You are on page 1of 18

Bayesian Estimation of Rock Mechanical Parameter and

Stability Analysis for a Large Underground Cavern


Quan Jiang1; Jian Liu2; Hong Zheng3; Bin Wang4; Zhi-Zhong Guo5; Tao Chen6; and Xian-Tao Xiong7

Abstract: The uncertainty rock mechanical parameters (i.e., deformation and strength parameters) is an important factor in the safety esti-
mation and support design of underground engineering. Ignoring this uncertainty could allow potential risks to the structure. To address this
challenge, this paper develops and verifies a Bayesian approach for a rock’s mechanical parameters estimation by integrating limited site data
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

and prior knowledge, and the integrated knowledge is then transformed into a large number of equivalent samples of the rock’s parameters.
The experimental data of marble from triaxial compression tests are first used to verify this method, and the results show that this method can
effectively estimate the distribution of the marble’s deformation and strength parameters under the condition of small samples. Further, the
probability distribution of rock mass parameters is obtained with the help of Hoek–Brown criterion. Then, the random field of the rock mass
with a large cavern is constructed according to the obtained parameter distribution, the influence of different autocorrelation distances is dis-
cussed, and the excavation-induced deformation’s statistical analysis is carried out. Finally, the failure of the surrounding rock is character-
ized probabilistically, which can be a reference to the reliability design of rock support in underground engineering. DOI: 10.1061/(ASCE)
GM.1943-5622.0002452. © 2022 American Society of Civil Engineers.
Author keywords: Underground engineering; Bayesian method; Parameters distribution; Markov Chain Monte Carlo simulation;
Deformation probability distribution.

Introduction factor (Chen et al. 2019; Cui et al. 2017; Jiang et al. 2016; Pinheiro
et al. 2016), and, according to current research, random field theory
The uncertainty of rock mechanical parameters has an important is widely used to characterize this uncertainty (Chen et al. 2019;
impact on the stability assessment and deformation prediction of Chen and Zhang 2021; Liu and Qi 2018). A vital problem of gen-
rock mass engineering, such as foundation (Asem and Gardoni erating a random field is how to determine the distributions of pa-
2019; Zhang et al. 2014), slope (Chen et al. 2020; Gravanis et al. rameters as accurately as possible (Ching et al. 2016; Phoon and
2014), underground cavity (Chen et al. 2019; He et al. 2020; Li Kulhawy 1999a; Wang et al. 2015). In addition, unreasonable pa-
et al. 2020), and tunnels (Li and Low 2010; Zheng et al. 2021; rameter distributions will put the whole project at risk and lead to
Zhou et al. 2021). The uncertainty of rock mechanical parameters unnecessary losses. It is also a matter of great concern to engineers
may come from many aspects, such as spatial variability, model how to use the probability method to characterize the excavation
transformation variability, and experimental errors. Given these un- displacement and failure of surrounding rock in large underground
certainties, the inherent spatial variability is the most important caverns, which has great significance for the support structure
design.
1
Professor, State Key Laboratory of Geomechanics and Geotechnical Some scholars have studied distributions of deformation pa-
Engineering, Institute of Rock and Soil Mechanics, Chinese Academy of rameters and strength parameters of rock; for example, Jiang
Sciences, Wuhan 430071, China (corresponding author). ORCID: https:// et al. (2016) tested a large number of marble samples under dif-
orcid.org/0000-0001-6039-9429. Email: qjiang@whrsm.ac.cn ferent levels of confining stress and carried out statistical anal-
2
Ph.D. Candidate, State Key Laboratory of Geomechanics and ysis to obtain the distribution of a rock’s strength parameters,
Geotechnical Engineering, Institute of Rock and Soil Mechanics, Chinese and Cui et al. (2017) further discussed the minimum number
Academy of Sciences, Wuhan 430071, China; Univ. of Chinese Academy of samples required to obtain sufficiently accurate rock param-
of Sciences, Beijing 100049, China.
3 eters distributions. Nevertheless, the traditional method to ac-
Associate Professor, State Key Laboratory of Geomechanics and
Geotechnical Engineering, Institute of Rock and Soil Mechanics, Chinese
quire sufficiently precise statistics requires a considerable
Academy of Sciences, Wuhan 430071, China. amount of data obtained via laboratory or field tests, which
4
Professor, State Key Laboratory of Geomechanics and Geotechnical will cost considerable manpower and costs. To obtain more
Engineering, Institute of Rock and Soil Mechanics, Chinese Academy of meaningful parameter distributions under small sample condi-
Sciences, Wuhan 430071, China. tions, it is reasonable to consider additional knowledge to sup-
5
Senior Engineer, Sichuan Huaneng Luding Hydropower Corporation plement the information contained in data according to the
Limited, Chengdu 610017, China. Bayesian method. The application of the Bayesian method in
6
Senior Engineer, Sichuan Huaneng Luding Hydropower Corporation parameters estimation of geotechnical engineering has made
Limited, Chengdu 610017, China. some progress. For example, Cao and Wang (2014a) used the
7
Senior Engineer, PowerChina Chengdu Engineering Corporation
Bayesian method for model selection and characterized the un-
Limited, Chengdu 610017, China.
Note. This manuscript was submitted on October 5, 2021; approved on drained shear strength of soft soil. Feng and Jimenez (2015) es-
February 25, 2022; published online on June 8, 2022. Discussion period timated the deformation modulus of a rock mass given a set of
open until November 8, 2022; separate discussions must be submitted for test data by Bayesian information criterion. Wang and Akeju
individual papers. This paper is part of the International Journal of Geo- (2016) used a Bayesian equivalent sample method to character-
mechanics, © ASCE, ISSN 1532-3641. ize the cohesion and friction angle of soil. Contreras et al.

© ASCE 04022129-1 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


(2018) used Bayesian regression method to characterize the pa- Sari and Karpuz 2006; Song et al. 2011). In this study, the
rameter uncertainty of Hoek–Brown criterion. Asem and Gar- same strategy is taken:
doni (2019) used the Bayesian approach to develop for E:
probabilistic models for the rock socket shear and normal stiff-    
nesses. The equivalent sample method based on a Bayesian 1 1 ln(E) − μE 2
f (E) = √ × exp − (1)
framework has proved that it can improve the accuracy of pa- 2π Eσ E 2 σE
rameter estimation of soil with limited data (Cao and Wang
2013; Wang and Akeju 2016; Wang and Cao 2013); however, for c and ϕ:
its effectiveness in rock mass must be verified and developed   
further. 1 1 c − μc 2
In addition, the application of the probability method in under- f (c, ϕ) =  exp −
2πσ c σ ϕ 1 − ρ2 2(1 − ρ2 ) σc
ground engineering has also made some progress. Nomikos     2 
and Sofianos (2011) used the analytical method to obtain the prob- c − μc ϕ − μϕ ϕ − μϕ
−2ρ + (2)
ability distribution of the safety factor of the supporting structure σc σϕ σϕ
aiming at the mine tunnel in a rock mass. Mollon et al. (2011)
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

used the response surface method to study the influence of cohesion where μE, μc, and μϕ = means of E, c, and ϕ, respectively; σE, σc,
and internal friction angle on tunnel stability. Chen et al. (2019) and σ ϕ = standard deviations (std) of E, c, and ϕ; and ρ = cross-
used the Karhunen–Loève (K–L) expansion discretizing the Gauss- correlation coefficient that describes the cross-correlation degree
ian random field and verified that the random parameters can rep- between c and ϕ (Wang and Akeju 2016); it is defined as (Cheung
resent the anisotropy of layered rock mass parameters in the case and Tang 2005)
of tunneling. Zheng et al. (2021) used the multivariate distribution
function to establish the probability model of related random vari- Cov(c, ϕ)
ρ= (3)
ables and obtained the nonlinear relationship between tunnel dis- σc σϕ
placements and rock mass parameters. These methods have made
some progress; however, a robust probabilistic method that can where Cov(c, ϕ) = covariance of c and ϕ.
deal with the excavation of complex underground caverns is still To construct the joint probability distribution, both prior
worth studying. knowledge and site-specific data are needed. Prior knowledge
This paper develops and verifies the equivalent samples ap- includes possible conjectures about the parameters (e.g., μE,
proach to acquire the parameters of rock, according to the data σE), which could be acquired from investigations, published re-
from a large number of triaxial compression tests of marble ports, and engineering experience (Cao et al. 2016a). By com-
from the Chinese Jinping II hydropower station, and transforms bining site-specific data with prior knowledge, the
these samples into parameters of rock mass. Then, combined corresponding probability of rock parameters can be defined as
with the acquired parameter distributions of rock mass, the defor- a joint conditional probability density function (PDF) of P(μE,
mation distribution and failure probability of surrounding rocks σE|Data) and P(μc , μϕ , σ c , σ ϕ , ρ|Data), where Data =
are estimated considering the influence of different autocorrela- site-specific data of field or laboratory tests. To construct the
tion distances. joint PDFs of the rock parameters, the theorem of total probabil-
ity is used first, which can be expressed as:
for E:
Probabilistic Modeling and Bayesian Framework

P(E|Data) = P(E|μE , σ E )P(μE , σ E |Data)dμE dσ E (4)


μE ,σ ϕ
Probabilistic Modeling of Rock Deformation and Strength
Parameters for c and ϕ:
As a typical natural material, various factors affect a rock’s de-

formation parameters (e.g., elastic modulus E) and strength pa- P(c, ϕ|Data) = P(c, ϕ|μc , μϕ , σ c , σ ϕ , ρ)
μc ,μϕ ,σ c ,σ ϕ ,ρ
rameters (e.g., cohesion c and friction angle ϕ), and these
parameters vary in space, which is commonly called “spatial × P(μc , μϕ , σ c , σ ϕ , ρ|Data)dμc dμϕ dσ c dσ ϕ dρ
variability.” It is more plausible to describe the rock parameters (5)
with probability theory and consider them as random variables.
It is worth noting that the transformation uncertainty is another where P(E|μE, σE) and P(c, ϕ|μc , μϕ , σ c , σ ϕ , ρ) = joint PDFs under
important uncertainty that measures the difference between the condition of prior knowledge based on the specific site. Accord-
model space and actual parameter space (Phoon and Kulhawy ing to Eqs. (1) and (2), the PDFs can be expressed as:
1999a). Since the transformation model is not included in this for E:
study, the transformation uncertainty is not discussed here.    
The deformation parameter E and the strength parameters c 1 1 ln(E) − μE 2
P(E|μE , σ E ) = √ × exp − (6)
and ϕ are generally treated as a one-dimensional (1D) lognormal 2π Eσ E 2 σE
distribution and two-dimensional joint normal distributions, re-
spectively (Ang and Tang 2007; Phoon and Kulhawy 1999b; for c and ϕ:

        
1 1 c − μc 2 c − μc ϕ − μϕ ϕ − μϕ 2
P(c, ϕ|μc , μϕ , σ c , σ ϕ , ρ) =  exp − − 2ρ + (7)
2πσ c σ ϕ 1 − ρ2 2(1 − ρ2 ) σc σc σϕ σϕ

© ASCE 04022129-2 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


The PDFs of P(μE, σE|Data) and P(μc , μϕ , σ c , σ ϕ , ρ|Data) in


Eqs. (4) and (5) reflect the probability distribution of parameters K= P(Data|μE , μc , μϕ , σ c , σ ϕ , ρ)
μc ,μϕ ,σ c ,σ ϕ ,ρ
under the site-specific data, and the Bayesian framework is an effec-
tive and powerful tool to acquire these PDFs (Cao et al. 2016b; Jiang × P(μc , μϕ , σ c , σ ϕ , ρ)dμc dμϕ dσ c dσ ϕ dρ (11)
et al. 2018). Here, P(μE, σE|Data) and P(μc , μϕ , σ c , σ ϕ , ρ|Data) can where K = normalizing constant of the Bayesian model; Data =
be expressed as: experimental data obtained from the tests; P(Data|μE, σE) and
for E: P(Data|μE , μc , μϕ , σ c , σ ϕ , ρ) = likelihood functions that reflect
the probability of acquiring Data; and P(μE, σE) and P(μc , μϕ ,
P(μE , σ E |Data) = K −1 × P(Data|μE , σ E ) × P(μE , σ E ) (8)
σ c , σ ϕ , ρ) = prior distribution.

Considering that experimental samples are sparse, and the sam-
pling distance is generally large, the data from the specific site
K= P(Data|μE , σ E ) × P(μE , σ E )dμE dσ E (9) could be deemed independent of each other, so the likelihood func-
μE ,σ ϕ
tion can be rewritten as:
for c and ϕ:    
nE
1 1 ln(Ei ) − μE 2
P(Data|μE , σ E ) = √ × exp −
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

P(μc , μϕ , σ c , σ ϕ , ρ|Data) = K −1 × P(Data|μE , μc , μϕ , σ c , σ ϕ , ρ) i=1 2π Ei σ E 2 σE


× P(μc , μϕ , σ c , σ ϕ , ρ) (12)
(10)

        

nc,ϕ
1 1 cj − μc 2 cj − μc ϕj − μϕ ϕj − μϕ 2
P(Data|μE , μc , μϕ , σ c , σ ϕ , ρ) =  exp − − 2ρ + (13)
j=1 2πσ c σ ϕ 1 − ρ
2 2(1 − ρ2 ) σc σc σϕ σϕ

where Data = {Ei}; i = 1, 2…nE = set of data E; and Data = {cj, ϕj}; are sufficient experiences and strong confidence in the parameter dis-
i = 1, 2…nc,ϕ = set of data pair c and ϕ. tribution, which could supplement the lack of information of data to
The prior distributions can be uniform distribution when there is the greatest extent.
not much prior knowledge about the distributions of the parame- So far, combined with Eqs. (4) and (5), the distributions of E, c,
ters, which is called “vague prior” or “noninformative prior.” Be- and ϕ conditioned on Data are acquired and expressed as:
cause it only provides the possible range of parameters values, for E:
and there is no specific preference for a certain parameter value

compared with other distributions (e.g., normal distribution). The P(E|Data) = K P(E|μE , σ E ) × P(Data|μE , σ E )
range of each variable can be taken as the interval between the pos- μE ,σ ϕ
sible minimum values and maximum values of a rock’s parameters × P(μE , σ E )dμE dσ E (16)
according to engineering experience or the relevant literature (Cao
et al. 2016a). The uniform distribution is adopted and can be ex- for c and ϕ:
pressed as:

for E: P(c, ϕ|Data) = K P(c, ϕ|μc , μϕ , σ c , σ ϕ , ρ)


μc ,μϕ ,σ c ,σ ϕ ,ρ
μE ∈ [μEmin , μEmax ] × P(Data|μc , μϕ , σ c , σ ϕ , ρ)
P(μE , σ E ) = [(μEmax − μEmin )(σ Emax − σ Emin )]−1
σ E ∈ [σ Emin , σ Emax ]
× P(μc , μϕ , σ c , σ ϕ , ρ)dμc dμϕ dσ c dσ ϕ dρ (17)
(14)
Note that the derivations of the formula are mainly from
for c and ϕ: Wang and Akeju (2016) and Wang and Cao (2013) and are mod-
P(μE , μc , μϕ , σ c , σ ϕ , ρ) = [(μcmax − μcmin )(σ cmax − σ cmin ) ified since the transformation uncertainty is not considered here.
The distributions of E, c, and ϕ are complex, and it is hard to
× (μϕmax − μϕmin )(σ ϕmax − σ ϕmin ) acquire any useful information from the distributions directly.
× (ρmax − ρmin )]−1 A widely used method is the Markov Chain Monte Carlo
(MCMC) sample from complex distributions (Wang et al.
μc ∈ [μcmin , μcmax ]
2015; Zhang et al. 2012). The samples drawing from the distri-
σ c ∈ [σ cmin , σ cmax ] butions [i.e., Eqs. (17) and (18)] contain information of both
μϕ ∈ [μϕmin , μϕmax ] test data and prior knowledge that benefits from the Bayesian
σ ϕ ∈ [σ ϕmin , σ ϕmax ] framework, so they can be deemed as the equivalent samples
of test data (Wang and Akeju 2016).
ρ ∈ [ρmin , ρmax ] (15)
In Eqs. (14) and (15), μEmin and μEmax = possible lower limit and
upper limit of the mean of E; and σEmin and σEmax = possible lower MCMC Simulation
limit and upper limit of the std of E. For c, ϕ, and ρ, the symbol
definitions are the same. Note that the choices of the prior distribu- The MCMC can generate a sequence of samples that obeys the
tion are flexible, a smaller intervals or informative distributions complex target distribution after the states of the Markov Chain
(e.g., normal or log-normal distribution) are recommended if there reaches stationary condition. In addition, the adoption of the

© ASCE 04022129-3 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


MCMC avoids the tedious calculation of the normalizing constant
K in Eqs. (9) and (11), which can improve the computational effi- mσ,c = (σ c,max − σ c,min )/Δσ c
ciency significantly.
In this study, the Metropolis–Hastings (M–H) algorithm mμ,ϕ = (μϕ,max − μϕ,min )/Δμϕ
(Hastings 1970) is used in MCMC to generate equivalent samples
of the rock’s parameters from Eqs. (16) and (17). For providing a mσ,ϕ = (σ ϕ,max − σ ϕ,min )/Δσ ϕ
starting point in the Markov Chain, the means of the prior is se-
lected as the first E1 and (c1, ϕ1). Then, candidate samples En* or mρ = (ρmax − ρmin )/Δρ (25)
(c*n , ϕ*n ) will be generated from the proposal PDF f (En* |En−1 ) or
f (c*n , ϕ*n |cn−1 , ϕn−1 ), where n = n-th state of the Markov Chain, where ΔμE, ΔσE, Δμc, Δσc, Δμϕ , Δσ ϕ , and Δρ = lengths of the in-
and the superscript * = candidate state. The proposal PDF can be tervals of grid corresponding to μE, σE, μc, σc, μϕ , σ ϕ , and ρ,
taken as a simple distribution in general. The Gaussian distribution respectively.
adopted here as the proposal distribution with a mean of the Therefore, μ jμ,E , σ jσ,E , μ jμ,c , μ jμ,ϕ , σ jσ,c , σ jσ,ϕ , and ρ jρ in j-th inter-
(n−1)-th state of En−1 or (cn−1, ϕn−1). The correlation coefficient val are given as:
for the proposal distribution of c and ϕ is equal to the mean of for E:
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

the prior of correlation coefficient. (2jμ,E − 1)ΔμE


Then, the acceptance ratio [raE for E and rac,ϕ for (c, ϕ)] can be μ jμ,E = μE,min +
2
calculated according to the M–H algorithm, which is expressed as:
for E: (2jμ,E − 1)Δσ E
  σ jσ,E = σ E,min + (26)
I(En* ) f (En−1 |En* ) 2
raE = min 1, × (18)
I(En−1 ) f (En* |En−1 ) for c and ϕ:

(2jμ,c − 1)Δμc
μ jμ,c = μc,min +
I= P(E|μE , σ E )P(Data|μE , σ E ) × P(μE , σ E )dμE dσ E (19) 2
μE ,σ ϕ

for c and ϕ: (2jσ,c − 1)Δσ c


σ jσ,c = σ c,min +
2
I(c*n , ϕ*n ) f (cn−1 , ϕn−1 |c*n , ϕ*n )
rac,ϕ = min 1, × (20) (2jμ,ϕ − 1)Δμϕ
I(cn−1 , ϕn−1 ) f (c*n , ϕ*n |cn−1 , ϕn−1 ) μ jμ,ϕ = μϕ,min +
2

I= P(c, ϕ|μc , μϕ , σ c , σ ϕ , ρ) (2jσ,ϕ − 1)Δσ ϕ


μc ,μϕ ,σ c ,σ ϕ ,ρ
σ jσ,ϕ = σ ϕ,min +
2
× P(Data|μE , μc , μϕ , σ c , σ ϕ , ρ)
(2jρ − 1)Δρ
× P(μc , μϕ , σ c , σ ϕ , ρ)dμc dμϕ dσ c dσ ϕ dρ (21) ρ jρ = ρmin + (27)
2
The integral item of I is calculated by the grid approximation Note that P(Data|μ jμ,E , σ jσ,E ) and P(Data|μ jμ,c , μ jμ,ϕ , σ jσ,c ,
method, and the corresponding functions are: σ jσ,ϕ , ρ jρ ) can be calculated before sampling because they do not
for E: change with different candidate samples.

mμ,E 
mσ,E Subsequently, a number from the uniform distribution within
I= P(E|μ jμ,E , σ jσ,E ) × P(Data|μ jμ,E , σ jσ,E ) the range of [0, 1] is randomly generated. If the acceptance ratio
jμ,E =1 jσ,E =1 is larger than the randomly selected number, then the candidate
× P(μ jμ,E , σ jσ,E )Δμ jμ,E Δσ jσ,E (22) samples will be accepted as the n-th samples; otherwise, the candi-
date will be rejected, and the (n−1)-th samples will be saved as the
for c and φ: n-th samples. In addition, there is a “burn-in process” in the early
mμ,c mσ,c mμ,ϕ mρ period of the MCMC simulation, and the samples do not converge
I= j =1 j =1 j =1 j =1
P(c, ϕ|μ jμ,c , μ jμ,ϕ , σ jσ,c , σ jσ,ϕ , ρ jρ ) to the target distributions (posterior distributions), which means the
μ,c σ,c μ,ϕ ρ
samples belonging to the burn-in process should be excluded from
× P(Data|μ jμ,c , μ jμ,ϕ , σ jσ,c , σ jσ,ϕ , ρ jρ ) the statistics. The specific length of the burn-in process cannot be
determined accurately, and a longer burn-in process is required
× P(μ jμ,c , μ jμ,ϕ , σ jσ,c , σ jσ,ϕ , ρ jρ )Δμc , Δμϕ , Δσ c , Δσ ϕ , Δρ for more complex distributions. In this study, the first 50% of the
(23) samples will be excluded as the samples corresponding to the
burn-in process; this strategy is consistent with the Contreras
where mμ,E, mσ,E, mμ,c , mσ,c, mμ,ϕ , mσ,ϕ, and mρ = numbers of in- et al. (2018).
tervals of the corresponding grids of parameters, which are calcu-
lated as:
for E:
Implementation Procedure
mμ,E = (μE,max − μE,min )/ΔμE
To further clarify the implementation process of the equivalent
for c and ϕ: mσ,E = (σ E,max − σ E,min )/Δσ E (24) sample approach, Fig. 1 shows the implementation procedure for
acquiring the equivalent samples. The main steps are as follows:
mμ,c = (μc,max − μc,min )/Δμc (i) Obtain ns parameter values of the rock from laboratory tests.

© ASCE 04022129-4 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. Flow chart of equivalent sample approach.

(ii) Obtain a set of appropriate prior knowledge of the parameter Rock samples were collected from the underground powerhouse,
values, including μE, σE, μc, σc, μϕ , σ ϕ , and ρ. and the triaxial compression tests for the marble specimens with
(iii) Choose the appropriate grid size for the grid approximation to 100 mm in height and 50 mm in diameter were executed in a rock
each parameter. mechanics testing system (Jiang et al. 2016). The confining stresses
(iv) Choose a proper initial state for the sampling, then a large num- are set as 5, 10, 20, 30, and 40 MPa, and more than 15 specimens
ber of Bayesian equivalent samples will be generated using the were tested under each confining stress. The typical stress–strain
M–H algorithm. curves are obtained via triaxial compression experiments (Fig. 2).
(v) Discard the samples belonging to the burn-in process and esti- Subsequently, the deformation parameter (e.g., E) and strength pa-
mate the parameter distribution of the rock using the Bayesian rameters (e.g., c and ϕ) are obtained by the test data analysis.
equivalent samples; for example, acquire the PDFs and cumu-
lative distribution functions (CDFs) of the parameters.
Uncertainty of Marble’s Deformation Parameter
Equivalent Samples of Jinping II Station Marble Based on the triaxial compression tests, 109 elastic moduli are ob-
tained, as plotted in Fig. 3. The mean of these 109 samples is
To verify the effectiveness of the equivalent sample approach for 49.98 GPa, and the std is 13.07 GPa. When enough samples are ac-
rock parameters, the test data of marble in the Jinping hydropower quired, statistical parameters could be estimated easily. In addition,
station’s underground powerhouse are used to estimate the distribu- 30 samples are randomly selected from all the samples to verify the
tion of marble strength parameters and deformation parameters. effectiveness of the method, as listed in Table 1.
The uniform distribution is used to be consistent with previous
research (Cao and Wang 2014b; Wang and Akeju 2016). The prior
Engineering Background range of μE is set as 0 to 100 GPa, and the prior range of σE is set as
0 to 50 GPa according to the experience (Behzadafshar et al. 2019;
The Jinping II hydropower station is located on the Yalong River in Feng and Jimenez 2014; Moradian and Behnia 2009). The intervals
Sichuan Province, China. Its main powerhouse is 352.4 m in of the grid (i.e., ΔμE and ΔσE) are defined in Eqs. (26) and (27) used
length, 28.3 m in width, and 72.2 m in height and is buried in a to calculate the integral, ΔμE is set as 0.2 GPa for μE, and ΔσE is set
marble stratum with a vertical burial depth of 400 m. For such a as 0.1 GPa for σE. Note that the interval of the grid does not have a
large-scale underground engineering cavern, the large deformation specific requirement and can be modified according to the com-
and failure risks of the surrounding rock should be taken into con- puter performance and the time cost. According to the study of
sideration cautiously to ensure safety. Wang and Cao (2013), it is recommended that the number of

© ASCE 04022129-5 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


(a) (b)
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 2. Typical results of the triaxial compression testing of Jinping II marble: (a) typical stress–strain curves under different confining stresses; and
(b) typical failure of marble specimens under different confining stresses.

Fig. 3. Histogram of elastic modulus results from the triaxial compres-


sion tests.

Table 1. Data of elastic modulus E randomly selected from the triaxial Fig. 4. Scatter plot of the elastic modulus results of the Bayesian equiv-
compression tests alent samples.

Parameter Values
E (GPa) 45.84 42.24 46.32 58.48 75.78
43.88 36.77 51.75 44.47 33.17
45.82 68.91 49.78 48.77 55.23
44.72 50.28 49.81 22.89 49.91
30.84 53.12 50.39 96.21 40.28
50.87 50.57 89.05 33.17 33.01

grids in each dimension should not be less than 150. Because the
uniform distribution only provides a possible range for the model
parameters, the results of the equivalent sample method will be
close to the frequency method, which is convenient for the compar-
ison of the results of the two methods.
There are a total of 60,000 equivalent samples for E generated by
Fig. 5. PDF of elastic modulus E.
MCMC, and the first 50% of the samples are discarded to avoid the
effects of the burn-in period in which the samples do not converge to
the target distribution; only the last 30,000 samples are selected for included in the figure, which are represented by circles. The rela-
analysis (Contreras et al. 2018). Fig. 4 shows the 30,000 samples tionship between the test data and probability density based on
with trace plot of sampling; it also includes the mean, 5% percentile, equivalent samples shows good consistency.
and 95% percentile presented by different line styles, which are Table 2 summarizes the statistical results. The mean and std of
49.88, 28.4, and 71.69 GPa, respectively. As we can see, the gener- 30 randomly selected data are 49.67 and 15.3 GPa, and the mean
ated samples are sparsely distributed below 28.4 and above 71.6 GPa and std of E based on the equivalent sample approach are 49.88
but are concentrated in the region from 28.4 to 71.6 GPa. and 13.40 GPa, respectively. The exact value of the parameter can-
Fig. 5 is the frequency histogram of 30,000 elastic modulus not be obtained, so the results of a large number of test data are set
equivalent samples. The 109 triaxial compression tests are also as the standard to measure the statistical error. The absolute

© ASCE 04022129-6 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Table 2. Summary of statistics of E
Statistics Triaxial compression tests samples (109) Randomly selected samples (30) Bayesian equivalent samples (30,000)
Mean (GPa) 49.98 49.67 49.88
Absolute difference (GPa) — 0.32 0.11
Relative difference (%) — 0.63 0.21
Standard variation (GPa) 13.07 15.30 13.40
Absolute difference (GPa) — 2.16 0.37
Relative difference (%) — 16.48 2.81

Table 3. Hypothesis testing for the distribution of E using the K–S test
Distribution format

Elastic modulus Index Normal Lognormal


Triaxial compression Dn 0.1931 0.1557
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

tests samples (109) Judgment Rejectable Acceptable


Best distribution Lognormal —
Bayesian equivalent Dn 0.1783 0.1258
samples (30,000) Judgment Rejectable Acceptable
Best distribution Lognormal —

differences of the means of randomly selected samples and equiv-


alent samples are 0.32 and 0.11 GPa; the absolute differences of the (a)
standard deviations of randomly selected samples and equivalent
samples are 2.16 and 0.37 GPa, respectively. The relative differ-
ences of the means of the random samples and equivalent samples
are 0.63% and 0.21%, respectively, and the relative differences of
the standard deviations of random samples and equivalent samples
are 16.48% and 2.81%, respectively. Compared with the randomly
selected samples, the results of equivalent samples are closer to the
true value, especially for the std. To obtain a deeper understanding,
the K–S test is adopted to analyze the parameter distribution. As
plotted in Table 3, the result demonstrates that the best distribution
of E is a lognormal distribution, and the Bayesian equivalent sam-
ples have the same distribution pattern as the test data.
To further verify the effectiveness of the method, the effects on (b)
the calculation results are studied when the sample size (ns) is set as
5 to 30 with the interval of 5 (i.e., ns = 5, 10, 15, 20, 25, 30). For Fig. 6. Effect of the number of test data of E: (a) estimates of the mean
each sample size, the estimations are performed 10 times respec- of E; and (b) estimates of the std of E.
tively using the randomly selected data from all test results. The es-
timation results are shown in Fig. 6, where the circles are the result
of direct statistical analysis, and the crosses are the results using the Table 4. Data pairs of c and ϕ acquired from the triaxial compression tests
Bayesian method. The dotted line and the solid lines are the mean
Parameters Values
and the 95% confidence interval, respectively, calculated from 109
test results. The results show that the results obtained by the Baye- (c (MPa), ϕ (°)) (30.90, 29.68) (37.83, 20.90) (40.60, 18.64)
sian method are generally closer to the results evaluated directly (25.35, 33.59) (33.17, 26.26) (35.92, 23.74)
using 109 test data compared with the direct estimations using ran- (26.37, 31.31) (39.68, 20.88) (29.91, 26.94)
(40.79, 19.43) (28.05, 28.76) (32.57, 24.56)
domly selected data. This advantage is more obvious under the con-
(40.49, 17.95) (36.01, 21.29) (31.43, 28.81)
dition of small samples since the Bayesian method incorporates the (29.71, 29.30) (32.45, 26.98) (31.70, 30.94)
influence of prior knowledge. (26.37, 33.98) (29.98, 30.57) (24.62, 33.67)
(30.83, 30.01) (27.25, 29.13) (39.27, 22.01)
(34.60, 26.93) (28.80, 30.61) (34.28, 24.63)
(38.50, 23.60) (35.19, 26.51) (33.00, 28.28)
Uncertainty of Marble’s Strength Parameters

According to the triaxial compression tests, the strength parameters


of Jinping marble are acquired, and the Bayesian equivalent sam- and the prior range of σ ϕ is set as 0° to 30°; and the prior range of ρ
ples approach is applied to characterize the strength parameters is set as 0 to –1. The intervals of the grid of c are set to 0.2 MPa for
of the Jinping marble according to 30 sample pairs randomly se- μc and 0.1 MPa for σc; the intervals of the grid of ϕ are set to 0.3°
lected from the complete data set. Table 4 lists the randomly se- for μϕ and 0.15° for σ ϕ ; and the intervals of the grid of ρ are set as
lected data pairs. 0.05. MCMC is executed 20,000 times, and the initial 10,000 sam-
The prior range of μc is set as 10 to 50 MPa, and the prior range ples are discarded to avoid the influence of the burn-in period.
of σc is set as 0 to 20 MPa; the prior range of μϕ is set as 10° to 70°, Ultimately, 10,000 equivalent samples are obtained and used.

© ASCE 04022129-7 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 7. Plot of the 10,000 equivalent samples and 30 test samples of c and ϕ.

Table 5. Summary of mean and standard deviation of c


Absolute Relative Absolute Relative
Method Mean of c (MPa) difference (MPa) difference (%) Std of c (MPa) difference (MPa) difference (%)
Triaxial compression tests (100) 33.25 — — 6.33 — —
Triaxial compression tests (30) 32.85 0.60 1.80 4.82 1.51 23.85
Bayesian equivalent samples (10,000) 33.60 0.35 1.05 5.56 0.77 12.16

Table 6. Summary of mean and standard deviation of ϕ


Mean of ϕ Absolute Relative difference Std of ϕ Absolute Relative difference
Method (°) difference (°) (%) (°) difference (°) (%)
Triaxial compression tests (100) 25.42 — — 5.85 — —
Triaxial compression tests (30) 26.66 1.24 4.88 4.56 1.29 22.05
Bayesian equivalent samples (10,000) 26.03 0.81 3.19 4.09 1.76 30.08

Fig. 7 includes the scatter plot, 1D histograms, kernel destiny The detailed statistics of c, ϕ, and ρ are presented in Tables 5–7,
curves, and statistics of the 10,000 equivalent samples of c and ϕ. respectively. Based on 100 complete data pairs, the mean of c and ϕ
The equivalent samples are represented by solid circles, and the are 33.25 MPa and 25.42°; the std of c and ϕ are 6.33 MPa and
30 triaxial compression test samples are represented by stars. Ac- 5.85°; and the cross-correlation coefficient (ρ) between c and ϕ is
cording to the equivalent samples, the means of c and ϕ are −0.925. For the randomly selected data pairs, the absolute differ-
33.6 MPa and 26.03°; the std of c and ϕ are 5.56 MPa and ences of means corresponding to c and ϕ are 0.6 MPa and 1.24°,
4.09°; and the cross-correlation coefficient between c and ϕ is respectively, and the relative differences are 1.8% and 4.88%; the
estimated to be −0.94. The cross-correlation coefficient of absolute differences of std corresponding to c and ϕ are
−0.94 indicates a strong negative dependence between c and ϕ, 1.51 MPa and 1.29°, and the relative differences are 23.85% and
which means when the value of ϕ decreases, the value of c 22.05%, respectively. For the equivalent sample, the absolute dif-
increases rapidly. ference of means corresponding to c and ϕ is 0.35 MPa and

© ASCE 04022129-8 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


0.81°, and the relative differences are 1.05% and 3.19%, respec- are performed based on randomly selected test data. The remaining
tively; the absolute differences of the std corresponding to c and calculation parameters of the equivalent sample are consistent, as
ϕ are 0.77 MPa and 1.76°, and the relative differences are before. The estimation results of mean and std are plotted in
12.16% and 30.8%, respectively. The cross-correlation coefficient Figs. 9 and 10, respectively. It can be seen that the variability of
(ρ) between c and ϕ based on randomly selected data and equiva- the std is greater than the means because there are more results
lent samples are −0.938 and −0.940, respectively. The statistics of of the std outside the 95% confidence interval. And the evaluation
c, ϕ, and ρ show that the samples of the equivalent sample approach results of equivalent samples have certain advantages over direct
maintain good consistency with the direct test. statistical analysis, especially under the condition of a small sam-
Fig. 8 also shows the joint probability distribution of c and ϕ of ples size (e.g., ns = 5, 10) because Bayesian method incorporates
the 10,000 Bayesian equivalent samples. The darker the color, the the prior knowledge. When the sample size is large (e.g., ns
greater the joint probability density of the samples. To further verify ≥25), the statistical results of the two methods tend to be consistent.
the distribution of the marble strength parameters, K–S testing is ex- This is because when the prior knowledge only contains weak in-
ecuted, as plotted in Table 8. The results based on test data show that formation, the equivalent samples will be mainly determined by
the best distribution for cohesive strength and internal friction are the test data if there is a large number of test data.
lognormal and normal, respectively, and the normal distribution is So far, the equivalent sample approach for inferring the de-
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

also acceptable for both parameters. The results based on the equiv- formation parameters, strength parameters, and correlation co-
alent samples show that the best distributions of cohesive strength efficients of marble are verified. After obtaining the
and internal fiction are both normal distributions, but the lognormal distributions of E, c, and ϕ, the deformation and failure proba-
distribution is also acceptable, which is slightly different from the re- bility of the underground cavern will be analyzed further using
sult of test data. However, considering the convenience of establish- these distributions.
ing a joint distribution of cohesion and internal friction angle in the
next section, the normal distributions for both parameters are chosen.
In addition, the impacts of different sample sizes (ns) on the es-
timation results of the strength parameters are studied, including six
cases with sample sizes varying from 5 to 30 with an interval of 5
(i.e., ns =5, 10, 15, 20, 25, 30). For each case, 10 times estimations

Table 7. Summary of cross-correlation coefficient ρ


Absolute Relative
Method ρ difference difference (%)
Triaxial compression tests (100) −0.925 — —
Triaxial compression tests (30) −0.938 0.013 1.4
Bayesian equivalent samples (10,000) −0.940 0.015 1.6
(a)

(b)

Fig. 9. Effect of the number of test data on the means of c and ϕ:


Fig. 8. Joint PDF of c and ϕ. (a) estimates of the mean of c; and (b) estimates of the mean of ϕ.

Table 8. Hypothesis testing for the distributions of strength parameters using the K–S test
Cohesive strength (c) Internal friction (ϕ)

Samples Index Normal Lognormal Normal Lognormal


Triaxial compression tests samples Dn 0.0273 0.0095 0.0411 0.0661
Judgment Acceptable Acceptable Acceptable Acceptable
Best distribution Lognormal Normal
Bayesian equivalent samples Dn 0.0133 0.0134 0.0143 0.0187
Judgment Acceptable Acceptable Acceptable Acceptable
Best distribution Normal Normal

© ASCE 04022129-9 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


′ ′ a−1
σ ci [(1 + 2a)s + (1 − a)mb σ3n ](s + mb σ3n )
c′ = 

(1 + a)(2 + a) 1 + (6amb (s + mb σ3n ) )/((1 + a)(2 + a))
a−1

(31)
where σ ′3n = σ ′3 max /σ ci , σ ′3max = upper limit of confining stress
when the relationship between the Hoek–Brown and the Mohr–
Coulomb criteria is considered; mb = reduced value of the material
constant mi; mi = parameter related to rock type, which is set as 10
for marble (Cai 2010); s and a = constants for the rock mass; and
mb, s, and a are given by the following relationships:
 
GSI − 100
mb = mi exp (32)
(a) 28 − 14D

 
GSI − 100
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

s = exp (33)
9 − 3D

1 1
a = + (e−GSI/15 − e−20/3 ) (34)
2 6
For simplicity, the parameter values required in parameters es-
timation of rock mass are listed in Table 9. According to the stated
relationship, the Em is approximately 11.45 GPa, and c′ and ϕ′ of
rock mass are estimated to be 4.97 MPa and 25.98°, respectively.
Further, to convert the parameters of intact rock into the param-
eters of rock mass, the reduction coefficient is defined as the param-
(b) eters of rock mass obtained by using the Hoek–Brown criterion
divided by the average value of intact rock’s parameters (e.g.,
Fig. 10. Effect of the number of test data on the std of c and ϕ: (a) es-
rE = Em/μE). Therefore, the reduction coefficients of Em are
timates of the std of c; and (b) estimates of the std of ϕ.
0.229, and the reduction coefficients of c′ and ϕ′ are 0.147 and
0.99, respectively. Then, each equivalent sample is multiplied
Parameter Reduction for Engineering Rock in an with the reduction coefficients and the parameter’s statistics is
Underground Cavern carried out. It is worth noting that the means of parameters is con-
sistent with the results obtained by Hoek–Brown criterion, the std
The parameters are measured from intact samples, but actual engi- and probability distributions of rock mass parameters are obtained
neering projects of underground caverns are located in rock masses further for the probability analysis of rock mass engineering.
affected by joints and fractures. Thus, parameter reduction between In addition, tensile failure is also a typical failure form for rock
intact rock and engineering rock mass is necessary (Hoek 2002; mass excavation. Direct tensile strength tests of rocks are recom-
Hoek and Brown 1997; Sonmez and Ulusay 1999). In practice, mended but not routinely conducted due to the difficulty in speci-
the Hoek–Brown criteria is often used to estimate the parameters men preparation. Brazilian tests are often used but the validity is
of rock mass generally (Cai et al. 2004; Eberhardt 2012; Hoek controversial (Cai 2010). Sheorey (1997) provided a method for es-
and Diederichs 2006). According to Hoek (2002), the deformation timating the tensile strength according to the strength ratio (R),
modulus of rock mass (Em) can be estimated as which is described as
σc
  =R (35)
D σ ci GSI −10 |σ t |
Em = 1 − · 10 40 for σ ci ≤ 100 MPa (28)
2 100
where σt = rock mass tensile strength; and σc = rock mass strength,
which is defined as (Hoek 2002)
  2c′ cosϕ′
D GSI−10
σc = (36)
Em = 1 − · 10 40 for σ ci > 100 MPa (29) 1 − sinϕ′
2
The recommended value of R is 10 (Sheorey 1997). Therefore, the
where σc = uniaxial compressive strength of the intact rock; GSI = tensile strength of rock mass is taken as 10% of its compressive
geological strength index, which is set as 60 according to the geo- strength in this study, which is −0.6 MPa.
logical exploration of Jinping II hydropower station; and D = factor
depending on the disturbance degree of the rock mass and is set as Table 9. Parameters of Hoek–Brown criteria
0.6. In addition, Hoek (2002) provides a method to obtain the Parameters Value
equivalent cohesion (c′ ) and internal friction angle (ϕ′ ) of rock
σci(MPa) 98
mass, the formula is given as
GSI 60
  D 0.6
6amb (s + mb σ ′ 3n )a−1 γ(kN/m3) 27
ϕ′ = sin−1 (30) mi 10
2(1 + a)(2 + a) + 6amb (s + mb σ ′ 3n )a−1

© ASCE 04022129-10 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Table 10. Deformation parameter and strength parameters of rock mass m = number of cross-correlation variables. Subsequently, the stan-
Statistics Em(GPa) ′
c (MPa) ′
ϕ (°)
dard Cholesky decomposition algorithm is used to decompose the
matrices R and C to obtain the cross-correlated standard Gaussian
Reduction coefficients 0.229 0.147 0.99 random fields, which can be expressed as
Means 11.45 4.97 25.98
Std 3.01 0.82 4.05 C = L1 × LT1 (39)

Finally, parameters required for numerical calculation are listed R = L2 × LT2 (40)
in Table 10 and will be used to analyze the deformation and failure where L1 and L2 = lower triangular matrices corresponding to C
probability of the Jinping II underground powerhouse. and R.
The cross-correlated standard Gaussian random fields H are de-
rived as
Stability Analysis for the Underground Cavern H = L1 ξL2 (41)
To quantitatively estimate the stability of the underground cavern where ξ consists of the vectors of independent normal random sam-
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

by probabilistic approaches, the random finite-element method ples according to their distributions.
(RFEM) (Griffiths and Fenton 2004; Griffiths and Fenton 2009)
is adopted to analyze the deformation distribution of the main pow-
erhouse and estimate the failure probability of the surrounding rock
Reliability Calculation for Underground Caverns
by considering the parameter uncertainty.
After the coupled random field and finite-element model are estab-
lished, the Monte Carlo simulation (MCS) can be executed easily
Generation of the Random Parameters Field for based on the repeated generations of random field elements. For
Jinping II Engineering Rock Nmcs times calculation of MCS, the number of times the element
enters the yield state (i.e., the plastic strain of the element is greater
The spatial variability reflects that the parameters of rock masses in than 0) is recorded as N. Then, the probability of failure Pf of each
different locations are various, and it is obvious that the rock param- element can be calculated as
eters at adjacent locations have more similarity than those at far po-
N
sitions (Chen et al. 2019; Liu and Qi 2018, Xu et al. 2021). Random Pf = (42)
fields are generally used to describe these similarities and differ- Nmcs
ences. Among them, the stationary random field is the most widely Note that the finite-element calculations are executed by the com-
used; the “stationary” in most studies indicates that the means and std mercial calculation software Abaqus (2011) to ensure the accuracy
of the rock mass maintain the same value in space. In addition, the of the finite-element calculation results. The cross section of the
relationship of the same parameters in the different locations can computational model is a height of 320 m in the x-direction and a
be fitted by the autocorrelation function (ACF); the difference in var- width of 320 m in the y-direction, with 28,356 finite elements
ious ACFs has been studied in the literature (Li et al. 2015). The sin- [Fig. 11(a)], and the powerhouse is placed in the center to eliminate
gle exponential ACF (SNX) is considered in this study, since it is the boundary effects. The elements close to the excavation boundaries
most widely used, which can be defined as of the powerhouse are scaled by approximately 0.5 m to satisfy the
   requirements of calculation accuracy (Ching and Phoon 2013). The
τx τy
ρ(lh , lv ) = exp − + (37) initial stress field is set as σ1 = 12.2 MPa and σ3 = 6.1 MPa accord-
lh ly ing to the results of the in situ measurements. The boundary of the
where τx and τy = distances between two locations in the horizontal module is taken as fixed conditions. Table 10 lists the statistics of
and vertical directions, respectively; and lh and lv = horizontal and the model parameters. The ACF is set as SNX, and the autocorre-
vertical autocorrelation distances, respectively. lation matrix C is calculated using the centroid coordinates (xi, yi)
Then, according to the centroid coordinates (xi, yi) of random of random field elements, in which i is the number of random
field elements and the ACF [Eq. (37)], the autocorrelation matrix field elements.
C can be obtained and expressed as Thus, most of the parameters used for the random field are ac-
⎡ ⎤ quired, but the autocorrelation distance of rock mass has not
1 ρ(τx12 , τy12 ) · · · ρ(τx1ne , τy1ne ) been fully studied yet (Liu and Qi 2018), so it is more plausible
⎢ ρ(τx21 , τy21 ) · · · ρ(τx12 , τy12 ) ⎥ to execute a parameter study than assume a confirmed value.
⎢ 1 ⎥
C= ⎢ ⎢ .. .. .. .. ⎥
⎥ Twenty group calculations are executed to study the effect of the
⎣ . . . . ⎦ horizontal autocorrelation distance (lh) and vertical autocorrelation
ρ(τxne 1 , τyne 1 ) ρ(τxne 2 , τyne 2 ) ··· 1 distance lv, as listed in Table 11. When lv is set as 30 m, lh varies
from 1 to 320 m, corresponding to calculation Cases R1 to R10,
(38)
which has enough range to cover the real autocorrelation distance
where i and j = number of random field elements; ρ(τxij, τyij) = au- of marble. When lh is equal to 320 m, its scale is the same as the
tocorrelation coefficient between elements i and j according to Eq. module scale, and the 2D random field degenerates to the 1D ran-
(37); τxij = |xi − xj| and τyij = |yi − yj| = relative distances of ele- dom field; the same calculation strategy is taken to determine lh cor-
ments i and j in the x-direction and y-direction, respectively; and responding to calculation Cases R11 to R20. The variations in
C =symmetric matrix, that is, ρ(τxij, τyij) = ρ(τxji, τyji). gravity and Poisson’s ratio of the rock mass are relatively small
Next, the cross-correlation matrix R is acquired as R = (ρk,l)m×m, and set as confirmed values in this study, which are 27 kN/m3
where ρk,l = cross-correlation coefficient between random variable and 0.3, respectively. It is worth noting that the deformation param-
k and random variable j (e.g., c, and ϕ) in the same location; and eters and strength parameters are set at the same autocorrelation

© ASCE 04022129-11 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

(c) (d)

Fig. 11. Outline of the powerhouse and a typical realization of a random field: (a) outline and size of the powerhouse and in situ stress distribution;
(b) typical realization of a random field of E; (c) typical realization of a random field of c; and (d) typical realization of a random field of ϕ.

Table 11. Calculation conditions corresponding to different autocorrelation distances


Autocorrelation
Statistics distance (/10 m)

Variables Mean std ρ Distribution lh lv Names


Em(GPa) 11.45 3.01 — Normal (SQX) 0.1 3 R1
— — — — 0.5 — R2
— — — — 1 — R3
c′ (MPa) 4.97 0.82 −0.94 Normal (SQX) 2 — R4
— — — — 3 — R5
— — — — 4 — R6
ϕ ′ (◦ ) 25.98 4.05 Normal (SQX) — 5 — R7
— — — — 8 — R8
— — — — 16 — R9
— — — — 32 — R10
Em(GPa) 11.45 3.01 — Normal (SQX) 3 0.1 R11
— — — — — 0.5 R12
— — — — — 1 R13
c′ (MPa) 4.97 0.82 −0.94 Normal (SQX) — 2 R14
— — — — 3 R15
— — — — 4 R16
ϕ ′ (◦ ) 25.98 4.05 Normal (SQX) — 5 R17
— — — — 8 R18
— — — — 16 R19
— — — — 32 R20

© ASCE 04022129-12 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Fig. 12. Means of deformation of monitoring points with the number of MCS calculations.
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

distances in this study, as both of them are always positively corre-


lated according to experience.
Fig. 11 also shows the discretization of the random field ele-
ments and a typical realization of the random field for E, c, and
ϕ, respectively. It shows that the parameters significantly vary
from location to location, as signified with different colors; for ex-
ample, the darker color in Fig. 11(c) indicates a larger c, the lighter
color indicates a smaller c, and the color changes more rapidly in
the x-direction than in the y-direction because its horizontal auto-
correlation distance is shorter than the vertical autocorrelation
distance. In addition, the locations of darker color in Fig. 11(c) cor-
respond to the location of lighter color in Fig. 11(d), which implies
a negative correlation between c′ and ϕ′ .
To determine the calculation times of MCS, the average dis-
placements of nine monitoring points are counted, as plotted in
Fig. 12. The results show that the variability of the means of defor- Fig. 13. Statistical deformation of the monitoring points of R1.
mation decreases with the increase of the number of MCS and
reaches stability when the calculation time is set as 8,000
(Fig. 12) for all nine monitoring points. Therefore, the number of average displacement at 8# is 58.03 mm when the autocorrelation
calculations of each case is taken as 8,000 in this study. distance is equal to 1 m, and the maximum mean at 8# is
60.06 mm when the autocorrelation distance is 320 m. However,
the std of displacement is 1.85 mm when the autocorrelation dis-
Statistical Analysis of Excavation-Induced tance is equal to 1 m, and the standard deviation of displacement
Deformation of Surrounding Rock reaches 10.08 mm when the autocorrelation distance is 320 m. In
addition, the results show that the horizontal autocorrelation dis-
For obtaining the deformation uncertainty of a cavern, the excava- tance and vertical autocorrelation distance have the same degree
tion displacements of monitoring points are analyzed. Fig. 13 dis- of effect on the cavern’s deformation, which is different from
plays the deformation of calculation Case R1 along the edge of the some presented results in soil engineering.
powerhouse by the violin plot. In this violin plot, the monitoring To study the influence of the autocorrelation distance on the dis-
point with the maximum mean is 8#, which is the midpoint of placement in detail, the violin plots show the displacement distribu-
the right sidewall, and the mean of displacement is 59.6 mm. How- tion of the 8#monitoring point in Figs. 15(a and b), considering that
ever, corresponding to different MCS, the minimum and maximum the displacement of the 8# monitoring point is the largest and most
displacements at 8# are 47.81 and 79.32 mm, respectively, and the discrete. The median displacement of different horizontal and ver-
upper bound and lower bound of the 95% confidence interval are tical autocorrelation distances is approximately 59 mm, and the
67.61 and 50.59 mm, respectively; it can be seen that the monitor- larger the autocorrelation distances, the more discrete the displace-
ing points with larger average displacement are more discrete. The ment data. In other words, deformation with a larger deviation from
minimum average displacement occurs at monitoring Point 1#, and the average value is more likely to occur in the site with a larger
the value is 19.18 mm. At the same time, the corresponding dis- autocorrelation distance.
placement dispersion is also the smallest.
Furthermore, the influence of autocorrelation distances on the
displacement of different monitoring points is studied. Generally
speaking, the excavation deformation of the cavern increases Probability Distribution of Element Failure
with the increase of the autocorrelation distance (for both lh and
lv). The results show that the autocorrelation distance has little ef- In the excavation process of the main powerhouse, the failure of
fect on the mean of the displacement of monitoring points, as surrounding rock is also one of the important references to the de-
shown in Fig. 14(a), but it has a great influence on the std of dis- sign parameters of underground engineering(Lin et al. 2015; Fan
placement, since the std increases significantly with autocorrelation et al. 2021). In this section, the RFEM is used to study the local fail-
distance as shown in Fig. 14(b). For example, the minimum ure probabilities of the surrounding rock.

© ASCE 04022129-13 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


(a)
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

(b)

Fig. 14. Variation of the deformation with autocorrelation distances: (a) variation with the horizontal autocorrelation distances; and (b) variation with
vertical autocorrelation distances.

(a) (b)

Fig. 15. Variation of the deformation with autocorrelation distances in the #8 monitoring point: (a) variation with horizontal autocorrelation dis-
tances; and (b) variation with vertical autocorrelation distances.

The typical results of MCS are plotted in Fig. 16 and, as we can the powerhouse. The failure probability of the element at a certain
see, for each calculation, the element of failure varies due to the depth near the excavation boundary of the right spandrel is equal to
change of parameters. Using Eq. (42), the failure probabilities of 100%, which means that although the strength parameters calcu-
each element are obtained, which are equal to the number of lated each time are different, the element at this position will always
each element entering the plastic zone divided by the total number enter the plastic state. After reaching a certain depth from the
of MCS. As seen from Fig. 17, the zones of high failure probability excavation boundary, the failure probability of the element de-
are mainly located at the spandrel on the right side and bottom of creases rapidly. Fig. 18 shows the variation of failure probability

© ASCE 04022129-14 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

(c) (d)

Fig. 16. Typical results of Monte Carlo simulations: (a) the case I of failure zone; (b) the case II of failure zone; (c) the case III of failure zone; and (d)
the case IV of failure zone.

10 m. The results show that in the range of 0 to 5 m, the failure


probability of the element is 100% no matter how the autocorrelation
distance changes. The failure probability of the element decreases
rapidly when the depth exceeds 5 m, and it decreases to 0 in all cal-
culation conditions when the depth reaches 10 m (Cases R1 to R20).
However, the correlation distance still has a significant effect on the
failure probability. For example, as plotted in Fig. 18(a), the failure
probability of surrounding rocks at the depth of 8 m is about 2.1%
for the calculated work condition of R1, which corresponds to lh =
1 m and lv = 30 m; however, the failure probability of element in-
creases to 28% in the calculated work condition of R10, correspond-
ing to lh = 320 m and lv = 30 m. Generally, the failure probability
tends to increase with the increasing autocorrelation distance.
In addition, the field investigation found that there are extensive
Fig. 17. Distribution of failure probability of surrounding rock based cracking and falling of lining at the right spandrel of the main pow-
on the R1 calculation case.
erhouse, such as Sections R0 + 050 and R0 + 105, shown in
Fig. 19. This phenomenon indicated that the local stability state
of element along with depth with different autocorrelation dis- at the right spandrel is not ideal, it is consistent with the calculation
tances. The location of the monitoring line is marked in Fig. 18 results shown in Fig. 19, as the local failure probability of this po-
with the straight line as well, the length of the monitoring line is sition is equal to 1. In the design process, the supports of this part

© ASCE 04022129-15 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


(a) (b)

Fig. 18. Variation of failure probability corresponding to different autocorrelation distances: (a) variation with horizontal autocorrelation distances;
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

and (b) variation with vertical autocorrelation distances.

site-specific distributions of rock parameters and the cross-


correlation coefficient between c and ϕ. The effectiveness of the
method is validated with the test data from numerous triaxial com-
pression tests of marble from the Chinese Jinping II hydropower
station, which proves that this method can provide a reasonable es-
timation of rock parameters distribution from limited test data.
The acquired rock parameters are used for probability analysis
of the Jinping II underground powerhouse. A method for evaluat-
ing the spatial failure probability of underground cavern excavation
is developed, which can be used to evaluate the probability of en-
tering the failure state at each position in the calculation area. The
effects of autocorrelation distance on the spatial failure probability
and boundary excavation deformation are also studied. Generally,
this paper presents a scheme of probability design that includes pa-
(a) rameter estimation, parameter reduction from the rock to a rock
mass, deformation probability characterization, and spatial failure
probability estimation, which can provide a reference for the reli-
ability design of underground engineering.

Data Availability Statement

All data, models, or code that support the findings of this study are
available from the corresponding author upon reasonable request.

Acknowledgments

The authors gratefully acknowledge the financial support from the


(b) National Natural Science Foundation of China (Nos. U1965205
and 51779251) and S&T program of China Huaneng Group (No.
Fig. 19. Damage investigation of underground powerhouse: (a) block HNKJ21-HF317). Jiang designed the study and wrote the content;
falling of lining at R0 + 50; and (b) cracking of lining at R0 + 105. Liu carried out the calculation, data analysis, and wrote the content;
Zheng organized the experiment and technical analysis; Wang im-
proved the data exhibition; Guo, Chen, and Xiong took part in the
need to be strengthened and the length of bolts should be able to experimental analysis. All authors have read and approved the final
control the failure depth as plotted in Fig. 17. manuscript.

References
Summary and Conclusion
ABAQUS. 2011. Abaqus 6.11 user’s manual. Providence, RI: SIMULIA.
First, the application of the equivalent sample approach in under- Ang, A. H. S., and W. H. Tang. 2007. Probability concepts in engineering:
ground engineering rock masses is extended. This approach inte- Emphasis on applications to civil and enviromental engineering.
grates the prior knowledge on parameters with limited test data Chichester, UK: Wiley.
and transforms them into a large number of equivalent samples Asem, P., and P. Gardoni. 2019. “Bayesian estimation of the normal and
based on MCMC simulation, which can obtain the meaningful shear stiffness for rock sockets in weak sedimentary rocks.”

© ASCE 04022129-16 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Int. J. Rock Mech. Min. Sci. 124: 104129. https://doi.org/10.1016/j Feng, X., and R. Jimenez. 2015. “Estimation of deformation modulus of
.ijrmms.2019.104129. rock masses based on Bayesian model selection and Bayesian updating
Behzadafshar, K., M. E. Sarafraz, M. Hasanipanah, S. F. F. Mojtahedi, and approach.” Eng. Geol. 199: 19–27. https://doi.org/10.1016/j.enggeo
M. M. Tahir. 2019. “Proposing a new model to approximate the elastic- .2015.10.002.
ity modulus of granite rock samples based on laboratory tests results.” Gravanis, E., L. Pantelidis, and D. V. Griffiths. 2014. “An analytical solu-
Bull. Eng. Geol. Environ. 78 (3): 1527–1536. https://doi.org/10.1007 tion in probabilistic rock slope stability assessment based on random
/s10064-017-1210-5. fields.” Int. J. Rock Mech. Min. Sci. 71: 19–24. https://doi.org/10
Cai, M. 2010. “Practical estimates of tensile strength and Hoek–Brown .1016/j.ijrmms.2014.06.018.
strength parameter mi of brittle rocks.” Rock Mech. Rock Eng. 43 (2): Griffiths, D. V., and G. A. Fenton. 2004. “Probabilistic slope stability anal-
167–184. https://doi.org/10.1007/s00603-009-0053-1. ysis by finite elements.” J. Geotech. Geoenviron. Eng. 130 (5): 507–
Cai, M., P. K. Kaiser, H. Uno, Y. Tasaka, and M. Minami. 2004. 518. https://doi.org/10.1061/(ASCE)1090-0241(2004)130:5(507).
“Estimation of rock mass deformation modulus and strength of jointed Griffiths, D. V., and G. A. Fenton. 2009. “Probabilistic settlement analysis
hard rock masses using the GSI system.” Int. J. Rock Mech. Min. Sci. by stochastic and random finite-element methods.” J. Geotech.
41 (1): 3–19. https://doi.org/10.1016/S1365-1609(03)00025-X. Geoenviron. Eng. 135 (11): 1629–1637. https://doi.org/10.1061
Cao, Z., and Y. Wang. 2013. “Bayesian approach for probabilistic site char- /(ASCE)GT.1943-5606.0000126.
acterization using cone penetration tests.” J. Geotech. Geoenviron. Eng. Hastings, W. K. 1970. “Monte Carlo sampling methods using Markov
139 (2): 267–276. https://doi.org/10.1061/(ASCE)GT.1943-5606 Chains and their applications.” Biometrika 57 (1): 97–109. https://doi
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

.0000765. .org/10.1093/biomet/57.1.97.
Cao, Z., and Y. Wang. 2014a. “Bayesian model comparison and character- He, L., Y. Liu, S. Bi, L. Wang, M. Broggi, and M. Beer. 2020. “Estimation
ization of undrained shear strength.” J. Geotech. Geoenviron. Eng. of failure probability in braced excavation using Bayesian networks
140 (6): 04014018. https://doi.org/10.1061/(ASCE)GT.1943-5606 with integrated model updating.” Underground Space 5 (4): 315–323.
.0001108. https://doi.org/10.1016/j.undsp.2019.07.001.
Cao, Z., and Y. Wang. 2014b. “Bayesian model comparison and selection Hoek, E. 2002. “Hoek-Brown failure criterion-2002 edition.” In Vol. 1 of
of spatial correlation functions for soil parameters.” Struct. Saf. 49: 10– Proc., 5th North American Rock Mech. Symp. and 17th Tunneling
17. https://doi.org/10.1016/j.strusafe.2013.06.003. Assoc. of Canada Conf. Mining Innovation and Technology, 167–
Cao, Z., Y. Wang, and D. Li. 2016a. “Quantification of prior knowledge in 273. Toronto, Canada: University of Toronto Press.
geotechnical site characterization.” Eng. Geol. 203: 107–116. https://doi Hoek, E., and E. T. Brown. 1997. “Practical estimates of rock mass
.org/10.1016/j.enggeo.2015.08.018. strength.” Int. J. Rock Mech. Min. Sci. 34 (8): 1165–1186. https://doi
Cao, Z.-J., Y. Wang, and D.-Q. Li. 2016b. “Site-specific characterization of .org/10.1016/S1365-1609(97)80069-X.
soil properties using multiple measurements from different test proce- Hoek, E., and M. S. Diederichs. 2006. “Empirical estimation of rock mass
dures at different locations – A Bayesian sequential updating ap- modulus.” Int. J. Rock Mech. Min. Sci. 43 (2): 203–215. https://doi.org
proach.” Eng. Geol. 211: 150–161. https://doi.org/10.1016/j.enggeo /10.1016/j.ijrmms.2005.06.005.
.2016.06.021. Jiang, Q., S. Zhong, J. Cui, X.-T. Feng, and L. Song. 2016. “Statistical
Chen, D., D. Xu, G. Ren, Q. Jiang, G. Liu, L. Wan, and N. Li. 2019. characterization of the mechanical parameters of intact rock under triax-
“Simulation of cross-correlated non-Gaussian random fields for layered ial compression: An experimental proof of the Jinping marble.” Rock
rock mass mechanical parameters.” Comput. Geotech. 112: 104–119. Mech. Rock Eng. 49 (12): 4631–4646. https://doi.org/10.1007/s00603
https://doi.org/10.1016/j.compgeo.2019.04.012. -016-1054-5.
Chen, F., and W. Zhang. 2021. “Influence of spatial variability on the uni- Jiang, S.-H., I. Papaioannou, and D. Straub. 2018. “Bayesian updating of
axial compressive responses of rock pillar based on 3D random field.” slope reliability in spatially variable soils with in-situ measurements.”
ASCE-ASME J. Risk Uncertainty Eng. Syst. Part A: Civ. Eng. 7 (3): Eng. Geol. 239: 310–320. https://doi.org/10.1016/j.enggeo.2018.03
04021035. https://doi.org/10.1061/AJRUA6.0001162. .021.
Chen, L., W. Zhang, Y. Zheng, D. Gu, and L. Wang. 2020. “Stability anal- Li, D.-Q., S.-H. Jiang, Z.-J. Cao, W. Zhou, C.-B. Zhou, and L.-M. Zhang.
ysis and design charts for over-dip rock slope against bi-planar sliding.” 2015. “A multiple response-surface method for slope reliability analysis
Eng. Geol. 275: 105732. https://doi.org/10.1016/j.enggeo.2020 considering spatial variability of soil properties.” Eng. Geol. 187: 60–
.105732. 72. https://doi.org/10.1016/j.enggeo.2014.12.003.
Cheung, R. W. M., and W. H. Tang. 2005. “Realistic assessment of slope Li, H.-Z., and B. K. Low. 2010. “Reliability analysis of circular tunnel
reliability for effective landslide hazard mangement.” Géotechnique under hydrostatic stress field.” Comput. Geotech. 37 (1–2): 50–58.
55 (1): 85–94. https://doi.org/10.1680/geot.2005.55.1.85. https://doi.org/10.1016/j.compgeo.2009.07.005.
Ching, J., and K.-K. Phoon. 2013. “Effect of element sizes in random field Li, S., H. Zhao, M. Zheng, and H. Ruan. 2020. “Probabilistic predictions of
finite element simulations of soil shear strength.” Comput. Struct. 126: the convergences of surrounding rock masses in underground rock cav-
120–134. https://doi.org/10.1016/j.compstruc.2012.11.008. erns.” Int. J. Geomech. 20 (5): 04020038. https://doi.org/10.1061
Ching, J., S.-S. Wu, and K.-K. Phoon. 2016. “Statistical characterization of /(ASCE)GM.1943-5622.0001659.
random field parameters using frequentist and Bayesian approaches.” Lin, P., H. Liu, and W. Zhou. 2015. “Experimental study on failure behav-
Can. Geotech. J. 53 (2): 285–298. https://doi.org/10.1139/cgj-2015 iour of deep tunnels under high in-situ stresses.” Tunnelling
-0094. Underground Space Technol. 46: 28–45. https://doi.org/10.1016/j.tust
Contreras, L. F., E. T. Brown, and M. Ruest. 2018. “Bayesian data analysis .2014.10.009.
to quantify the uncertainty of intact rock strength.” J. Rock Mech. Liu, H., and X. Qi. 2018. “Random field characterization of uniaxial com-
Geotech. Eng. 10 (1): 11–31. https://doi.org/10.1016/j.jrmge.2017.07 pressive strength and elastic modulus for intact rocks.” Geosci. Front.
.008. 9 (6): 1609–1618. https://doi.org/10.1016/j.gsf.2017.11.014.
Cui, J., Q. Jiang, S. Li, X. Feng, M. Zhang, and B. Yang. 2017. “Estimation Mollon, G., D. Dias, and A.-H. Soubra. 2011. “Probabilistic analysis of
of the number of specimens required for acquiring reliable rock me- pressurized tunnels against face stability using collocation-based sto-
chanical parameters in laboratory uniaxial compression tests.” Eng. chastic response surface method.” J. Geotech. Geoenviron. Eng.
Geol. 222: 186–200. https://doi.org/10.1016/j.enggeo.2017.03.023. 137 (4): 385–397. https://doi.org/10.1061/(ASCE)GT.1943-5606
Eberhardt, E. 2012. “The hoek–brown failure criterion.” Rock Mech. Rock .0000443.
Eng. 45 (6): 981–988. https://doi.org/10.1007/s00603-012-0276-4. Moradian, Z. A., and M. Behnia. 2009. “Predicting the uniaxial compres-
Fan, Q., P. Lin, P. Wei, N. Zeyu, and G. Li. 2021. “Closed-loop control the- sive strength and static young’s modulus of intact sedimentary rocks
ory of intelligent construction.” J. Tsinghua Univ. 61 (07): 660–670. using the ultrasonic test.” Int. J. Geomech. 9: 14–19. https://doi.org
Feng, X., and R. Jimenez. 2014. “Bayesian prediction of elastic modulus of /10.1061/(ASCE)1532-3641(2009)9:1(14).
intact rocks using their uniaxial compressive strength.” Eng. Geol. 173: Nomikos, P. P., and A. I. Sofianos. 2011. “An analytical probability distri-
32–40. https://doi.org/10.1016/j.enggeo.2014.02.005. bution for the factor of safety in underground rock mechanics.”

© ASCE 04022129-17 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129


Int. J. Rock Mech. Min. Sci. 48 (4): 597–605. https://doi.org/10.1016/j Wang, Y., and Z. Cao. 2013. “Probabilistic characterization of Young’s
.ijrmms.2011.02.015. modulus of soil using equivalent samples.” Eng. Geol. 159: 106–118.
Phoon, K.-K., and F. H. Kulhawy. 1999a. “Characterization of geotechni- https://doi.org/10.1016/j.enggeo.2013.03.017.
cal variability.” Can. Geotech. J. 36 (4): 612–624. https://doi.org/10 Wang, Y., T. Zhao, and Z. Cao. 2015. “Site-specific probability distribution
.1139/t99-038. of geotechnical properties.” Comput. Geotech. 70: 159–168. https://doi
Phoon, K.-K., and F. H. Kulhawy. 1999b. “Evaluation of geotechnical .org/10.1016/j.compgeo.2015.08.002.
property variability.” Can. Geotech. J. 36 (4): 625–639. https://doi Xu, J., L. Zhang, J. Li, Z. Cao, H. Yang, and X. Chen. 2021. “Probabilistic
.org/10.1139/t99-039. estimation of variogram parameters of geotechnical properties with a
Pinheiro, M., J. Vallejos, T. Miranda, and X. Emery. 2016. “Geostatistical trend based on Bayesian inference using Markov chain Monte Carlo
simulation to map the spatial heterogeneity of geomechanical parame- simulation.” Georisk: Assess. Manage. Risk Eng. Syst. Geohazards
ters: A case study with rock mass rating.” Eng. Geol. 205: 93–103. 15 (2): 83–97. https://doi.org/10.1080/17499518.2020.1757720.
https://doi.org/10.1016/j.enggeo.2016.03.003. Zhang, J., J. P. Li, L. M. Zhang, and H. W. Huang. 2014. “Calibrating cross-
Sari, M., and C. Karpuz. 2006. “Rock variability and establishing confining site variability for reliability-based design of pile foundations.” Comput.
pressure levels for triaxial tests on rocks.” Int. J. Rock Mech. Min. Sci. Geotech. 62: 154–163. https://doi.org/10.1016/j.compgeo.2014.07.013.
43 (2): 328–335. https://doi.org/10.1016/j.ijrmms.2005.06.010. Zhang, J., W. H. Tang, L. M. Zhang, and H. W. Huang. 2012.
Sheorey, P. R. 1997. Empirical rock failure criteria. London: CRC “Characterising geotechnical model uncertainty by hybrid Markov
Press. Chain Monte Carlo simulation.” Comput. Geotech. 43: 26–36. https://
Downloaded from ascelibrary.org by Quan Jiang on 06/07/22. Copyright ASCE. For personal use only; all rights reserved.

Song, K. I., G. C. Cho, and S. W. Lee. 2011. “Effects of spatially variable doi.org/10.1016/j.compgeo.2012.02.002.
weathered rock properties on tunnel behavior.” PrEM 26 (3): 413–426. Zheng, M., S. Li, H. Zhao, X. Huang, and S. Qiu. 2021. “Probabilistic anal-
Sonmez, H., and R. Ulusay. 1999. “Modifications to the geological strength ysis of tunnel displacements based on correlative recognition of rock
index (GSI) and their applicability to stability of slopes.” Int. J. Rock mass parameters.” Geosci. Front. 12 (4): 101136. https://doi.org/10
Mech. Min. Sci. 36 (6): 743–760. https://doi.org/10.1016/S0148 .1016/j.gsf.2020.12.015.
-9062(99)00043-1. Zhou, J., Y. Qiu, S. Zhu, D. J. Armaghani, M. Khandelwal, and E. T.
Wang, Y., and O. V. Akeju. 2016. “Quantifying the cross-correlation be- Mohamad. 2021. “Estimation of the TBM advance rate under hard
tween effective cohesion and friction angle of soil from limited rock conditions using XGBoost and Bayesian optimization.”
site-specific data.” Soils Found. 56 (6): 1055–1070. https://doi.org/10 Underground Space 6 (5): 506–515. https://doi.org/10.1016/j.undsp
.1016/j.sandf.2016.11.009. .2020.05.008.

© ASCE 04022129-18 Int. J. Geomech.

Int. J. Geomech., 2022, 22(8): 04022129

You might also like