Inverse analysis on two geotechnical works: a tunnel and a

cavern.

S. Eclaircy-Caudron, D. Dias & R. Kastner

LGCIE laboratory, Site Coulomb 3, INSA Lyon, 20 avenue Albert Einstein, 69100 Villeurbanne, France
T. Miranda & A. Gomes Correia
University of Minho, Department of Civil Engineering, Guimares, Portugal
L. Ribeiro e Sousa
University of Porto, Department of Civil Engineering, Porto, Portugal

ABSTRACT: One of the major difficulties for geotechnical engineers during project phase is to
estimate in a reliable way the mechanical parameters values of the adopted constitutive model.
In project phase, they can be evaluated by laboratory and in situ tests. But, these tests lead to
uncertainties due to the soil reworking and to local character of the test which is not representa-
tive of the soil mass. Moreover for in situ tests interpretation difficulties exist due to the non
homogeneous character of the strain and stress fields applied to the soil mass. In order to reduce
these uncertainties, geotechnical engineers can use inverse analysis processes during construc-
tion. This article shows the application of two of these processes (a deterministic and a probabil-
istic method) on convergence leveling measurements realized during the excavation of the Bois
de Peu tunnel (France). Moreover, these two processes are also applied on displacements meas-
ured by inclinometers during the excavation of the hydroelectric powerhouse cavern Venda
Nova II (Portugal). The two inverse analysis methods are coupled with two geotechnical soft-
ware (CESAR-LCPC and FLAC3D) to identify soil parameters. Numerical and experimental re-
sults are compared.

1 INTRODUCTION

The purpose of an inverse analysis process is the identification of parameters using tests results
or/and experimental measurement carried out during works. Various methods exist. Hicher
(2002) distinguishes three kinds: analytical methods, correlation and optimization methods. Op-
timization methods are applied when model parameters are not appropriate to be used in a direct
approach by graphic construction or in an analytical approach. By this method, the inverse prob-
lem is solved using an algorithm which minimizes a function depending on all parameters. This
function is generally called cost function and corresponds to the difference between numerical
and experimental measurements. The experimental results can come from various origins: labo-
ratory tests, in situ tests or work measurement data. Two types of optimization methods are dis-
tinguished: deterministic and probabilistic methods. The deterministic methods include gradient,
Newton, Gauss-Newton or Liebenberg Marquardt optimization algorithms. Several researches
aiming at parameter identifications based on deterministic methods were carried out in the last
years. Zentar (2001) tried to identify some mechanical parameters of the Saint-Herblain clay us-
ing the results of pressuremeter tests and the optimization software SiDoLo (Pilvin 1983). He
considered an elastoplastic model with or without hardening. In spite of the identification meth-
ods development, a few of them were applied to real cases like tunnels or deep excavations
(Jeon et al. 2004, Finno et al. 2005).
The deterministic methods present some advantages. The iteration number required to
achieve the optimization process is relatively low. But, if the cost function presents several local
minima, the deterministic methods can converge towards the first found minimum. This major
drawback explains why these methods are seldom used for complex problems. For such prob-
lems probabilistic methods are preferred. Probabilistic methods include evolutionary algorithms
such as genetic algorithms and evolution strategies. Evolutionary algorithms reproduce the natu-
ral evolution of the species in biological systems and they can be used as a robust global optimi-
zation tool. The major principles of genetic algorithms were developed by Goldberg (1991) and
renders (1995). Evolution strategies (Schwefel, 1985) start searching from an initial population
(a set of points) and transition rules between generations are deterministic. The search of new
points is based on mutation and recombination operators. Recent researches applying genetic
algorithms to soil parameters identification for a constitutive model have been undertaken in
geotechnic (Levasseur et al. 2005, Samarajiva et al. 2005). Evolution strategies were also re-
cently applied to problems in many domains (Costa et al. 2001).
First, this paper presents briefly the used optimization software, SiDoLo and the used evolu-
tionary algorithm. Then the numerical modeling of the two geotechnical works and the followed
approach are detailed. Finally, results obtained by the two processes are compared and influence
of some data in the evolutionary algorithm is highlighted.

2 PRESENTATION OF THE OPTIMIZATION METHOD USED

2.1 The optimization software, SiDoLo
SiDoLo is coupled with the geotechnical software CESAR-LCPC (Itech 2002) or FLAC3D
(Itasca 2005). The coupling principle is illustrated in figure 1. Numerical results
DATA FILES

CESAR-LCPC
Set of initial
parameters

INTERFACE

SiDoLo
Experimental
measurements

OPTIMUM SET OF
PARAMETERS

Figure 1. Principle of the coupling

Through this coupling, SiDoLo compares numerical results with experimental measurements,
in order to calculate the cost function. When the cost function is lower than a fixed value, then
the optimization process stops. The cost function L(A) is expressed by the following finite sum:
N 1 Mn
L( A) = T [Z s ( A, t i ) Z s* (t i )].Dn .[Z s ( A, t i ) Z s* (t i )]
n =1 Mn i =1
(1)
where A are the model parameters, N is the number of experimental measurements; [Zs(A,ti)-
Zs*(ti)] is the difference between numerical and experimental results evaluated only at Mn ob-
servation steps ti and Dn is the weighting matrix of the nth test. Measurements accuracy can be
taken into account by weighting coefficients.
SiDoLo uses a hybrid optimization algorithm, which combines two typical minimization
methods: the gradient method and a variant of the Lavenberg-Marquardt method to accelerate
the convergence when the solution is close. More details on the SiDoLo approach can be found
in Eclaircy-Caudron et al. (2006).
2.2 The algorithm based on evolution strategies
The algorithm used was developed by Costa et al. (2001). This algorithm was adapted to the
problems studied in this paper, and as for SiDoLo, requires creating an interface with computer
codes. The population entities are vectors of real coded decision variables that are potential op-
timal solutions. An initial population is generated and then each following generation, off-
spring are generated from progenitors by mutation and recombination. Then the best entities
are selected for next generation among the + members according to their cost function value.
Finally, the best of all the + members become the parents of the next generations. Important
features of evolution strategies are the self adaptation of step sizes for mutation during the
search and the recombination of entities that is performed between individuals. This algo-
rithm, named generally (/ +) algorithm, is illustrated in figure 2.

Figure 2. Principle of the (/ +) evolution strategies (Costa et al., 2001)

The algorithm stops when one of the following conditions is verified:

The maximum number of generations is reached
The difference between the two extreme values of the cost functions is lower than 10-5
This difference divided by the average of the cost functions values is lower than 10-5
Various error functions are considered. First, the same function than in SiDoLo is introduced
in the algorithm. Then, the following error function is considered.
1 N [ Z s ( A, ti ) Z s* (ti )]2 (2)
LE ( A) =
N i = 1 [ + .Z s* (ti )]2

The and variables represent respectively the absolute and relative error of measurements.
Different values of these two variables are tested.

3 THE FOLLOWED APPROACH

3.1 Case 1: Application to the Bois de Peu tunnel
3.1.1 Presentation of the tunnel
The Bois de Peu tunnel is situated near Besanon in France. It is composed of two tubes. The
excavation length is about 520 meters per tube for a cover height which varies between 8 meters
and 140 meters. Four kinds of supports were foreseen in the project phase for four materials
types: limestones, marls, interbeding of marls and limestones and clays. The studied section is
located near one of the two tunnel portals where interbeding of marls and limestones were ex-
pected. A face levelling of this section is showed in figure 3. The lining support set up in this
area is composed of a mixed tunnel support by shotcrete with a thickness of 0.2 m and steel ribs.
Figure 3. Face leveling

The tunnel is dug with a constant step of 1.5 m. The cover height is equal to twenty six me-
ters. From the site investigations, two kinds of design characteristics were defined: probable and
exceptional. They are resumed in table 1.

Table 1. Design characteristics

_____________________________________________________________
Parameters E C
kN/m3 MPa MPa -
_____________________________________________________________
Probable 24.8 1600 0.70 40 0.3
Exceptional 24.0 750 0.21 36 0.3
_____________________________________________________________

3.1.2 Numerical model

Due to time consuming of three dimensional calculations, tunnel excavations are often taken
into account by two dimensional modelling. We adopted a transverse plane strain model, taking
into account the real geometry, the excavation and tunnel support being modelled by an uncon-
finement ratio (Panet, 1995). Only a half section is represented due to symmetry. The grid ex-
tents are 75 m (=6D) in all directions in order to avoid the influence of the boundaries. The grid
(figure 4) contains around 3900 nodes and 1850 elements with quadratic interpolation (six
nodes triangles and 8 nodes quadrangles). The initial stress field is anisotropic. An earth pres-
sure ratio of 0.7 is adopted. This value was obtained from in situ tests. Steel ribs and sprayed
concrete are simulated by a homogeneous support with equivalent characteristics. This support
is assumed to have a linear elastic behaviour. A linear elastic perfectly plastic model with a
Mohr Coulomb failure criterion and non associate flow rule is considered to represent the soil
mass.
Three computing phases are defined. First the initial stress field is applied. After, the tunnel is
excavated. In this phase, the unconfinement ratio at the time of the support installation is ap-
plied. Finally the homogeneous support with 0.24 m width is activated. A total unconfinement is
applied.
 H=26 m

Pt 1

6D Pt 2

Pt 3

6D

Figure 4. The grid and its dimensions

3.1.3 Experimental results

Figures 4 and 5 locate the three experimental measurements points used for optimization. Five
measurements are observed for these three points: levelling for these three points and horizontal
displacements for the points 2 and 3. They correspond to convergence and levelling measure-
ments of the tunnel wall in a given cross section (point 1 to 3 in figures 4 and 5).

Figure 5. The observations points and deformed section

The convergence and levelling measurements of this section are reported in figures 6 and 7.
These measurements permit to deduce the horizontal and vertical displacements of each point.
The convergence is maximal for wires 4 and 5. It reaches 30 mm for wire 4. Moreover leveling
is more important for targets 3, 4 and 5 than for the others. It reaches 13 mm. So, these meas-
urements showed a dissymmetrical deformation of the section. In optimization process average
displacements measured at equilibrium are used. The dissymmetry is confirmed by the section
deformation presented in figure 5 where the crown displacement is assumed purely vertical. The
experimental results considered in the optimization process correspond to the average displace-
ments observed at the two sides.
 5

0
0 50 100 150 200 250 300 350 400 450 500
-5
Convergence (mm)

-10

-15

-20

-25

-30 Wire 1
Wire 2
Wire 3
-35 Wire 4
Wire 5
Wire 6 Distance from the face (m)
Figure 6. Convergence measurements

5 Levelling points

0
0 50 100 150 200 250 300 350 400 450 500
Levelling (m m

-5

-10 1 2
3 4
5

-15
Distance from the face (m )
Figure 7. Levelling measurements

3.1.4 Unknown parameters

In plane strain calculation, the introduced unconfinement ratio at the time of support installa-
tion might be calculated by the convergence-confinement method (Panet, 1995) if the constitu-
tive model parameters were well known. But in this study, it is considered as an additional pa-
rameter to identify. Before using inverse analysis, a sensibility study was performed. The
Poisson ratio and the dilatancy angle are parameters which have low influence on the observed
variables. So, these parameters could not be identified with these measurements. Only parame-
ters , C, E and , might be evaluated.
Validation studies (Eclaircy-Caudron et al. 2006) were performed with SiDoLo on this nu-
merical model in order to define which parameters might be identified according to the available
experimental measurements. These studies showed that with convergence and levelling meas-
urements no more than two parameters could be identified. But SiDoLo always succeeded to
identify and even if they were unknown because they have more influence on the conver-
gence and levelling measurements than the others. In order to verify the provided solution an
evolution strategies algorithm is used in complement of the software SiDoLo.
3.2 Case 2: Application to the Venda Nova II hydroelectric powerhouse cavern
3.2.1 Presentation of the powerhouse cavern
The Venda Nova II powerhouse cavern is located in the North of Portugal, about 55 km of the
town Braga. The powerhouse complex includes two caverns interconnected by two galleries.
The caverns have respectively the dimensions as follows: 19.00 m 60.50 m and 14.10 m
39.80 m. Their axes are spaced by 45.00 m. They are situated at a 350 m depth. Both caverns
have vertical walls and arch roofs. The roof of both caverns is situated at different levels. More-
over, the main cavern has various floors while the other has a single floor. The geometry is il-
lustrated in figure 8. The powerhouse complex is situated in a rock mass characterized by me-
dium-size grain granite of a porphyritic trend with quartz and/or pegmatitic veins and beds,
which are, sometimes, rose. The rock mass also includes embedment of fairly quartzitic mica-
schist. In order to investigate the geotechnical characteristics of the rock mass in depth, four
deep and subvertical boreholes were performed. Laboratory tests were done on rock samples re-
sulting from boreholes. These tests permitted to identify three geological and geotechnical
zones. The zone where the caverns are located presents a Young modulus of 54 GPa and a Pois-
son ratio of 0.17. Moreover, in situ tests are carried out before the beginning of the main work
and after completing the access tunnel to the powerhouse. In fact, an exploration gallery was
excavated from the top of the access tunnel and parallel to the caverns axis. In situ tests confirm
the presence of four main families of discontinuities and led to values ranging from 33 to 40
GPA for the deformability modulus. The in situ stress state of the rock mass was also character-
ized. The vertical and horizontal stresses aligned with the longitudinal axes of the caverns corre-
spond to the earth weight. The horizontal stress normal to the longitudinal axes of the caverns is
2 to 3 times higher than the other stresses.

Figure 8. The powerhouse cavern geometry

3.2.2 Numerical model

The caverns excavation was modelled in 3D with the finite difference computer code
FLAC3D in order to simulate the real complex geometry of them and the excavation stages.
Figure 9 shows the mesh performed with the hexahedral-Meshing Pre-processor, 3DShop.
The grid dimensions are 300 m in the transversal direction, 460 m in the longitudinal direc-
tion and its height is 275 m. A layer of 75 m is modelled under the caverns. A cover height of
350 m is simulated. The mesh contains around 47000 grid-points, 44000 zones and 1100
structural elements. The simulated excavation stages are lightly different from the real stages
but the rock mass staying in elasticity, final displacements are the same. Figure 10 shows the
excavation sequences modelled.
 300 m 460 m

200 m

75 m

Figure 9. The 3D mesh.

Stage 1 Stage 2 Stage 3

Stage 4 Stage 5 Stage 6

Figure 10. The excavation stages.

The excavation steps are the following:

1: Total excavation of the main cavern arch. Set up of fibre sprayed concrete (thickness: 25
cm) on the arch.
2: Excavation of the main cavern arch until 1.5 m below the base level of the support beams
of the rail tracks and set up of fibre sprayed concrete on the walls cavern (thickness: 25 cm).
3: Excavation of the main cavern until the level of the interconnecting galleries.
4: Excavation of the interconnecting galleries and set up of fibre sprayed concrete on the gal-
leries arch (thickness: 25 cm).
5: Achievement of the main cavern excavation
6: Excavation of the transformer cavern and set up of fibre sprayed concrete (thickness: 25
cm) on the arch
In the numerical modelling, an elastic perfectly plastic constitutive model with a Mohr Cou-
lomb failure criterion is assumed to represent the rock behaviour. The sprayed concrete is simu-
lated by shell elements with a linear elastic and isotropic constitutive model, with a Young
modulus of 15 GPa and a Poisson ratio of 0.2. The rock bolts are simulated by cable elements,
with two nodes and one axial free degree. These elements can yield a tensile stress.
3.2.3 Experimental results
A monitoring program was established in order to evaluate the behaviour of the rock mass and
the support system and to observe displacements in the rock during and after construction. Con-
vergence targets were installed in several sections. Besides the convergence measurements, 11
extensometers were installed in two sections along the caverns axis (figure 11). In the power-
house cavern extensometers were installed in two sections, while only one in the other cavern
was installed. Almost all the extensometers are double. Just the ones installed in the wall on the
main caver are triple and of larger length (EF1 and EF5 in Figure 11). Figure 12 shows the evo-
lution of displacements in extensometers 1 and 5. Figure 13 gives the displacement value meas-
ured at the last stage for each extensometer.

Figure 11. Location of the extensometers in the two cross sections

EF1.5
EF1.15
-0.5 EF1.40
EF5.5
-1.5 EF5.15
EF5.40
Displacements (mm)

-2.5

-3.5

-4.5

-5.5

-6.5

-7.5
3/juil/01

11/oct/01

7/aot/02
29/avr/02

15/nov/02

3/juin/03

11/sept/03

20/dc/03
19/janv/02

23/fvr/03

29/mars/04

Date
Figure 12. Evolution of the measured displacement in extensometers 1 and 5

In optimization process, displacements measured at the last excavation stage are used. In
validation study, all extensometers are considered. So, 24 displacements are available in the op-
timization process. In the case of the application on the in situ measurements, only 20 displace-
ments are used in the optimization process because some values are not considered due to
measurement errors (EF 3.8, 3.3, 4.8 and 4.3).
 16
EF x.xx
14 x - n extensometer
xx - Depth (m)
12
Displacements (mm)

10

1.51.151.40 2.8 2.3 3.8 3.3 4.8 4.3 5.55.15 5.4 6.8 6.3 7.8 7.3 8.8 8.3 9.6 9.210.610.211.6 .18
EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF EF F 11
E
Figure 13. Displacements value measured in each extensometer at the last excavation stage.

3.2.4 Unknown parameters

Previously studies permitted to obtain mechanical properties of the rock by using Artificial In-
telligence techniques (Miranda et al. 2005). These are the following: 45 GPa for E, 54 for
and 4 MPa for C. Due to the good strength of the rock, a few plastic zones appear. So the value
of the two parameters C and has low influence on the numerical results. A value of 0.2 is
adopted for the Poisson ratio and the dilatancy angle is taken equal to 0. These two parameters
have also low influence on the numerical result. Then, only the Young Modulus might be identi-
fied by inverse analysis. Besides, many uncertainties reside on the in situ stresses ratio value R.
So, as this parameter influences the numerical results, it might be also identified by inverses
analysis.
Validation studies were performed with the two optimization processes on this numerical
model. These studies show that it is possible to identify the Young modulus and the in situ
stresses ratio just from the displacements measured by extensometers. So, this paper presents
the identification results performed on these two parameters and compares the results provided
between the two processes.

4 RESULTS
4.1 Case 1:Bois de Peu tunnel (France)
4.1.1 Comparisons between the two methods
Several identification attempts were performed with SiDoLo, confirming that SiDoLo provides
approximately the same values of and in all cases. They also showed that the friction angle
was lower than the project value (between 10 and 20).
One example of identification is presented here, where three vertical and two horizontal dis-
placements are used in the optimization process. For each identification several initial values
(referred as a, b and c) of the unknown parameters are tested in order to show the influence of
the initial value. Only the Young modulus E and the cohesion C are identified. and are fixed
respectively to the average of values provided by identifications of the four parameters (E, C, ,
). A friction angle of 14 and an unconfinement ratio of 0.7 are adopted. The enabled ranges in
SiDoLo are resumed in table 2. Table 2 gives also the obtained values. Several solutions are
found by the optimization process according to the introduced initial values. SiDoLo does not
succeed to find the best couple. It provides local minima. Displacements computed after optimi-
zation are showed in figure 14.
Table 2. The identification results realized with SiDoLo
___________________________________________________
Set E C
MPa MPa
___________________________________________________
Range 100-2000 0.10-1
a 217 0.13
b 100 0.19
c___________________________________________________
102 0.19

Points
Pt1-Vertical Pt2-Horizontal Pt2-Vertical Pt3-Horizontal Pt3-Vertical
0

-2

-4
Displacements (mm)

-6

-8

-10

-12

-14
In situ measurements
Probables characteristics
-16 Exceptionals characteristics
Optimization a
-18 Optimization b
Optimization c
-20 Optimization ES
Figure 14. Comparisons between the measured and computed displacements after optimization by Si-
DoLo

For points 2 and 3 numerical results are close to the measured displacements. The vertical
displacements computed at point 1 are lower than the one measured. The vertical displacement
measured at the tunnel crown appears difficult to reproduce in the numerical model. Although
the face leveling showed in figure 3 highlights a disturbed geology with folds and faults, an ho-
mogeneous medium is considered in the model. A more complex model seems to be required to
simulate the real behavior of the tunnel crown.
In order to avoid local minima, the evolution strategies algorithm is used. The same identifi-
cation is realized. This identification allows verifying the solutions found by SiDoLo. The same
ranges are enabled for parameters. The number of generations is limited to 50 in order to keep
acceptable computation times. The parent and recombination population size are 10 and the off-
spring population size is 20. The evolution strategies algorithm stopped when the maximum
number of generations is reached, the cost function being still relatively important and the stop
criteria on this function cannot be reached. The design characteristics, the best member provided
by the Evolution Strategies algorithm (ES) and solutions provided by SiDoLo are resumed in
table 3.

Table 3. Design characteristics and optimized values.

_____________________________________________________________
Parameters E C
MPa MPa -
_____________________________________________________________
Probable 1600 0.7 40 -
Exceptional 750 0.21 36 -
SiDoLo 100-217 0.13-0.19 14* 0.7*
ES 242.7 0.12 14* 0.7*
_____________________________________________________________
* Fixed values
 Figure 15 presents the evolution of the cost function. The cost function is not convex and
shows a valley. A decrease of parameter E is counterbalanced by an increase of C in a small
variation domain. So it is difficult to identify these two parameters with only convergence and
leveling measurements. Moreover the value of the cost function is important, confirming that
the numerical model should be improved. Figure 15 shows also the parents at initial and final
generations. At initial generation parents are dispersed in all research space. At later generation
they are located in the valley and are close to each other.

Error (102)
4
0.8
0.6
2
0.5 1

Initial generation

Final generation

Figure 15. Evolution of the cost function and representation of initial and final generations parents

Figure 16 locates the three couples found by SiDoLo and the best member found by the evo-
lution strategies algorithm after identification. The error function value is also reported. Solu-
tions found by SiDoLo are situated in the valley and correspond to local minima. The evolution
strategies algorithm provides a global minimum.

Sc (46 )
Sb (46 )

Sa (37 )

ES (36 )

Figure 16. Location of the solutions found by SiDoLo (Sa, Sb and Sc) and the best member found by the
evolution strategies algorithm (ES).

4.1.2 Influence of the error function

The two error functions presented in equations 1 and 2 are compared. For the second error func-
tion, different values of and are tested. All the error functions tested present a valley but the
valley extend depends on the error function and on the and values. Figure 17 showed the
evolution following E of the different cost functions for a fixed value of C close to the global
minimum. The error value reported in figure 17 is not the real value. It corresponds to the real
value minus the minimum value in order to set all curves at the origin. The minimum error is
reported on each curve. In this figure the different valleys are visible. For an absolute value of 1
mm and a relative value of 0 the valley is more condensed than for the others. So it is easier to
find the global minimum with this error function. The algorithm convergence is similar for all
the error function.

2.5
Mini. error
2.25
=0 mm =1
C=120 kPa; e=0 a=1 0.22
2 =0 mm =0.5
C=120 kPa; e=0 a=0.5 0.46

1.75 C=120 kPa; =1 mm =0

e=1 mm a=0 2.69

1.5
Corrected error

1.25

0.75

0.5

0.25

0
150 175 200 225 250 275 300 325 350 375 400

E ( MPa)
Figure 17. Evolution of the LE(A) error function for several couples of parameters (, ) following the
Young modulus for a fixed cohesion value and their minimum value

The solution found by the evolutionary algorithm varies according to the values of absolute
and relative errors. Table 4 resumes the found solutions and the computation time. The compu-
tation time with SiDoLo is also given. The same solution is found if just a relative error or an
absolute error is considered but this solution changes according to the considered type of error.
The calculation time is similar for all the error function. But it is very important compared to the
computation time required by SiDoLo.

Table 4. Found solutions according to the error function and calculation time
_____________________________________________________________
Error E C Error Time
MPa MPa Hours
_____________________________________________________________
=0 ; =1 332 0.118 0.22 56
=0 ; =0.5 332 0.118 0.46 55
=1 mm ; =0 289 0.12 2.69 53
SiDoLo 100-217 0.13-0.19 37-46 1,5
_____________________________________________________________

4.1.3 Influence of the population size

Table 5 gives the solutions provided by the algorithm and the calculation time in two cases. In
the first case a parent and recombination population size equal to 10 and a offspring population
size of 20 is adopted. In the second case, values of 15 and 30 are respectively adopted for these
population sizes.

Table 5. Found solutions according to the population size and calculation time
_____________________________________________________________
Cases E C Error Time
MPa MPa Hours
_____________________________________________________________
Case 1 288 0.120 2,69 53
Case 2 252 0.125 2.69 84
Difference 12 % 4% 0% 50 %
_____________________________________________________________
 For each calculation, the stop criteria (10-5) cannot be reached; then it is the fixed maximum
number of generations which stopped the calculation. For the second case the time is greater
about 50 %. With a difference of 12 % on the Young modulus and 4 % on the cohesion, these
two calculations give the same error function value. So, due to the computation, it does not
seem necessary to increase the population size in our studied case.

4.2 Case 2: hydroelectric powerhouse cavern (Portugal)

4.2.1 Identifications with SiDoLo
The initial values introduced in SiDoLo are the following: 45 GPa for E and 2 for the in situ
stress ratio R. First, the twentieth values of displacements are considered in the optimization
process. Figure 18 compares the measured displacement to the computed displacements with the
initial values and with the optimized values of the two parameters. It also gives the respective
error function values. Displacements computed with the optimized values are closer to the ex-
perimental measurements than those computed with initial values excluded for 1.5, 1.15, 5.5,
5.15, 9.2, 10 and 11 (so for 9 observations points). Despite of that, the error function computed
with the optimised value is lower than with initial values. So, the using of the optimization
software SiDoLo permits to improve the fitting of the numerical values on experimental values.
Extensometer 1 and 5 are the same location in the two main cavern cross sections. Extensom-
eters 9, 10 and 11 are located in the small cavern. It could be interesting to do identification only
on the main cavern monitoring and in two steps: first, on the monitoring of the first cross sec-
tion (EF 1 to 4) and after on the other section (EF 5 to 8).
The difference [Zs(A,ti)- Zs*(ti)] between the computed value Zs(A,ti) and the experimental
value Zs*(ti) is always of the same sign excluded for values EF 1.15 and EF 5.4. In fact, for these
observations points, the computed value before optimization is more important than the meas-
ured values while after optimization, this value is lower than the measurement. In a second time,
these two displacements will not be considered in the optimization process. So, only 18 dis-
placements will be used for the identification.
Figure 19 shows the error function value for each displacement and in the two cases: before
and after optimization. With initial values, maximal error functions are obtained for EF 1.4 and
2.3 while they are obtained for EF 1.5 and 11.6 with optimized values.

10 Measurements -6
Computed response before optimization, error=1.7610
-6
9 Computed response after optimization (SiDoLo), error=1.2510

7
Displacement (mm)

0
E F 1.5

EF .4
EF .15

EF 0
EF .8
EF .3
EF 5.5

EF 5

EF .8
EF .3
EF .8
EF .3
EF .8
EF .3
EF .6
EF 9.2
E F 0 .6
E F 0 .2

6
8
4

EF 11.
.1
2

5
6
6

8
9
2

8
1.

5.
1

1
1

11
EF

Figure 18. Comparisons between experimental and numerical measurements after the first identification
attempt with SiDoLo.
 12 Before optimization
After optimization
10
Maximum values before optimization
Error function (10-6)

6
Maximum values after optimization
4

0
EF 5

EF 0.2
EF .15

EF 0

E F 0.6

11 6
EF .8

E F 5.5

EF .8

EF .8

EF .8

EF .6
EF 1.5

EF .3

EF .4

EF .3

EF .3

EF .3

EF 9.2

8
1
4

EF 1 1 .
.1
5

8
2
2

9
5.
1.

1
1

1
EF

Figure 19. The error function values during the first identification attempt with SiDoLo

In a second time, identifications are performed only on 18 displacements. Table 6 resumes

the parameters values provided by SiDoLo at the end of the two identification attempts, the re-
spective error function and the iteration number required to achieve the identification process.
For the second attempt the iteration number is not so important than for the first attempt. But,
the error function is more important than for the first identification attempt. So, the fitting of the
numerical values on experimental values is not improved.

Table 6. Results of the two identification attempts with SiDoLo

________________________________________________________________________________________
Cases E R Error function iteration Time
-6
GPa - 10 Hour
________________________________________________________________________________________
Initial values 45.0 2 1.76 or 1.7 - -
Attempt 1 55.0 1.98 1.25 25 27
Attempt 2 56.7 1.90 1.34 12 14
Difference 3 % -4.2 % 7.2 % -108 % -93 %
________________________________________________________________________________________

Figure 20 compares the measured displacement to the computed displacements with the ini-
tial values and with the optimized values of the two parameters. It also gives the respective error
function values. Similar observations to those obtained for the first identification attempt can be
made. The error function computed with the optimised value is lower than with initial values.
The difference [Zs(A,ti)- Zs*(ti)] between the computed value Zs(A,ti) and the experimental value
Zs*(ti) is always of the same sign.
A similar error function value for each displacement to the one computed in the first identifi-
cation attempt is obtained.
Figure 21 shows the plastic zones computed at the last excavation stage with the optimized
values obtained for the second identification attempt (E=56.7 GPa and R=1.9). Plastic zones are
very limited with the two optimized parameters sets. So, it could be very difficult to identify C
and . These parameters have no influence on the results. Maybe, it might be interesting to
adopt an elastic linear constitutive model to represent the rock behaviour and to compare the
computed displacements. Besides, the computation time will be shorter.
 10 Measurements -6
Computed response before optimization, error=1.7 10
-6
9 Computed response after optimization (SiDoLo), error=1.3410

7
Displacement (mm)

0
EF .5

EF 8

EF 3
EF .5

EF 8

EF 3
8

EF 3

EF 8

EF 3

EF 6
EF .2

8
40

15

EF .6

EF .2
EF 1.6
.1
2.

2.

6.

6.

7.

7.

8.

8.

9.
1

9
10

10
1.

5.

1
11
EF

EF

EF

EF

Figure 20. Comparisons between experimental and numerical measurements after the second identifica-
tion attempt with SiDoLo.
Excavated powerhouse cavern
Plastic zones

Figure 21. Plastic zones at the last excavation stage in the case of the second identification attempt.

4.2.2 Identifications with the evolution strategies algorithm and comparisons between the two
optimizations methods
In order to verify that SiDoLo does not provide a local minimum, a probabilistic optimization
method is performed in a second step. An evolution strategies algorithm is used to find the op-
timized values of the two parameters and obtained results are compared between the two meth-
ods. Only 18 displacements are used in the optimization process.
The computed displacements with the optimized values provided by the two methods are
close to each other. Table 7 compares the parameters value provided by SiDoLo at the end of
the second identification attempt to the values provided by the evolution strategies algorithm
(ES), the respective error functions and the computation time required to achieve the identifica-
tion process.
Table 7. Comparisons between the two optimization methods
________________________________________________________________________
Cases E R Error function Time
-6
GPa - 10 Hour
________________________________________________________________________
SiDoLo 56.7 1.90 1.34 14
ES 52.1 1.72 1.37 49
Difference -8.1 % -9.4 % 2.2 % 250 %
________________________________________________________________________

The error functions are quite similar. A difference of 8.1 % is observed on the Young
modulus value provided by the two methods and for the in stresses ratio a difference of 9.4 % is
obtained. The time required by the evolution strategies algorithm is more than 3 times higher in
comparison with the time required for the other method. Only one generation is required in the
evolution strategies algorithm to achieve the optimization process due to that for the parents of
the first generation the stop criterion is verified. In order to verify that the solution provides by
the evolution strategies algorithm corresponds to the global minimum it is necessary to decrease
the tolerance criteria in order to obtain error function lower than 10-7 and to launch a new identi-
fication. But, the computation time will be important.

5 CONCLUSION

The use of inverse analysis on in situ measurements carried out during the construction of a real
work is rare. This might be due to the difficulty met to apply inverse analysis (measurements
quality, complexity of the geometry and of the excavation stages, ). In an underground works
case, the numerical model required many simplifications in order to obtain acceptable computa-
tion times. The adopted constitutive model should not be too sophisticated with many parame-
ters but should properly simulate the soil behaviour. For the studied cases, a linear elastic per-
fectly plastic model with a Mohr Coulomb failure criterion is considered to represent the soil
mass.
This article shows the coupling of two optimization methods with 2D and 3D numerical mod-
elling of two geotechnical works: a tunnel and underground powerhouse caverns.The use of a
deterministic method may provide local minima. The probabilistic method based on evolution
strategies permits to find global minimum. But this method required more important computa-
tion time. In the tunnel case, with two unknown parameters, the studied cost functions are not
convex and present a valley. Moreover the error values are important for all the studied cost
function (> 20 %). Thus the problem is difficult to solve. The solution found by the evolutionary
algorithm varies according to the studied error function. For the second used function LE(A) it
depends on the values of absolute and relative errors. But if just a relative error or an absolute
error is considered the same solution is found. The solution changes according to the considered
type of error. The calculation time is similar for all the error functions. This article shows also
that the parent, combination and offspring population size influence the results. With an in-
crease of the population size, the same error function value is obtained at the same number of
generations. The provided value sets are different of about 10 % and the calculation time can in-
crease of 50 %. In order to improve results a non linear elasticity and a Young modulus which
varies with depth should be introduced. Moreover, maybe the evaluation of the earth pressure
ratio by means of in situ tests needs to be more investigated. This ratio might be considered as a
complementary parameter to identify.
In the complex powerhouse caverns case, a linear elastic constitutive model could be consid-
ered to represent the rock behaviour because a few plastic zones appear. And, this model should
permit to reduce computation time. Moreover, in order to verify that the solution provides by
the evolution strategies algorithm corresponds to the global minimum in the case of the complex
powerhouse caverns, it is necessary to decrease the tolerance criteria and launch some new iden-
tifications.
AKNOWLEDGMENTS

The authors wish to express their acknowledge to Professor Lino Costa for providing the evolu-
tion strategy algorithm.

REFERENCES

Costa, L. & Oliveira, P. 2001. Evolutionary algorithms approach to the solution of mixed integer non lin-
ear programming problems. Computers and Chemical Engineering 25: 257-266.
Eclaircy-Caudron, S., Dias, D., Kastner, R. & Chantron, L. 2006. Identification des paramtres du sol ren-
contr lors du creusement dun tunnel par analyse inverse. Proc. JNGG 2006. Lyon, France.
Finno, R.J. & Calvello, M. 2005. Supported Excavations: the Observational Method and Inverse Model-
ing. Journal of Geotechnical and Geoenvironmental Engineering 131(7): 826-836.
Goldberg, D. 1991. Algorithmes gntiques : exploration, optimisation et apprentissage automatique.
Adisson-Wesley Edition.
Hicher, P.Y. & Shao, J.F. 2002. Modles de comportement des sols et des roches 2: lois incrmentales,
viscoplasticit, endommagement. Herms Science Publications.
Itasca Consulting group 2005. FLAC3D users manual.
Itech 2002. Cleo2D users manual.
Jeon, Y.S. & Yang, H.S. 2004. Development of a back analysis algorithm using FLAC. International
Journal of Rock Mechanics and Mining Sciences 41(1): 447-453.
Levasseur, S., Malecot, Y., Boulon, M. & Flavigny, E. 2005. Analyse inverse par algorithme gntique en
gotechnique : application un problme dexcavation. 17th Congrs Franais de Mcanique. Troyes,
France.
Miranda, T., Gomes Correira, A., Ribeiro e Sousa, L. & Lima, C. 2005. Numerical modelling of a hy-
droelectric underground station using geomechanical parameters obtained by artificial intelligence
techniques. 4th Portuguese-Mozambican of Engineering, Maputo,30 August-1 September 2005:807-
816.
Panet, M. 1995. Le calcul des tunnels par la mthode convergence-confinement. Presse de lcole natio-
nale des Ponts et Chausses.
Pilvin, P. 1983. Modlisation du comportement dassemblages des structures barres. Ph. D. thesis. Uni-
versit Paris VI. Paris, France.
Renders, J.M. 1995. Algorithmes gntiques et rseaux de neurones. Herms Science Publications.
Samarajiva, P. 2005. Genetic algorithms for the calibration of constitutive models for soils. International
Journal of Geomechanics. ASCE. September 2005: 206-217.
Schwefel, H.-P. 1985. Evolution and optimal seeking. John Wiley and Sons.
SiDoLo version2.4495. 2003. Notice dutilisation. Laboratoire Gnie Mcanique et Matriaux de
lUniversit de Bretagne-Sud, Lorient.
Zentar, R. 1999. Analyse inverse des essais pressiomtriques : Application largile de Saint-Herblain.
Ph. D. thesis. Ecole Centrale. Nantes, France.