Computers and Geotechnics 76 (2016) 23–32

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

A novel parameter identification approach for buffer elements involving

complex coupled thermo-hydro-mechanical analyses
Long Nguyen-Tuan a,⇑, Tom Lahmer a, Maria Datcheva b,d, Eugenia Stoimenova c, Tom Schanz b
a
Institute of Structural Mechanics, Bauhaus-Universität Weimar, Germany
b
Chair of Foundation Engineering, Soil and Rock Mechanics, Ruhr-Universität Bochum, Germany
c
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, Bulgaria
d
Institute of Mechanics, Bulgarian Academy of Sciences, Bulgaria

a r t i c l e i n f o a b s t r a c t

Article history: Coupled thermo-hydro-mechanical (THM) analyses contain a large set of constitutive parameters and
Received 22 May 2015 this requires numerous experiments to determine these parameters. This study contributes to the
Received in revised form 20 January 2016 identification of parameters of a coupled THM constitutive model via back analysis of information-rich
Accepted 10 February 2016
experiments. An iterative sampling based back analysis approach is proposed comprising both the model
Available online 3 March 2016
parameter identification and the assessment of the reliability of identified model parameters. The results
obtained in the context of buffer elements indicate that sensitive parameter estimates generally follow
Keywords:
the normal distribution. According to the sensitivity of the parameters and the probability distribution
Parameter identification
Coupled THM problem
of the samples we can provide confidence intervals for the estimated parameters and thus allow a qual-
Sensitivity analysis itative estimation on the identified parameters which are in future work used as inputs for computational
Particle swarm optimisation predictions in high-risk situations.
Confidence interval Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction Generally, there are several challenges while conducting model-

Back analysis based on non-linear optimisation techniques is
nowadays a well-accepted approach to calibrate models against
sets of measured data in geotechnical engineering. Several authors
highlight such types of parameter identification approaches as
suitable ways to solve geotechnical problems in higher accuracy,
including [1–7]. The advantage of using means of non-linear opti-
misation is that there is no need to invert the governing equations
with respect to the involved model parameters, as such inversion is
not always easy or possible. The main discussion in the literature
with respect to back analysis approaches has so far addressed
the formulation of the objective function and the optimisation
algorithm. In the field of the optimisation based back analysis,
some papers focus on the development of a superior algorithm of
the optimisation which has global-search capacity and acceptable
computational costs, e.g., [5,8–10]. Other authors worked on the
formulation of complex objective functions, so that the parameter
set minimising the objective function can be reliably used for
numerical simulations and for making predictions based on these
simulations [2,3,11].
⇑ Corresponding author.
based identifications of the constitutive parameters of complex
models. One main challenge is that the provided measurements
usually cannot be expected to be in the range of the operator link-
ing model input parameters to model output. This is due to noise
generally present in the measurements and abstractions during
the model building process. Therefore, a direct invertibility of the
systems is not possible and we seek for solutions of the inverse
(i.e. the calibration) problem in the sense of least-squares solu-
tions. Such solutions always exist, however, there is no guarantee
that they will be close to the exact solution and that they are
unique. The uniqueness issue can be addressed by restricting
search spaces or by providing good initial guesses, if available,
and regularisation, for instance according to Tikhonov.
Accepting the concept of a least-squares solution for the inverse
problem, it raises another question of how to efficiently find the
solution. According to the ‘‘no-free-lunch-theorem” [12] there is
no strategy which performs better than others on average for all
possible optimisation problems. Inspired by good experiences from
researchers in the field of geotechnics, we employ meta-heuristic
approaches which in particular guarantee good solutions for non-
convex complex problems. Besides the above mentioned problems
another challenge arises, namely the stable dependency of the
solution on the data provided. Assume that there are small changes

http://dx.doi.org/10.1016/j.compgeo.2016.02.005
0266-352X/Ó 2016 Elsevier Ltd. All rights reserved.
24 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32

in the data, will this lead to visible changes in the numerical values 2. Parameter identification via back analysis
of the solution? What happens, if we run the applied meta-
heuristic search method, which is based on random number gener- 2.1. Back analysis strategy
ators, several times? To quantify these effects, repeated parameter
identifications are performed and statistics on the results are con- The back analysis strategy employed in this study is illustrated
ducted and interpreted. in Fig. 1. Firstly, the mathematical models for the forward calcula-
It is necessary to point out that the applications of estimating tion are selected. In coupled THM analysis, we use multi-physical
the model parameters for a geotechnical problem based on a statis- relations described in Table 1, which are implemented in the finite
tical approaches, e.g. the maximum likelihood method applied by element code, CODE_BRIGHT [17]. Afterwards, the numerical solu-
[11] or the Bayesian method applied by [3], have difficulties in tion of the forward problem is compared with experimental data
the considered models due to the lack of prior information at least by means of an objective function i.e. a weighted sum of squared
about the empirical parameters involved in the constitutive equa- errors. The objective function is minimised by means of non-
tions. Some researchers suggested that the confidence in the iden- linear optimisation algorithms, in particular by the PSO method
tified parameters can be assessed based on sensitivity measures [15]. The initial values of parameters are randomly selected follow-
e.g. [4] or [7]. However, their approaches do not cope with the dif- ing the uniform distribution within their prescribed ranges of vari-
ficulty of estimating a general sensitivity index for the responses of ation. The sampling process generates np samples of parameters by
a model in a multi-field context as studied in this work. For means of performing np times of the optimisation using the PSO
instance, one parameter might be a sensitive parameter in a partic- method. Next, a sensitivity analysis is carried out and the confi-
ular single-field model but it may be a non-sensitive parameter for dence intervals are determined. The confidence interval theory is
any other quantity in a coupled system. Therefore, estimating the based on assumption that the samples follow a normal distribu-
confidence in identified parameters only based on sensitivity tion, therefore, a normality test has to be performed before calcu-
indices for one output quantity might be less reliable for multi- lating the confidence intervals. A model parameter xj ðj ¼ 1; . . . ; JÞ
field problems. is called a sensitive parameter when the sum of its sensitivity
Particularly, identifying parameters for the coupled THM analy- indices ðRSk Þj is greater than a predefined value S0 , where k denotes
sis in unsaturated soils is a complex problem due to a large set of
the considered model responses (k ¼ 1; 2; . . . ; K). The confidence
parameters and a variety of variables in the forward calculation
intervals are used to assess the reliability of the identified param-
(i.e. displacements, temperature, pore water pressure and air pres-
sure). Some authors attempted to identify constitutive model eters only for those parameters that are sensitive (i.e. ðRSk Þj > S0 )
parameters for unsaturated soils by means of back analysis, for and their samples follow the normal distribution. The normality
instance [13,14]. They drew an objective function in different sub- test is done according to Shapiro–Wilk [18]. The confidence inter-
spaces and found the minimum of the objective function assuming vals are determined based on the probability distribution of the
that the other parameters are kept constant. In fact, model param- samples, see [16]. The back analysis procedure is presented in
eters vary in the search space during searching process, besides detail in the next subsections.
that the confidence of the identified parameters has not been
assessed. In the paper [10], model parameters for describing the
2.2. Objective function
elasto-plastic behaviour have been identified for unsaturated soils.
The authors have chosen six parameters for identification based on
Let ymeas be a set of observed data from an experiment and
qualitative arguments. However, the quality of the identified
ycalc be a set of obtained data by numerical simulation of this
parameters has not been assessed quantitatively.
experiment depending on a vector of model parameters
Therefore, in this paper, a novel back analysis approach is pro-
posed comprising a model parameter identification and an assess-
ment of the reliability of the identified model parameters.
Parameter sampling process is carried out based on meta-
heuristic optimisation methods, in which parameters are varied
under the control of the computational paradigms such as the Par-
ticle Swarm Optimisation (PSO) [15]. The confidence intervals are
determined based on these parameter samples by means of prob-
ability distribution functions, see e.g. [16]. The approach can be
applied for back analysing a variety of geotechnical problems, in
particular it is well suited for parameter identification problems
including large sets of model parameters and multi-physical
simulations.
To show the applicability of the proposed method, it is applied
to identify the model parameters for the simulation of the beha- Fig. 1. Schematic presentation of the back analysis strategy.
viour of buffer elements in high-level waste facilities. In proposed
designs of repositories for the isolation of high-level radioactive
waste, the clayey buffer elements play the role as engineered bar-
riers. The behaviour of the clay barrier is highly complex. It Table 1
Constitutive relations in coupled THM analysis.
involves coupled THM phenomena, which take place due to the
simultaneous heating (generated by the radioactive waste) and Variables Constitutive equation Notation
hydrating of the barrier (due to the inflow of water from the sur- Liquid and gas advective flux Darcy’s law ql ; qg
rounding rock) and mechanical forces (due to swelling phe- Vapour and air non-advective Fick’s law iw a
g ; il
nomenon of the buffer). It requires a fully coupled non-linear fluxes
THM numerical analysis for simulating water/vapour transport, Conductive heat flux Fourier’s law ic
Liquid phase degree of saturation Retention curve Sl ; Sg
heat conduction, and modelling of complex thermo-elasto-plastic
Stress tensor Mechanical constitutive model r
stress–strain behaviour.
 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32 25

x ¼ fx1 ; x2 ; . . . ; xJ g. The difference between the calculated and mea-  We consider the function f : X # RJ ! R as the objective func-
sured values defines the error tion to be minimised due to prescribed constraints.
meas calc
 We start with n parameter vectors. Each vector has values
d
f tm ðxÞ ¼ fydtm g  fydtm ðxÞg : ð1Þ defined by the vector of model parameters xi 2 X and is associ-
ated with an updating vector v i for i ¼ 1; 2; . . . ; n.
A function of the model parameters is expressed as a nor-
 The updating vector of the l  th iteration is computed as:
malised weighted sum of the squared errors
   
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u 2
v i;lþ1 ¼ Ul v i;l þ c1 r1 pil  xi;l þ c2 r 2 plg  xi;l ð4Þ
uXP XM ½f tm ðxÞ xdtm
d
f MAE ðxÞ ¼ u
d
t t¼1 m¼1 P P h i2 ; ð2Þ and the updated entries of every parameter vector at l  th iter-
P M d meas
t¼1 m¼1 ðytm Þ ation computes as
xi;lþ1 ¼ xi;l þ v i;lþ1 : ð5Þ
with t ¼ 1; 2; . . . ; P, where P is the number of measured data accord-
ing to time, m ¼ 1; 2; . . . ; M, where M is the number of selected
points for observation, d ¼ 1; 2; . . . ; D is the number of serial mea- The algorithm is described as follows:
surements, for instance temperature, degree of saturation, or stress
in one location and xdtm is the weighting factor for each measure- 1. Initialise:
ment. Values of xdtm depend on the importance and reliability of (a) Set constants lmax ; c1 ; c2 .
the analysed data. In Eq. (2), it is assumed that (b) Randomly initialise the parameter vectors xi0 2 D½xmin ; xmax 
PP PM h d meas i2 in RJ for i ¼ 1; 2; . . . ; n.
t¼1 m¼1 ytm is larger than zero.
(c) Randomly initialise the initial parameter updated
Finally, considering the multi-field data, such as D different 0 6 v i0 6 v max
0 for i ¼ 1; 2; . . . ; n.
serial measurements, the objective function is defined as (d) Set l ¼ 1.
2. Minimise:
1X D
ðxÞxd ;
d
f ðxÞ ¼ f ð3Þ i
(a) Evaluate the function value f l for values xil according to Eq.
D d¼1 MAE
(3).
i i i i
where xd is the weighting factor for each serial measurement. (b) If f l 6 f best then f best ¼ f l ; pil ¼ xil .
i g g i
(c) If 6 fl f best
then ¼ f l ; plg ¼ xil .
f best
2.3. Optimisation algorithms (d) If a stopping criterion is satisfied then go to 3.
(e) Compute all parameter updates v il for i ¼ 1; 2; . . . ; n.
In general, the optimisation methods can be classified in the fol- (f) Update all parameter values xil for i ¼ 1; 2; . . . ; n.
lowing groups: sampling-based stochastic methods, gradient- (g) Increase l.
based methods, direct search methods and population-based (h) Go to 2(a).
methods. Among these groups, population-based methods provide 3. Terminate the program with f opt ¼ f best .
g

global search and the results are independent of the initial guess,
however, they are relatively expensive. These methods are suitable
for the models having a high number of model parameters for
which initial guesses are difficult to establish. PSO [15] was
selected as an optimisation algorithm due to the fact that for some
parameters no initial guess can be provided, only parameter
ranges, and the complexity of the forward problem supports the
assumption of a non-convex optimisation problem.
Particle swarm optimisation works basically by considering a
population (called a ‘‘swarm”) of candidate solutions of the optimi-
sation problem (called particles). These particles are moved around
in the search space according to several simple laws. The move-
ments of the particles are guided by their own best known position
in the search space as well as the entire swarm’s best known posi-
tion. Translating this to parameter identification problems, we con-
sider a collection (swarm) of parameter vectors. Now, the entries in
the vectors are repeatedly updated by combining local and global
information about the values of the objective functions for the dif-
ferent parameter vectors. When better vectors are discovered, they
will determine the further updates of the parameters. This process
is repeated iteratively and by doing so, it is expected that a satis-
factory solution of the calibration problem will eventually be dis-
covered with high probability.
The PSO algorithm is presented basically as follows: Let us call
pil the actual best configuration of the i  th parameter vector and
plg is the parameter value which provides the smallest value of
the objective function (globally best parameter vector) in the
l  th iteration. Let now c1 and c2 be two constants and r1 and r 2
are random numbers between 0 and 1, U is called the inertia func-
tion which is also modelled as uniformly distributed random num-
ber. In detail:
We chose the values of optimisation parameters as
c1 ¼ c2 ¼ 0:8. There is no general rule to select these values opti-
mally. For further information about these parameters, one may
see e.g. [19,20] for particular problems. The above algorithm is
implemented in the software VARO2PT, see [21], values recom-
mended therein have been used which provided a sufficiently effi-
cient algorithm.
2.4. Sensitivity analysis
The LSA (Local Sensitivity Analysis) is carried out for the point
in the parameter space that was identified as an optimal after per-
forming the back analysis. The sensitivity analysis is presented
hereafter based on the partial derivative based method described
in [6]. For each increment of a parameter xj with the value of @xj ,
we receive a value of model responses yktm and the deviation @yktm .
By utilising the one-at-a-time sampling technique, using the centre
value of the parameter xj , the parameter xj changes the amount of a
step size @xj , and the other parameters are fixed at the centre value.
The centre value is chosen as a mean value of the parameter sam-
ples obtained by the sampling process. The procedure is as follows:
2.4.1. Determining scaled sensitivity (SS)
th
The scaled sensitivity of the j parameter is written as follows,
@yktm
SSktmj ¼ xj ð6Þ
@xj
where yktm is an entry of a vector of the model response at time t
(t ¼ 1; 2; . . . ; P) at the selected point m of the geometrical model, k
26 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32
denotes the model response (k ¼ 1; 2; . . . ; K) as in Section 2.1, K ¼ 3
(saturation, stress and temperature) in this paper.
2.4.2. Determining composite scaled sensitivity (CSS)
The overall model sensitivity to a given model parameter xj is
assessed by a composite scaled sensitivity over the entire duration
of the numerical simulation:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 X P X M h k i2
CSSkj ¼ SStmj :
PM t¼1 m¼1
2.4.3. Determining the sensitivity index SðjÞ for each of the parameters
The sensitivity index is used as a normalisation of CSSðjÞ in the
range from zero to one
CSSkj
Skj ¼ : ð8Þ
maxj ðCSSkj Þ
It is suggested in this paper that a parameter xj is considered a sen-
P k
sitive parameter when the sum of its sensitivity indices k S is
greater than S0 ¼ 0:5, which is a value chosen by experience and
which may be adapted for other problems.
2.5. Confidence intervals of the parameters
According to the scheme presented in Fig. 1, the confidence
interval is used to assess the reliability of the parameter only if
the parameter is a sensitive parameter and the distribution of its
samples follows the normal distribution. Otherwise, the samples
of the parameter cannot be used for determining the confidence
interval. It will be discussed in detail in Section 3.3.
The distribution of the parameter samples is verified through
the test of Shapiro–Wilk [18], in order to confirm the normal distri-
bution of the samples. For the normality test, the W statistic is cal-
culated as follows,
Pnp 2
n¼1 an xj ðnÞ
Wj ¼ h i2 ;
Pnp
n¼1 xj ðnÞ  lj
where lj is the mean value of parameter xj ; n is number of the sam-
ples, an is constant generated from the means, variances and covari-
ances of the order statistics of a sample of size np and is given in
[22]. It is suggested that if W j > W 0 ¼ 0:9, the samples are consid-
ered to follow a normal distribution in this paper.
When the number of samples is smaller than 30, the Student t-
distribution is recommended to estimate the confidence intervals
[16,23]. Assuming an unbiased estimate of the sample variance
(sj ) of the j  th parameter, the confidence interval is estimated by
sj sj
lj  tu ðnp Þ pffiffiffiffiffi 6 xj 6 lj þ t u ðnp Þ pffiffiffiffiffi ;
np np
where t u ðnp Þ is the Student t percentile dependent on a percentile of
xj and sample size np ; t u ðnp Þ is the critical value of the Student t-
distribution. When the number of samples (np ) is greater than 30,
normal distribution is recommended to estimate the confidence
intervals. The formulas in [16,23] can be used to estimate the con-
fidence interval. Certainly, the greater the number np the more
accurate are the estimates of the confidence intervals. However,
np is limited by the computational costs of the forward problems.
For instance, each forward computation of the THM problem
expends 15 min, PSO method with 10 different parameter vectors
requires 150 iterations to attain the convergence criteria, conse-
quently one sample costs 15 days. It could cost more than a year
for np > 30 when computing on a single processor machine.
3. Back analysis of thermo-hydro-mechanical column
experiments
Back analysis based parameter identification is used widely for
calibrating parameters for geotechnical models. However, there are
still only few papers on the identification of model parameters for
coupled THM model based on back analysis approaches. The cou-
pled THM phenomenon is recently attracting the attention of the
researchers and engineers involved in the conceptual design of
ð7Þ
engineered barriers for storage of spent nuclear fuel and nuclear
waste. The effects of the THM phenomena play an important role
to optimise the performance of the buffer system for the nuclear
waste containers and in order to reduce uncertainties in any pre-
dictions for such systems. In order to reproduce most closely in
the lab the behaviour of the buffer material, a THM column device
was constructed and its setup is given in [24]. The column device
enables to obtain variety of transient information about tempera-
ture, water content, soil suction and swelling pressure at different
locations along the column. The column can control at the same
time the boundary conditions for temperature, mechanical load
and water content in the sample. The measurement data can be
used to reveal the interaction among stress–strain distribution,
water/moisture flow, and heat conduction. Therefore, the THM col-
umn device can be considered as an information-rich experimental
setup in the context of multi-field problems.
Concerning the mathematical modelling of the behaviour of
buffer materials obeying high swelling potential, widely used for
THM analysis, a constitutive model developed at UPC Barcelona
[25,26]. In the following subsections, the constitutive models used
for the coupled THM analyses are presented.
3.1. Formulation of the forward model
The THM problem is simulated by solving a system of balance
equations enriched by a set of constitutive equations as described
in the following subsection. The system of balance equation is
updated and solved iteratively in the time-dependent problems.
Equations for mass balance were established following the compo-
ð9Þ
sitional approach [27]. The mass balance is prescribed for both liq-
uid and gas phases. The equation for the balance of energy is
established for the whole medium. The equation of momentum
balance in the porous medium is reduced to stress equilibrium
equation. Constitutive equations are used to connect the primary
unknowns (displacements, liquid pressure, gas pressure and tem-
perature) with parameters and dependent variables like water sat-
uration and energy flux. In the scope of this paper, we only
introduce constitutive equations to clarify the relations and
parameters.
3.1.1. Constitutive equations of the coupled thermo-elasto-plastic
model
ð10Þ
Gens et al. [28,29] drew the constitutive model under the
thermo-elasto-plastic theory for unsaturated soil (BBM-TEP). The
BBM-TEP model is adopted for stress–strain behaviour in CODE_B-
RIGHT Finite Element code. The increment of the strain is a sum of
the increment of the elastic and plastic strain components as
follows,
de ¼ dee þ dep : ð11Þ
Following the two stress variable concept in unsaturated soil
mechanics, the elastic part of the strain increment is taken to be
a sum of the increments of suction induced dese , net stress
induced dere and the strain increment due to temperature change
deTe . The final relation for the elastic strain increment reads:
dee ¼ dere þ dese þ deTe : ð12Þ
 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32 27

The nonlinear elastic law for the volumetric strain induced by The tensile strength equation is adopted from [29]. The tensile
the net stress is expressed in Eq. (13). strength ps follows a linear relationship with suction and is an
exponential function of temperature:
ji ðsÞ dp0
derv e ¼ and p0 ¼ p  maxðPg ; Pl Þ ð13Þ ps ¼ ps0 þ ks s expðqDTÞ; ð23Þ
1 þ e p0
where ks is the parameter that takes into account the increase in
where ji ðsÞ is defined by: tensile strength due to suction, ps0 is tensile strength in a saturated

jio ð1 þ ai sÞ if 1 þ ai s P 0:001 state, q is a parameter that takes into account the decrease in
ji ðsÞ ¼ ð14Þ
0:001 kio if 1 þ ai s < 0:001 tensile strength due to the temperature increment. The isotropic
hardening depends on the plastic volumetric strains epv as follows:
where p is the total mean stress and p0 ¼ p  maxðPg ; Pl Þ; Pg and P l
1þe
are the gas pressure and the liquid pressure, e is the void ratio and dpo ¼ p dep : ð24Þ
jio and ai are the model parameters. kso  jio o v
Suction and temperature induce only volumetric strains with
constitutive equations given as follows: 3.1.2. Constitutive equations of the hydraulic model
js ðp0 ; sÞ
ds Advective flow of the water phase is described by the gener-
dev
se
¼ ; dev
Te
¼ ao dT ð15Þ alised Darcy’s law as follows:
1 þ e s þ pat
kkrl
with ql ¼  ðrPl  ql g Þ; ð25Þ
ll
js ðp ; sÞ ¼ jso jsp exp ðass sÞ
0
ð16Þ
where ll is the dynamic viscosity of the pore liquid, g is the gravity
and
acceleration, ql is the liquid density. The tensor of intrinsic permeabil-
8 !
> 1020 ity k usually depends on porosity according to the Kozeny’s model:
>
> 1 þ a ln if p0 6 1020
>
> sp
>
> pref /3 ð1  /o Þ2
>
>  k ¼ ko ; ð26Þ
< 1 2
jsp ¼ 0 if p0 P pref exp ð17Þ ð1  /Þ /3o
>
>
>
asp
>
> ! where / is the porosity, /o is the reference porosity,
>
> p0
> ko ¼ diagðkxx ; kyy ; kzz Þ is the intrinsic permeability tensor for a media
: 1 þ asp ln
> elsewhere
pref with porosity /o . Assuming that the material is isotropic, ko is
described by solely one permeability value, i.e. ko ¼ kxx ¼ kyy ¼ kzz
The parameters involved are: ao is the thermal expansion coef-
where x; y and z are the Cartesian coordinates. The relative perme-
ficient, jso is the elastic stiffness parameter when changing suction
ability krl , is derived from Mualem-van Genuchten closed form
at zero net stress, pat is the atmospheric pressure, finally, ass and asp
model, [31]:
are model parameters. The elastic parameters ji and js may be
considered independent of temperature in the case of moderate pffiffiffiffiffi
k 2
temperature gradients. krl ¼ Se 1  1  S1=k
e ; ð27Þ
In summary, the volumetric elastic strain is the sum of volumet-
ric strain induced by effective stress, volumetric strain induced by where k is a shape parameter for the retention curve and Se is
suction and expansion/shrinkage due to temperature of material as defined as:
follows: 0 1k
 1 
ji ðsÞ dp 0
js ðp ; sÞ
0
ds Sl  Srl B Pg  Pl 1  k C rT
deev ¼ þ þ ao dT; ð18Þ Se ¼ ¼B
@1þ C
A and P ¼ P0 ;
1 þ e p0 1 þ e s þ pat Sls  Srl P rT0
The yield surface in BBM-TEP model is given in the deviatoric
ð28Þ
plane q  p via the following equation:
where Sl ; Sls and Srl are the current, the maximum and the residual
F ¼ q2  M 2 ðp0 þ ps Þðpo  p0 Þ ¼ 0; ð19Þ
liquid degree of saturation, respectively. P 0 is a model parameter, rT
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where q¼ 3 D
r : rD , with deviatoric stress defined as is the surface tension of liquid and rT0 is the surface tension of liq-
2
uid at which P0 was measured. The parameter rT is calculated by
rD ¼ r0  13 r0 : I. The pre-consolidation pressure po depends on suc- empirical relation as in Eq. (29).
tion and it is defined according to [28] as: 
252:93
rT ¼ 0:03059 exp
  kso  jio
: ð29Þ
273:15 þ T
p ðTÞ ks  jio
po ¼ pc o c ; ð20Þ
p Molecular diffusion of vapour in air is governed by Fick’s law.
v
The diffusive flux of water vapour i is written as:
where p is the reference pressure, kso is the plastic stiffness param-
c
 
v
eter, po ðTÞ is the pre-consolidation pressure in the saturated state. i ¼  /qv Sl Dm I rxv ; ð30Þ
po ðTÞ is defined as a non-linear function of temperature as in [30]
where qv is the vapour density, xv is the mass fraction of the
as follows:
vapour, I is the identity matrix and Dm is the diffusion coefficient
po ðTÞ ¼ po þ 2ða1 DT þ a3 DTjDTjÞ ð21Þ of vapour in m2 =s and is defined by:
where a1 and a3 are model parameters. The stiffness parameter as a ð273:15 þ T ÞnD
function of suction is defined as: Dm ¼ sDr ; ð31Þ
Pg
ks ¼ kso ½ð1  r Þ exp ðbsÞ þ r ; ð22Þ
where s is the tortuosity, Dr is the molecular diffusion coefficient at
where r and b are model parameters. temperature 273.15 K and nD is a model parameter.
28 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32

3.1.3. Constitutive equations for the heat conduction

Fourier’s law is adopted for the heat conduction flux, ic as
follows:
ð1S Þ
ic ¼ kT rT
S
where kT ¼ ksat
l
kdry l ; ð32Þ

where kT is the soil thermal conductivity, ksat and kdry are soil ther-
mal conductivity in the saturated and dry state, respectively.

3.1.4. Coupled THM forward analysis

The solution of the coupled THM problems is performed via the
finite element method. The constitutive relations are summarised
in Table 1. Due to the complexity of the constitutive model there
is a large number of model and material parameters involved.
The model parameters in the THM analysis are classified in groups
depending on their relations i.e. the stress–strain relation (net
stress driven processes), the hydraulic relation (suction driven pro-
cesses), and the temperature relation (temperature driven pro-
cesses). The aim is to clarify the influence of each group of
parameters on each response of the THM analysis. The input vector
(x) is composed of the following parameter sets:
Fig. 2. Description of numerical model geometries and boundary conditions: (a)
HM test, (b) THM test.
 Parameters involved in modelling net stress driven processes
(dr – 0):
 
M ¼ jio ; ai ; pref ; kso ; r; b; ks ; ps0 ; pc ; M; eo ; po . Table 2
Test and procedures.
 Parameters involved in modelling suction driven processes
(ds – 0): Test name Temperature Hydration Initial w(%)
 
H ¼ P 0 ; k; /0 ; ko ; js0 ; ass ; asp . (°C) (kPa)

 Parameters involved in modelling temperature driven processes
(dT – 0):
 
T ¼ s; Dr ; nD ; ksat ; kdry ; a0 ; a1 ; a3 ; q .
Therefore the vector of model parameters reads:
x ¼ ðH; T ; MÞ
The vector of serial model responses corresponding with mea-
surements ycalc is composed as follows:
 calc
ycalc ¼ ydtm ðxÞ ¼ ðSl ðtmÞ; ryy ðtmÞ; TðtmÞ; sðtmÞÞcalc
where Sl ðtmÞ is the degree of saturation, ryy ðtmÞ is the vertical
stress, TðtmÞ is the temperature and sðtmÞ is suction over time t at
observation point m.
3.2. THM column experiment and numerical simulation
Two different tests have been performed: (1) a hydro-mechanical
(HM) test and (2) a thermo-hydro-mechanical (THM) test, see [24]
and Table 2. The volume of the soil samples were kept constant dur-
ing these tests. In the HM test, water was supplied from the bottom
of the sample at 15 (kPa) water pressure (Fig. 2a). The temperature
was kept constant at 20 (°C) during the test. In the THM test, the pro-
cess was performed in two phases. The column sample was heated
from the bottom at 80 (°C) and cooled from the top at 20 (°C) in
the first phase. In the second phase, water was supplied from the
top at 5 (kPa) water pressure while keeping the thermal condition
as in the first phase (Fig. 2b). The temperature was measured by
thermocouple devices (TH1, TH2 and TH3). The water content was
measured by time domain refractometry (TDR) devices (TDR1,
TDR2 and TDR3). The suction was measured by relative humidity
sensors (RH1, RH2 and RH3). The vertical stress is measured by the
load cell at the top of the column. Therefore, there are totally 10 ser-
ies of measurements (D ¼ 10) in this study. The test results provide a
vector of serial measurements ymeas as follows
Top Bottom Top Bottom
HM test 20 20 No 15 10
THM test Phase 1 20 80 No No 9
Phase 2 20 80 5 No
ð33Þ Numerical forward simulations were performed for two tests by
means of Finite Element code. The code allows simulating the cou-
pled THM behaviour of sand-bentonite mixture during the tests.
Because the column sample is axis symmetric, the model was built
ð34Þ in the X–Y plane, but the problem was solved as an axis-symmetric
problem with a Y axis-symmetry. Details on the initial and bound-
ary conditions can be found in [24]. There are in total 28 parame-
ters in the systems, however the values of some parameters are
determined exactly from other experiments on the same specimen
(i.e. eo ¼ 0:88; /o ¼ 0:48 and pref ¼ 10 (kPa)). It experiences from
experiments and literature [28] that other parameters do not influ-
ence the analysis of the studied materials (i.e. ks ; q; pso in Eq.
(23)). Hence, finally 22 parameters are considered in the back anal-
ysis. More so, according to literature and experiments in our lab on
similar soils, the ranges of the parameter variations (search space)
have been defined and are presented in Tables 3–5.
3.3. Results and discussions
In this work, a minimisation of the objective function was
always successful and a significant reduction in its value by PSO
method was achieved, see Fig. 3. The objective function decreases
quickly from the first iteration to the 60th iteration. The objective
function decreases slowly from the 60th to the 80th iteration and
after the 80th iteration the objective function is nearly unchanged.
Therefore, the maximum iteration number selected here is 120
iterations for PSO optimisation, and the swarm of parameter vec-
tors includes 10 vectors. For the sampling process, we performed
12 iterations (np ¼ 12) for each forward problem (HM and THM).
Consequently, there are in total 24 samples of the identified

meas parameters occurring in two forward problems. The initial param-
ymeas ¼ ydtm ¼ fSl ðtmÞ; ryy ðtmÞ; TðtmÞ; sðtmÞgmeas : ð35Þ eter values were randomly selected according to a uniform
 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32 29

Table 3 g
f best converges to almost identical values. This might be generally
Parameters involved in modelling net stress driven processes.
because of the non uniqueness solution of the problem or
M jio (–) ai (–) kso (–) r (–) b (MPa1) pc (MPa) M (–) p0 (MPa) insensitivity of these parameters. In the scope of this paper, it is
Min 0.005 0.030 0.050 0.75 0.005 0.008 0.90 0.15 supposed that these parameters are insensitive parameters. If the
Max 0.020 0.000 0.150 0.99 0.500 0.070 1.40 0.50 model is not sensitive with these parameters, these parameters
cannot be identified by back analysis based on these experiments.
However, such a model parameter may be identified by back
analysis of other experiments.
Table 4 The proposed sensitivity analysis was used to assess the sensi-
Parameters involved in modelling suction driven processes.
tivity of the model response with respect to the model input
H asp (–) ass (–) js0 (–) P 0 (MPa) k (–) ko (m2) parameters. The results of the sensitivity analysis are presented
Min 0.1 0.008 0.001 0.30 0.30 1.0E17 in Fig. 4, which shows that each parameter influences some partic-
Max 0.5 0.000 0.025 1.70 0.65 1.0E15 ular processes.
According to the scheme in Fig. 1, any parameter is only
selected for the confidence interval analysis if it satisfies two crite-
Table 5 ria: (i) it has a sum of sensitivity indices greater than S0 ¼ 0:5, (ii)
Parameters involved in modelling temperature driven processes. the distribution of its samples obeys a normal distribution
T Dr s nD ksat kdry ao a1 a3 (W j > 0:9). Therefore, the selected parameters are
(m2 =s) (–) (–) (W/mK) (W/mK) (°C1) (MPa/°C) (MPa/°C2) P0 ; k; ko ; ksat ; js0 ; nD . Some parameters (e.g. asp ; a0 ) have the
Min 1.0E10 0.5 0.5 0.8 0.4 5.0E6 8.6E3 4.5E3
sensitivity index higher than 0.5, but they could not pass the nor-
Max 9.0E06 1 2.0 1.5 0.9 5.0E5 8.3E3 4.3E3 mality test. The reason for this might be that they are strongly
influenced by the inherent uncertainties in the measurements of
the HM and THM test, respectively. The two tests produce two
independent sets of samples. The probability distribution of these
100 two sets may not be the same, therefore when combining two sets
Sample 1
Objective function fbest

of samples in a histogram, it will not show a normal distribution or

g

Sample 2 not pass the normality test. It indicates solely that the normality
Sample 3
test is indispensable in order to avoid identifying the confidence
Sample 4
interval for parameters influenced by the uncertainty in the
measurements.
10−1 Fig. 5 presents the histograms and scaled Student probability
density distribution (PDF), in which model parameters are nor-
malised from 0 to 1 corresponding to their minimum and maxi-
mum boundaries, respectively. From Fig. 5, it is clear that these
parameters have their distribution close to the normal distribution.
0 20 40 60 80 100 120
The figure also indicates that the high sensitive parameters have
Iteration step
narrow ranges of their probability density functions (i.e.
g
Fig. 3. Objective function f best versus iteration step. P0 ; k; ko ). Less sensitive parameters have wider ranges of their
probability density distributions (i.e. ksat ; js0 ; nD ). For the non-
sensitive model parameters, the probability distributions show
distribution for each forward computation. The sampling process no pattern and significantly deviate from the normal distribution.
yields different sets of parameters. Some of the obtained parame- From this, we derive that the confidence intervals for these param-
ters vary significantly among the different samples, even though eters may not be reliably estimated. The confidence intervals are

Saturation
1.5
HM test Stress
1
Sj

0.5

2 Saturation
Stress THM test
Temperature
Sj

0
κi0 κs0 αss αi αsp α0 M pc p∗0 P0 λ ko Dr τ nD λdry λsat
Parameter
Fig. 4. Sensitivity indices Sj .
30 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32

10
P0 λ ko
8
W = 0.92 W = 0.91 W = 0.95
Frequency 6

10
nD λsat κs0
8
W = 0.96 W = 0.91 W = 0.95
6
Frequency

0
0 0.5 1 0 0.5 1 0 0.5 1
Normalised parameter value
Fig. 5. Histogram and PDF of the parameter samplings.

Table 6
Confidence intervals of the parameters. estimated considering typical confidence coefficients (cc ¼ 95%),
which means that the expected value of each of the identified
P0 k ko ksat js0 nD
model parameters is with a 0.95 probability within the confidence
Mean 1.51 0.339 2.03E16 1.33 0.0071 0.79 interval. Table 6 presents the mean value, the upper (CIU ) and the
CIL 1.45 0.328 1.55E16 1.28 0.0012 0.73
lower (CIL ) bound of the confidence interval calculated using Eq.
CIU 1.58 0.349 2.50E16 1.38 0.013 0.86
(10). The result in Fig. 5 shows that in general the parameters,

1
Measurement - load cell
Simulation
0.8 40
Vertical sress (kPa)
Saturation (-)

0.6
Measurement - TDR 1
Measurement - TDR 2 20
0.4 Measurement - TDR 3
Simulation - TDR 1
Simulation - TDR 2
0.2 Simulation - TDR 3
0
0 20 40 60 80 100 10−1 100 101 102

22
Measurement - RH 1 Measurement - TH 1
Measurement - RH 2 Measurement - TH 2
20 Measurement - RH 3 Measurement - TH 3
Temperature (◦C)
Suction (MPa)

Simulation - RH 1 21 Simulation
Simulation - RH 2
Simulation - RH 3

10
20

0 19
0 20 40 60 80 100 0 20 40 60 80 100
Time (h) Time (h)
Fig. 6. HM test: curve fitting with mean parameter values.
 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32 31

1 80
Measurement - TDR 1 Hydration → Measurement - load cell
Measurement - TDR 2 Simulation
0.8

Vertical sress (kPa)

Measurement - TDR 3
Hydration →
Saturation (-)
Simulation - TDR 1 60
0.6 Simulation - TDR 2
Simulation - TDR 3
0.4
40
0.2

0 20
0 500 1,000 1,500 2,000 2,500 101 102 103 104

200 70
Measurement - RH 1 Measurement - TH 1
Measurement - RH 2 Measurement - TH 2
Hydration →
Measurement - RH 3 Hydration → 60 Measurement - TH 3
150 Simulation - RH 1 Simulation - TH 1

Temperature (◦ C)
Simulation - RH 2 Simulation - TH 2
Suction (MPa)

Simulation - RH 3 Simulation - TH 3
50
100
40

50
30

0 20
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
Time (h) Time (h)
Fig. 7. THM test: curve fitting with mean parameter values.

which are highly sensitive (e.g. ko ; k; P 0 ), have a more pronounced
peakedness. In contrast, the low sensitivity parameters (e.g.
ksat ; js0 ; nD ) have less pronounced peakedness.
The mean value of the samples is now used to perform a numer-
ical simulation for the two tests. The results of the numerical sim-
ulation are presented in Figs. 6 and 7. In these figures, the
simulation are compared with the measurements provided by
[24]. Each experiment point represents a measurement. The num-
ber of measurement points is 272 in the HM test and 583 in the
THM test. From the plots, it can be seen that the numerical simu-
lation data has generally a good agreement with the experimental
data for both tests. The temperature measured in HM test has a
fluctuation in a range of 1 (°C) due to the accuracy of thermocouple
measurement devices and the uncertainty of the temperature
controller.
Based on the above presented analysis, it can be suggested that
the following issues should be considered when identifying soil
model parameters via back analysis:
 In the current approach, there are two methods to improve the
reliability of the identified parameters. The first method, a
nowadays popular method, is to increase the number of the lab-
oratory tests with suitable boundary conditions. The second
method, as discussed in this paper, is to increase the number
of the samples (np ) in the sampling process and then to assess
the results statistically. However, it is not guaranteed that the
latter method can improve significantly the solution in case of
uncertainties or errors in some of the measured data or in case
of strong abstractions of the model.
 The proposed back analysis approach may be considered expen-
sive because it requires a significant number of parameter sam-
ples (np ) to approximate the distributions of the parameter.
 For all statistical tests and derived conclusions, when the num-
ber of samples is not sufficiently large due to the expensive
inverse computations, any results are needed to be interpreted
with caution.

 Generally, the THM analysis contains a large set of input param-
eters due to consideration of many physical processes which
possibly occur in the unsaturated media. The model sensitivity
with respect to a given model parameter depends on the bound-
ary conditions and the physical process demanding the employ-
ment of this parameter in the mathematical description.
Therefore, the boundary value problems must be strongly cou-
pled with the targeted parameters for identification.
 The parameter sensitivity determines the confidence in the
parameters obtained by the optimisation method. However,
ranking the general sensitivity of a parameter should be carried
out with the awareness that each parameter influences differ-
ently the particular types of the responses in a multi-physical
problem.
4. Conclusions
In this paper, a novel back analysis approach is proposed com-
bining model parameter identification and the assessment of the
reliability of identified model parameters. The obtained results
show that it is a promising method for model parameter identifica-
tion, especially in coupled multi-physical problems involving com-
plex non-linear constitutive laws. The reliability of an identified
parameter depends on its sensitivity revealed under the particular
boundary conditions employed in the forward problem. It can be
confirmed in general that the higher the parameter sensitivity,
the more narrow is the parameter confidence interval. The
approach can be applied for back analysing a variety of geotechni-
cal problems. It is especially designed for parameter identifications
32 L. Nguyen-Tuan et al. / Computers and Geotechnics 76 (2016) 23–32

of problems having a large set of constitutive parameters and a [12] Wolpert D, Macready W. No free lunch theorems for optimization. IEEE Trans
Evol Comput 1997;1:67–82.
large set of model responses.
[13] Hofmann M, Most T, Hofstetter G. Parameter identification for partially
saturated soil models. In: 2nd international conference on computational
Acknowledgements methods in tunneling, Ruhr University Bochum; 2009.
[14] Schanz T, Zimmerer M, Datcheva M. Identification of coupled hydro-
mechanical parameters with application to engineered barrier systems. In:
This research was supported by the Deutsche Forschungsge- Toll et al., editors. Unsaturated soils: advances in geo-engineering – 1st
meinschaft (DFG) through the research grants SCHA 675/17-1 European conference on unsaturated soils, Durham, UK; 2008. p. 797–803.
and LA 2869/4-1. These financial supports are gratefully [15] Kennedy J, Eberhart R. Particle swarm optimization. In: Proceedings of IEEE
international conference on neural networks. IV; 1995. p. 1942–8.
acknowledged. [16] Montgomery DC, Runger GC. Applied statistics and probability for
engineers. John Wiley & Sons Inc.; 2002.
References [17] DIT-UPC. CODE_BRIGHT user’s guide. Universitat Politècnica de Catalunya,
Barcelona, Spain; 2009.
[18] Shapiro SS, Wilk MB. An analysis of variance test for normality (complete
[1] Neuman SP, Yakowitz S. A statistical approach to the inverse problem of
samples). Biometrika 1965;52(3/4):591–611.
aquifer hydrology: 1. Theory. Water Resour Res 1979;15(4):845–60.
[19] Shi Y, Eberhart R. Parameter selection in particle swarm optimization. In:
[2] Cividini A, Maier G, Nappi A. Parameter estimation of a static geotechnical
Proceedings of evolutionary programming VII (EP98); 1998. p. 591–600.
model using a Bayes’ approach. Int J Rock Mech Min Sci 1983;20(5):215–26.
[20] Trelea I. The particle swarm optimization algorithm: convergence analysis and
[3] Honjo Y, Wen-Tsung L, Guha S. Inverse analysis of an embankment on soft clay
parameter selection. In: Information processing letters, vol. 85; 2003. p. 317–
by extended Bayesian method. Int J Numer Anal Methods Geomech 1994;18
25.
(10):709–34.
[21] Zimmerer MM. VARO2PT. User Manual; 2010.
[4] Ledesma A, Gens A, Alonso EE. Estimation of parameters in geotechnical
[22] Royston P. As r94 – calculates Shapiro–Wilk normality test and p-value. Appl
backanalysis – I. Maximum likelihood approach. Comput Geotech 1996;18
Stat 1995;44(4):547–51.
(1):1–27.
[23] Papoulis A. Probability, random variables, and stochastic processes. 3rd
[5] Pichler B, Lackner R, Mang HA. Back analysis of model parameters in
ed. McGraw-Hill Series in Electrical Engineering; 1991. Chapter 9.
geotechnical engineering by means of soft computing. Int J Numer Meth Eng
[24] Schanz T, Nguyen-Tuan L, Datcheva M. A column experiment to study the
2003;57(14):1943–78.
thermo-hydro-mechanical behaviour of expansive soils. Rock Mech Rock Eng
[6] Schanz T, Datcheva M, Nguyen-Tuan L. Back analysis of a coupled thermo-
2013;46(6):1287–301.
hydro-mechanical model based on instrumented constant volume column
[25] Alonso E, Alcoverro F, et al. The FEBEX benchmark test: case definition and
test. In: Papadrakakis M, Onate E, Schrefler B, editors. IV int. conf. on comput.
comparison of modelling approaches. Int J Rock Mech Min Sci 2005;42(5–
methods for coupled problems in sci. and eng., 20–22 June, 2011, Kos Island,
6):611–38.
Greece; 2011. p. 76–89.
[26] Sánchez M, Gens A, Olivella S. THM analysis of a large-scale heating test
[7] Knabe T, Datcheva M, Lahmer T, Cotecchia F, Schanz T. Identification of
incorporating material fabric changes. Int J Numer Anal Meth Geomech
constitutive parameters of soil using an optimization strategy and statistical
2012;36(4):391–421.
analysis. Comput Geotech 2013;49(0):143–57.
[27] Olivella S, Carrera J, Gens A, Alonso EE. Numerical formulation for a simulator
[8] Tang Y-G, Kung GT-C. Application of nonlinear optimization technique to back
(CODE_BRIGHT) for the coupled analysis of saline media. Eng Comput
analyses of deep excavation. Comput Geotech 2009;36(1–2):276–90.
1996;13:87–112.
[9] Rechea C, Levasseur S, Finno R. Inverse analysis techniques for parameter
[28] Alonso E, Gens A, Josa A. A constitutive model for partially saturated soils.
identification in simulation of excavation support systems. Comput Geotech
Géotechnique 1990;40(3):405–30.
2008;35(3):331–45.
[29] Gens A. Constitutive laws. In: Gens A, Jouanna P, Schrefler B, editors. Modern
[10] Zhang Y, Gallipoli D, Augarde C. Parameter identification for elasto-plastic
issues in non-saturated soils. Springer-Verlag; 1995.
modelling of unsaturated soils from pressuremeter tests by parallel modified
[30] Hueckel T, Pellegrini R. Thermo-plastic modelling of untrained failure of
particle swarm optimization. Comput Geotech 2013;48(0):293–303.
saturated clay due to heating. Soils Found 1992;31:1–16.
[11] Ledesma A, Gens A, Alonso EE. Parameter and variance estimation in
[31] van Genuchten MT. A closed-form equation for predicting the hydraulic
geotechnical back analysis using prior information. Int J Numer Anal Meth
conductivity of unsaturated soils. Soil Sci Soc Am J 1980;44:892–8.
Geomech 1996;20:119–41.