we 240
emphasise that the data and summary statistics reported in the t
tme n
literature allow for calculation of a mean transformation factor but a djus
does not allow for variability in Td and Ti data to be accounted for. Ti data adjusted by
As the Forsmark data contain both Td and Ti data, we take the 220
multiplying by the
opportunity to investigate the nature of the uncertainty that they
deterministic ratio
bring to strength analysis, and also to determine whether published
mean(Td)/mean(Ti)=0.58
transformation factors are adequate for practical application. 200
The HeB criterion was developed with emphasis on shear 10 15 20 25 30
strength in the compressive regime, and recent work questions its Posterior m
validity in the tensile regime (Hoek and Martin, 2014; Langford and (b) Using indirect tensile and adjusted indirect tensile strength data.
Perras, 2014). In particular, as there is evidence that fitting the HeB
criterion to shear strength data and extrapolating to the s1 ¼ 0 axis Fig. 3. Joint posterior distributions of m and sc for Forsmark data.
may underestimate or overestimate the tensile strength of the rock,
the use of a tensile cut-off has been proposed and is adopted in the
most recent version of the HeB criterion (Hoek and Brown, 2019). Table 1
Posterior summary statistics for Forsmark data (see Fig. 3).
As the Forsmark data are sufficiently comprehensive to allow us to
assess uncertainty surrounding adoption of a tensile cut-off, we Data Parameter Mean st. dev. 95% credible interval Correlation
take the opportunity to do so. Cu-Ct m 23.92 1.86 (20.38, 27.72) 0.68
sc (MPa) 241.16 6.22 (228.8, 253.3)
4.1. Analysis using direct tensile data st. dev. (MPa) 34.42 3.55 (28.32, 42.22) N/A
Cu-Td m 21.69 0.86 (20.07, 23.45) þ0.53
sc (MPa) 239.18 5 (229.3, 249.1)
For clarity, the results of the analyses applied in this section are st. dev. (MPa) 22.66 3.8 (16.68, 31.5) N/A
shown as joint probability distributions of m and sc. Fig. 3a shows Cu-Ct-Td m 22.44 0.68 (21.16, 23.82) þ0.1
joint posterior distributions for three different combinations of sc (MPa) 245.26 4 (237.07, 252.7)
st. dev. (MPa) 34.26 3.51 (28.24, 41.93) N/A
compressive and direct tensile strengths for the Forsmark data, and
CueTi m 14.41 0.43 (13.59, 15.27) þ0.7
Table 1 shows summary statistics of these posterior distributions. sc (MPa) 239.15 5.01 (229.2, 249)
These results show that, depending on the data combination st. dev. (MPa) 22.68 3.81 (16.67, 31.48) N/A
used, different best estimates (i.e. posterior mean, shown as crosses Cu-Ct-Ti m 16.1 0.42 (15.3, 16.93) þ0.55
at the centres of the iso-density contours) of pairs of m and sc are sc (MPa) 263.3 4.83 (253.9, 272.8)
st. dev. (MPa) 41.24 4.26 (33.87, 50.55) N/A
obtained. Furthermore, the uncertainties (i.e. standard deviations
of posterior m and posterior sc, shown by the extent of the con- Note: st. dev. is the standard deviation, and N/A means that the results are not
available.
tours), and correlations between the estimates (shown by the ec-
centricity of the elliptical contours) are also different for the three
data combinations. The close mean values and the notable overlap single group. The figure also supports the general observation that
of the three joint posterior distributions in Fig. 3a suggest similarity more data (even when provided in different categories of tensile
in strength, and thus support the earlier decision of combining the and compressive strengths) tend to result in reduced uncertainty:
data from boreholes KFM01A and KFM01D and analysing them as a here, the case of Cu and Ct data (n ¼ 52) shows the largest
618 N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627
uncertainty, with the case of Cu, Ct and Td (n ¼ 72) exhibiting the induced by these test configurations: the indirect (Brazilian) tensile
smallest uncertainty. test induces a combination of compressive and tensile stresses,
It is to be expected that different combinations of data generate whereas the direct tensile test induces tensile stress only. As the
different correlations between the posterior m and sc from a purely Bayesian model requires pure tensile rather than a combined stress
theoretical point of view: for Cu and Ct data, the dependent variable state, it incorrectly analyses the indirect tensile data. Nevertheless,
s1 increases with an increase in the independent variable s3 as it is indirect tensile data that are generally encountered in
(Eq. (1b)), and thus the parameter estimates are negatively corre- practice, it is important to overcome this problem in order to make
lated; and for Cu and Td data, the estimates are positively corre- best use of this widely available data type.
lated because, by Eq. (5), m is defined as a proportion of sc. Finally, A simple potential solution to this problem is to factor indirect
for the data analysed here, the combination of Cu, Ct and Td data tensile strength measurements (Bozorgzadeh et al., 2018a), and for
produces a posterior distribution that shows no significant the Forsmark data, a published value for the ratio of mean direct
correlation. tensile strength to mean indirect tensile strength is 0.58 (Read and
Although the significant overlap of the regions of high posterior Richards, 2014). Fig. 3b shows the effect of applying this trans-
probability indicates similarities between these estimates, overall formation factor, and comparison with Fig. 3a clearly represents a
these results appear problematic for rock strength testing. In significant improvement in the fit of the data to the HeB criterion.
addition to the previously known effect of sample size The distributions of sc shown in Fig. 3b also allow us to assess the
(Bozorgzadeh et al., 2017), we now see that the estimates of validity of combining the Ti data from boreholes KFM01A and
parameter value and variability are dependent on the type of KFM01D. As there is no indication that these distributions are
testing used. There are two principal conclusions to be drawn from bimodal, we assume that the data in each of the two boreholes may
these results. Firstly, as uncertainty is dependent on the type of be regarded as random samples from the same distribution of rock
testing used, for RBD purposes, it must always be reported as part of strength, and thus may be combined.
the analysis. Secondly, a combination of tensile with both uniaxial Although the transformation factor value of 0.58 applied here is
and triaxial compressive results is required to obtain minimal un- within the range of such values (0.5e1.1) reported in the literature
certainty in m and sc estimates. Once again, these suggest that for different types of granite, Bozorgzadeh et al. (2018a) suggested
revisions are required for guidelines such as the ISRM suggested caution when using these factors, and pointed out that the uncer-
methods. tainty associated with them should be taken into account. These
authors did not discuss this further, but the following shows how a
4.2. Analysis using indirect tensile data Bayesian analysis may be used to assess and incorporate such
uncertainty.
Fig. 3b shows the joint posteriors of m and sc for different data Unfortunately, and possibly due to the deterministic tradition of
combinations for the case where indirect tensile data are used rock engineering, variability in these transformation factors is not
instead of direct tensile data. Unlike the earlier results of Cu, Ct and always reported. However, in addition to reporting the mean factor
Td data, these three joint posteriors are different to the extent of 0.58 (Ti/Td), Read and Richards (2014) also reported the standard
where there is no overlap between them, and are therefore seen to deviation and sample size of both direct and indirect tensile test
generate different estimates of strength parameters. results. We will discuss later how this information may be used in
This misfit is due to the indirect tensile strength being greater the context of our regression model, but first examine the more
than direct tensile strength, to the extent that when combined with straightforward case of how the Forsmark direct and indirect ten-
the compressive data, they no longer fit well as predictors of an sile data allow us to investigate uncertainty in the value of trans-
HeB criterion. The boxplots presented in Fig. 4 clearly show the formation factor.
difference between the direct and indirect tensile strengths. This Our analysis incorporating variability in Ti/Td begins by
difference is well known, and is due to the different stress regimes randomly separating the indirect tensile data into two subsets of
equal size: the first subset is used in conjunction with the direct
tensile data to estimate the distribution of strength ratio, while the
second subset is pooled with the compressive data to estimate the
25 HeB strength parameters. This separation ensures that the same
data are not used twice in the same analysis. Two normal distri-
maximum butions e with vague prior distributions on their mean and vari-
ance e are fitted to the first subset of indirect and direct tensile
Tensile strength (MPa)
tmen While the outcome of this analysis is not that different from the
adjus
one presented earlier with the Normal (0.64, 0.032), the fact that it
was possible at all is only due to the authors reporting standard
240 t deviation and sample size along with the mean value. This em-
tmen
adjus phasises the critical importance of reporting not only measured
Ti data adjusted by data but also sufficient summary statistics to allow others to make
220 multiplying by the use of either the data or the statistics in their analyses. This is not a
deterministic ratio new reporting requirement, but is given new importance by the
mean(Td)/mean(Ti)=0.64 move to RBD. It is unfortunate that rock engineering generally lags
200 behind other engineering disciplines in terms of today’s standards
10 15 20 25 30 of reporting data and data summaries.
Posterior m These analyses go some way to address the problem of per-
(a) Using deterministic transformation factor. forming the statistical analyses required by RBD, and also support
Tye (1944)’s observation (noted above) that designers would need
280 Cu & Ct to collect substantial data related to material properties. We also
Cu & Ti believe that we have been able to wring more knowledge out of the
Cu & Ct & Ti data than is customary. However, as important as these analyses are
260 tmen
t e particularly in terms of input data for Monte Carlo simulation as
adjus
Posterior σ c (MPa)
Approximated using ratio of sc/m is found to have a posterior mean of 10.16 MPa and a
20 Normal(μ=0.58,ς2=0.022 ) 95% credible interval of (8.39 MPa, 12.29 MPa). This is shown in
Fig. 8 as estimate B. These two estimates differ visually, and the
Probability Density
figure shows that the mean value of estimate B falls outside the 95%
15 credible interval of estimate A. However, the non-negligible overlap
of the two credible intervals indicates that these data do not offer
10 strong evidence to refute the use of the approximation jstj z sc/m
to determine the tensile strength.
Applying the tensile cut-off of Eq. (7) with B0 ¼ 7 and B1 ¼ 0.81
5 to the same distributions of m and sc yields for st a posterior mean
of 9.19 MPa and a 95% credible interval of (7.88 MPa, 10.66 MPa).
This is shown as estimate C in Fig. 8, and is seen to differ more from
0 estimate A than estimate B. Nevertheless, and similarly to estimate
0.50 0.55 0.60 0.65 0.70 0.75 0.80
mean(Td)/mean(Ti) B, its 95% credible interval overlaps with that of estimate A e albeit
only slightly. Also, it is important to note that this credible interval
(a) Histogram of transformation factor simulated using information from the
is only based on the uncertainty associated with the posteriors of m
literature.
and sc, and does not include any uncertainties associated with the
linear cut-off relationship (i.e. those of the estimated values of m, sc
280 Cu & Ct and st used in fitting the equation, and also those of the slope and
Cu & Ct intercept of the fitted line, none of which are reported or discussed
Cu & Ct & Ti in the literature). The omission of these means that the uncertainty
t
260 tmen
adjus
Posterior σ c (MPa)
Estimate E
Estimate A Estimate C
σt from tensile cut-off,
mean of σt from tensile cut-off,
13 m, σt & σc estimated
direct tensile m & σc estimated
simultaneously
data from Cu & Ct
12 from Cu, Ct & Td
11
σ t (MPa)
10
Estimate D
σt=m/σc,
9 m, σt & σc estimated
Estimate B simultaneously
8 from Cu, Ct & Td
σt=σc/m,
m & σc estimated
7
from Cu & Ct
Fig. 8. Mean tensile strength of Forsmark data estimated from different data combinations and assumptions.
nonlinear strength criterion, these cases require curves represent- Characteristic curves:
Probability density
ing increasingly weaker materials.
With regard to rock engineering, the large volume case may be Lower bound of
Distribution of σ1 confidence interval
applicable to the rock mass surrounding an underground excava-
tion, while the small volume case may apply to cases like the rock for the mean
surrounding the anchor region of rock anchors, small pillars in hard
rock mines, and also the rock beneath the toes of piles bearing on 5th percentile curve
rock. Also, although rock engineering design generally concentrates Fitted H-B
on lower bound estimates of strength, upper bound estimates are (mean curve)
appropriate for design situations where attainment of peak
strength is an engineering objective, for example, caving tech-
niques in mining and rock excavation (both by mechanical means
and by blasting). For brevity, and also to avoid visual clutter in the
figures to follow, we focus our discussion on the lower bounds (for
both the small and large volume cases), although all the concepts
presented are equally applicable to the upper bounds.
Although it is rock mass strength that is of concern for most
engineering design situations, there are cases such as well-bore σ1
ax is
breakout, hydraulic fracturing or rock and strain bursting for is σ 3 ax
which intact rock strength is a design parameter. For this strength
parameter, Bozorgzadeh and Harrison (2016) discussed and
showed how to utilise the above definitions to estimate charac-
teristic values according to EC7 applying nonlinear regression (i.e. a Fig. 10. Schematic representation of characteristic values in the frequentist concept
(after Bozorgzadeh and Harrison, 2016).
frequentist approach) to a large data set comprising uniaxial and
triaxial compressive strength data. The authors particularly note
that, because strength is a function of confining pressure s3, char-
acteristic triaxial strength s* is also a function of s3. This indicates Posterior
that for the case of large ground volumes, we have mean
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi strength 90% credible Mean
Probability density
s*1 ¼ s*3 þ mbs b c s*3 þ b s 2c ta; n2 SEðE½s1 Þ (8) interval strength
criterion
and for small ground volumes, we have 5th percentile
Posterior of the posterior
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi predictive mean strength
s*1 ¼ s*3 þ mbs b c s*3 þ b s 2c ta; n2 SEðs1 Þ (9) strength
5th percentile
where s* is the value of confining pressure for which the charac-
of the posterior
teristic strength is to be determined; m b and sb c are the point esti-
predictive strength
mates; ta; n2 is the critical t-value for n data that gives a (1
a) 100% one-sided confidence interval (for example, a ¼ 0.05 for a Shaded area:
95% confidence interval); SE (E [s1]) and SE(s1) are the estimated 5% probability
standard error of the expected value or mean curve, i.e. the fitted
strength criterion, and that of the estimation of the axial strength,
respectively, evaluated at confining pressure s*. Note that for the σ1
ax is
case of nonlinear HeB regression, they must be approximated using is 90% credible σ 3 ax
numerical sampling methods like bootstrap or approximation interval σ3*
methods like the Delta method, details about which may be found
in e.g. Efron (1979) and Casella and Berger (2002), respectively.
Application of Eqs. (8) and (9) at various values of s* allows con- Fig. 11. Schematic representation of characteristic criteria in the Bayesian context.
struction of characteristic curves (Bozorgzadeh and Harrison,
2016). The resulting two curves are shown schematically in
Fig. 10 as the dash-dot and dashed curves, respectively. the first of which is a deterministic value and the other two are
It is necessary to recognise that the above discussion of char- characterised using probability distributions in the Bayesian
acteristic values from Bozorgzadeh and Harrison (2016) is strictly models of Eqs. (2a)e(2c) and illustrated in Fig. 3a. In the context of
frequentist. Recasting this discussion into a Bayesian context results this analysis, these are posterior distributions as they have been
in Fig. 11, and allows us to determine the two characteristic curves obtained from analysis of previously obtained strength data. Now,
in terms of the Bayesian regression models introduced earlier. In at a given value of s* (i.e. the confining stress at which the char-
the discussion that follows, we apply these concepts to the analysed acteristic strength is required) and sampling from the joint distri-
data, on the basis that intact rock strength may be a design bution of m and sc (which also accounts for any correlation
parameter in certain endeavours. between these parameters), we obtain an estimate of the mean
Firstly, consider the case where we are dealing with a large strength. In other words, at each MCMC iteration, a pair of m and sc
volume of ground. Following the discussion above, we need a lower values is sampled and used together with a value of s* in Eq. (1b) to
bound (or upper bound in cases where rock breakage is the design obtain a simulated value of mean axial strength ms1 . Many such
objective) estimate of the mean strength. Eq. (1b) shows that the iterations generate the posterior distribution of ms1 , the mean of
mean strength is a function of the three parameters s3, m and sc, which represents the best estimate of axial strength (as shown in
N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627 623
Cu & Td
Axial strength , σ 1 (MPa)
Cu & Td
Axial strength , σ 1 (MPa)
0 0 0
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Confining pressure, σ 3 (MPa) Confining pressure, σ 3 (MPa) Confining pressure, σ 3 (MPa)
(a) Mean (best- it) criteria for different data (b) 5th percentile of the posterior mean for (c) 5th percentile of the posterior predictive strength for
combinations. different data combinations. different data combinations.
Fig. 12. Mean and characteristic strength criteria for the Forsmark data in the compressive region.
624 N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627
where mk and sck are the HeB parameter values that result in an
HeB curve equivalent to that calculated using Eqs. (8) and (9). For
190 large numbers of compressive strength data, values of mk and sck
95% uncertainty may be calculated using frequentist quantile regression
range (Bozorgzadeh and Harrison, 2016), but when both tensile and
185 compressive data are required to be combined, or only small
numbers of results are available, a Bayesian model is appropriate.
As discussed in the previous sub-section, the results (i.e. pos-
teriors) of the Bayesian model allow the 5th percentile values at
180
Estimated value large numbers of s* values to be simulated, as shown in Fig. 13 for
the posterior distributions resulting from an analysis of uniaxial
and triaxial compressive and direct tensile Forsmark data. A curve
175 of HeB form is then fitted to these discrete points, the fitted pa-
5 10 20 50 200 500 2000 rameters of which may be regarded as mk and sck : Performing the
Number of σ 3 values used in simulation
simulation at large numbers of s* produces estimates of mk and sck
(b) Sensitivity analysis of estimates of σck for approximation of characteristic
that have negligible errors in comparison to the other uncertainties
(percentile) curves.
involved in the original regression model. This is demonstrated in
Fig. 14, which shows estimates and corresponding 95% uncertainty
bounds of mk and sck for the 5th percentile curve of the mean
Fig. 14. Sensitivity analysis of estimates of mk and sck for approximation of charac-
strength of Forsmark data (uniaxial and triaxial compressive and
teristic (percentile) curves.
direct tensile strengths) using a range of 5e2000 values for s*. We
300
Posterior σ c (MPa)
×Km0.05
250
×Kσc0.05
200
150
20 25 30 20 25 30 20 25 30
Posterior m Posterior m Posterior m
Fig. 15. H-B parameter values that express characteristic strength criteria (direct tensile strength data).
N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627 625
Table 3
Characteristic parameter values for Forsmark data (see Fig. 15).
Data Parameter Posterior mean (large volume of ground) Posterior predictive (small volume of ground)
Table 4
Multipliers for obtaining characteristic values for Forsmark data (see Fig. 15).
Data Parameter Posterior mean (large volume of ground) Posterior predictive (small volume of ground)
A lack of data for other rock types has prevented us from m and sc and the ratios mk/m and sck =sc are shown to depend on
examining other forms of adjustment and correction based on the testing regime used to obtain the strength data.
factors that may lead to practical lower bound characteristic curves. Overall, the paper shows how, with careful use of appropriate
One such example would be to factor measured values of strength statistical analysis, the HeB criterion can be applied within the
(both tensile and compressive) such that the HeB curve fitted to the framework of RBD, and concludes that modifications are needed to
factored strengths closely approximates the curve found from testing guidelines such as the ISRM suggested methods in order to
formal statistical analysis. To be of practical value, the factors would render them suitable to support RBD. Furthermore, as this is but a
need to either be constant across all rock types, or have different first attempt at placing the HeB criterion within the RBD frame-
values that apply to customary rock type groups (as is currently the work, it is clear that significant further analysis is required in order
case with look-up tables for values of m). The search for such factors to fully understand the implications of uncertainty on predicting
is perhaps a task that the wider rock mechanics community could rock strength and to generate the knowledge necessary to confi-
be usefully engaged in. dently apply such nonlinear criteria to the whole gamut of rock
types and rock masses.
6. Summary and conclusions
Conflicts of interest
It is over 70 years since the foundations of RBD were laid, and
not only does it now form the basis of structural engineering design The authors wish to confirm that there are no known conflicts of
codes globally, but also it is being adopted by geotechnical design interest associated with this publication and there has been no
codes worldwide. It is therefore necessary that rock engineering significant financial support for this work that could have influ-
practice evolves to embrace RBD, both by modification of current enced its outcome.
practices and development of new ones.
Among other things, at its core RBD requires acquisition and
References
probabilistic characterisation of material properties. This includes
rock strength, and as the HeB strength criterion remains the ACI 318-14. Building code requirements for structural concrete. Farmington Hills,
foremost model for this in rock engineering, in moving to RBD, it is USA: American Concrete Institute (ACI); 2014.
ACI 318R-14. Commentary on building code requirements for structural concrete.
critical that the rock engineering community begins to characterise
Farmington Hills, USA: ACI; 2014.
it in probabilistic terms. Such characterisation is the theme of this Aladejare AE, Wang Y. Evaluation of rock property variability. Georisk: Assessment
paper. and Management of Risk for Engineered Systems and Geohazards 2017;11(1):
In this context, we have presented three distinct analyses, all of 22e41.
American Association of State Highway and Transportation Officials (AASHTO).
which use a Bayesian approach to model probability and perform AASHTO LRFD bridge design specifications. 7th ed. Washington, D.C., USA:
regression. The choice of a Bayesian approach is necessary because AASHTO; 2016.
of both the complexity of variability encountered and the limited Ang AH, Tang WH. Probability concepts in engineering: emphasis on applications to
civil and environmental engineering. 2nd ed. New York: John Wiley & Sons;
data generally available from rock strength testing campaigns. 2006.
The first of these analyses uses a compilation of compressive Becker D. Limit states design for foundations. Part I. An overview of the foundation
strength testing results of intact rock to examine the variability in design process. Limit states design for foundations. Part I. An overview of the
foundation design process. Canadian Geotechnical Journal 1996a;33(6):956e
the HeB parameters m and sc, and the predicted axial strength for 83.
rock types including coal, granite, limestone, marble, mudstone and Becker D. Limit states design for foundations. Part II. Development for national
sandstone. The results suggest that, in the context of RBD, vari- building code of Canada. Canadian Geotechnical Journal 1996b;36(3):984e
1007.
ability shows so little similarity within and between rock types that
Bond A, Harris A. Decoding Eurocode 7. London: Taylor & Francis; 2008.
it will not be possible to make use of reference values for this, and Bozorgzadeh N, Harrison JP. Characteristic triaxial strength of intact rock for LSD.
that it will need to be determined during each rock testing Proceedings of the Institution of Civil Engineers e Geotechnical Engineering
2016;169(3):291e8.
campaign.
Bozorgzadeh N, Yanagimura Y, Harrison JP. Effect of small numbers of test re-
The second of the analyses uses an extensive set of compressive sults on accuracy of HoekeBrown strength parameter estimations: a sta-
and tensile (both direct and indirect) strength data for a granodi- tistical simulation study. Rock Mechanics and Rock Engineering 2017;
orite, together with a Bayesian regression model, to develop joint 50(12):3293e305.
Bozorgzadeh N, Escobar MD, Harrison JP. Comprehensive statistical analysis of
probability distributions of m and sc. This model is an enhancement intact rock strength for reliability-based design. International Journal of Rock
to the one developed in Bozorgzadeh et al. (2018a) in that it im- Mechanics and Mining Sciences 2018a;106:374e87.
plements a probabilistic transformation factor for mean indirect Bozorgzadeh N, Harrison JP, Escobar MD. Hierarchical Bayesian modelling of
geotechnical data: application to rock strength. Geotechnique 2018. https://doi.
tensile strength data rather than a fixed single value. These joint org/10.1680/jgeot.17.P.282.
distributions are suitable for further engineering analysis, including Bozorgzadeh N. Contributions in uncertainty quantification towards reliability-
design, and are used here to confirm the efficacy of applying a based rock engineering design. PhD Thesis. Toronto, Canada: University of
Toronto; 2018.
transformation factor to indirect tensile results when fitting the Casella G, Berger RL. Statistical inference. 2nd ed. Pacific Grove, USA: Duxbury
HeB criterion to such data. The work also shows that deterministic Press; 2002.
and probabilistic factors generate different results, and probabi- Contreras LF, Brown ET, Ruest M. Bayesian data analysis to quantify the uncertainty
of intact rock strength. Journal of Rock Mechanics and Geotechnical Engineer-
listic factors can only be made available if variability in direct and ing 2018;10(1):11e31.
indirect tensile strengths is reported. CSA-A23.3-14. Design of concrete structures. Canadian Standards Association (CSA);
Finally, the granodiorite data are used to investigate the 2014a.
CSA-S6-14. Canadian highway bridge design code (CHBDC). CSA; 2014b.
important matter of developing characteristic strength criteria.
Douglas KJ. The shear strength of rock masses. The University of New South Wales;
Using definitions from EC7, a formal Bayesian interpretation of 2002. PhD Thesis.
characteristic strength is proposed and used to show how strength Efron B. Bootstrap methods: another look at the jackknife. The Annals of Statistics
data can be robustly analysed to generate a characteristic criterion. 1979;7(1):1e26.
Elorant P. Forsmark site investigation e drill hole KFM01A: indirect tensile strength
Formal probabilistic sampling allows these criteria to be presented test (HUT). SKB P-04-171. Swedish Nuclear Fuel and Waste Management
in terms of characteristic parameters mk and sck . The values of both Company (SKB); 2004a.
N. Bozorgzadeh, J.P. Harrison / Journal of Rock Mechanics and Geotechnical Engineering 11 (2019) 612e627 627
Elorant P. Forsmark site investigation e drill hole KFM01A: triaxial compression test Lunn D, Jackson C, Best N, Thomas A, Spiegelhalter D. The BUGS book: a practical
(HUT). SKB P-04-177. SKB; 2004b. introduction to Bayesian analysis. CRC Press; 2012.
Elorant P. Forsmark site investigation e drill hole KFM01A: uniaxial compression Orr T. Selection of characteristic values and partial factors in geotechnical designs to
test (HUT). SKB P-04-176. SKB; 2004c. Eurocode 7. Computers and Geotechnics 2000;26(3e4):263e79.
EN 1990. Eurocode e basis of structural design. Brussels, Belgium: European Perras MA, Diederichs MS. A review of the tensile strength of rock: concepts and
Committee for Standardisation (CEN); 2002. testing. Geotechnical and Geological Engineering 2014;32(2):525e46.
EN-1997-1. Eurocode 7: geotechnical design e Part 1: general rules. Brussels, Phoon KK, Kulhawy FH. Evaluation of geotechnical property variability. Canadian
Belgium: CEN; 2004. Geotechnical Journal 1999;36(4):625e39.
Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Bayesian data Phoon KK. Role of reliability calculations in geotechnical design. Georisk: Assess-
analysis. 3rd ed. CRC Press; 2013. ment and Management of Risk for Engineered Systems and Geohazards
Gorski B, Conlon B. Forsmark site investigation: determination of the direct and 2017;11(1):4e21.
indirect tensile strength on cores from borehole KFM01D. SKB P-07-76. SKB; Read S, Richards L. Correlation of direct and indirect tensile tests for use in the
2007. HoekeBrown constant mi. In: Alejano R, Perucho Á, Olalla C, Jiménez R, editors.
Hoek E, Brown ET. Empirical strength criteria for rock masses. Journal of the Rock engineering and rock mechanics: structures in and on rock masses. CRC
Geotechnical Engineering Division, ASCE 1980a;106(9):1013e35. Press; 2014. p. 161e6.
Hoek E, Brown ET. Underground excavations in rock. London: Institution of Mining Richardson LF. Weather prediction by numerical process. Cambridge University
and Metallurgy; 1980b. Press; 1922.
Hoek E. Strength of jointed rock masses. Geotechnique 1983;33(3):187e223. Spross J, Johansson F. When is the observational method in geotechnical engi-
Hoek E. Strength of rock and rock masses. ISRM News Journal 1994;2(2):4e16. neering favourable? Structural Safety 2017;66:17e26.
Hoek E, Carranza-Torres C, Corkum B. HoekeBrown failure criterion e 2002 edition. Tye W. Factor of safety e or of habit? The Aeronautical Journal 1944;48(407):487e
In: NARMS-TAC 2002: Mining and Tunnelling Innovation and Opportunity, 94.
Proceedings of the 5th North American Rock Mechanics Symposium and the Wang L, Gong W, Luo Z, Khoshnevisan S, Juang CH. Reliability-based robust
17th Tunnelling Association of Canada Conference (NARMS-TAC 2002), vol. 1. geotechnical design of rock bolts for slope stabilization. In: Iskander M,
Toronto, Canada: University of Toronto Press; 2002. p. 267e73. Suleiman MT, Anderson JB, Laefer DF, editors. Proceedings of the international
Hoek E, Martin CD. Fracture initiation and propagation in intact rock e a review. foundations congress and equipment expo 2015 (IFCEE 2015). American Society
Journal of Rock Mechanics and Geotechnical Engineering 2014;6(4):287e300. of Civil Engineers (ASCE); 2015. p. 1926e35.
Hoek E, Brown ET. The HoekeBrown failure criterion and GSI e 2018 edition.
Journal of Rock Mechanics and Geotechnical Engineering 2019;11(3):445e63.
Jacobsson L. Forsmark site investigation e borehole KFM01A: triaxial compression
test of intact rock. SKB P-04-227. SKB; 2004a.
Nezam Bozorgzadeh is currently a post-doctoral fellow at
Jacobsson L. Forsmark site investigation e borehole KFM01A: uniaxial compression
the GeoEngineering Centre at Queen’s-RMC, Kingston,
test of intact rock. SKB P-04-223. SKB; 2004b.
Canada. He obtained his PhD degree from University of
Jacobsson L. Forsmark site investigation e drill hole KFM01A: indirect tensile
Toronto. His research interests are probabilistic design, un-
strength test. SKB P-04-170. SKB; 2004c.
certainty quantification, and Bayesian data analysis for
Jacobsson L. Forsmark site investigation e borehole KFM01D: triaxial compression
soil/rock engineering. His current research is probabilistic
test of intact rock. SKB P-06-214. SKB; 2006.
LRFD calibration of internal limit states of mechanically
Jacobsson L. Forsmark site investigation e boreholes KFM01A and KFM02B: micro
stabilised earth (MSE) walls for the Canadian Highway
crack volume measurements and triaxial compression tests on intact rock. SKB
Bridge Design Code.
P-07-93. SKB; 2007.
Jaeger JC, Cook NG, Zimmerman R. Fundamentals of rock mechanics. 4th ed. Wiley-
Blackwell; 2007.
Jimenez-Rodriguez R, Sitar N. Rock wedge stability analysis using system reliability
methods. Rock Mechanics and Rock Engineering 2007;40(4):419e27.
Kovari K, Tisa A, Einstein HH, Franklin J. Suggested methods for determining the
strength of rock materials in triaxial compression: revised version. Interna-
tional Journal of Rock Mechanics and Mining Sciences & Geomechanics Ab-
stracts 1983;20(6):285e90. John P. Harrison holds the W.M. Keck Chair of Engineering
Langford JC, Perras MA. Obtaining reliable estimates of intact tensile strength. In: Rock Mechanics at the University of Toronto, Canada. He
The 48th US rock mechanics/geomechanics symposium. Minneapolis, USA: obtained his PhD degree in rock mechanics from the
American Rock Mechanics Association; 2014. University of London, UK. His current principal research
Langford JC, Diederichs MS. Quantifying uncertainty in HoekeBrown intact strength interest is developing reliability-based design techniques
envelopes. International Journal of Rock Mechanics and Mining Sciences for rock engineering, and since 2011, he has been actively
2015;74:91e102. participating in the revision of Eurocode 7 to support rock
Liu H, Low BK. System reliability analysis of tunnels reinforced by rockbolts. engineering design.
Tunnelling and Underground Space Technology 2017;65:155e66.
Low BK, Phoon KK. Reliability-based design and its complementary role to Euro-
code 7 design approach. Computers and Geotechnics 2015;65:30e44.
Lunn D, Spiegelhalter D, Thomas A, Best N. The BUGS project: evolution, critique
and future directions. Statistics in Medicine 2009;28(25):3049e67.