You are on page 1of 8

Optimisation procedure for choosing Cam clay parameters

V. Navarro
*
, M. Candel, A. Barenca, A. Yustres, B. Garc a
Geoenvironmental Group, Civil Engineering School, University of Castilla-La Mancha, Avda. Camilo Jose Cela s/n, 13071 Ciudad Real, Spain
Received 27 February 2006; received in revised form 24 January 2007; accepted 25 January 2007
Available online 23 March 2007
Abstract
This paper analyzes the application of a grid-search approach for the estimation of modied Cam clay parameters from triaxial tests.
By means of the systematic sampling of the error, in addition to locating the area presenting the smallest error, its roughness, is also
characterized. This is a valuable information to evaluate the quality of the identication that has been carried out. The methodology
proposed here does not aspire to be the solution to the problem of parameter identication. The aim is simply to provide a tool which
may aid users with criteria.
2007 Elsevier Ltd. All rights reserved.
Keywords: Parameter estimation; Modied Cam clay model; Triaxial; Grid-search; Error topology
1. Introduction
Since their rst introduction (see [1,2]), critical state
models have successful in describing many of the most
important features of the mechanical behavior of soil [3].
Although several dierent formulations have been put
forth to improve the quality of the predictions of this
behavior, the modied Cam clay model (MCC) is still
widely referenced and used in solving boundary value
problems in geotechnical engineering [4,5]. The overwhelm-
ing acceptance of this model has even led to it being applied
to simulate compaction, tillage, stresses around growing
roots and other deformation events in agricultural engi-
neering [6]. Like other soil models, the MCC model
improves its performance if parameter estimation is carried
out by choosing experimental data obtained from tests with
stress levels, stress states and stress paths close to those for
which numerical predictions are subsequently required [7].
Moreover, predictions are also improved if, when estimat-
ing model parameters, inverse analysis or identication
techniques are applied to the interpretation of the experi-
mental data [8,6]. However, this gives rise to a minimiza-
tion problem which is generally not too easy to solve. To
collaborate on the resolution of this problem, in this paper
we present a tool aimed to support the identication of
MCC parameters from triaxial tests.
2. Triaxial tests: error topology for the MCC model
The denition of a critical state model requires the deter-
mination ve constitutive parameters: (i) slope M of the
critical state line in the p
0
q eective stress space (where
p
0
= (r
1
+ r
2
+ r
3
)/3 and q = r
1
r
3
, being r
1
, r
2
and r
3
the principal stresses); (ii, iii) slopes j and k of virgin com-
pression and unloadreload in the e lnp
0
space (where e
is the void ratio); (iv) the location of the normal consolida-
tion line in the compression plane dened by the void ratio
e
1
at p
0
equal to 1 kPa; and, (v) some elastic property, such
as the Poissons ratio m. When the results of the triaxial
tests to identify these parameters are analyzed, all the
experimental data are usually equally reliable. Therefore,
the usual procedure is to select a least-squares tting crite-
ria, adopting square error SE as the objective or merit
function:
SE
X
n
i1
e
aT;i
e
aM;i

2
1
0266-352X/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2007.01.007
*
Corresponding author. Tel.: +34 926 295 453; fax: +34 926 295 391.
E-mail address: vicente.navarro@uclm.es (V. Navarro).
www.elsevier.com/locate/compgeo
Available online at www.sciencedirect.com
Computers and Geotechnics 34 (2007) 524531
where n is the number of data points, and e
a
is the axial
strain. Subscripts T and M dene Test and Model val-
ues respectively. In certain cases, instead of using the axial
strain, the deviatoric strain e
S
is used. It is dened for triax-
ial conditions as e
S
= 2 (e
1
e
3
)/3, where e
1
and e
3
are the
principal strains. The root of the mean square error
RMSE = (SE/n)
1/2
is also commonly used as the merit
function. In any event, experience tells us that these types
of objective functions have a very irregular shape for the
MCC model, and therefore may be sensitive to the starting
point, and converge on a local minimum rather than the
global minimum of the function [7]. This can be seen in
Fig. 1, which synthesizes a comprehensive simulation exer-
cise. Here, parameters from column 2 in Table 1 were used
to simulate the conventional triaxial tests dened in Table
2. For each numerical test we obtained deformations
e
aT
= [e
aT,1
, . . . , e
aT,n
], which were adopted as a reference.
We then carried out the simulations associated with each
one of the vectors of parameters x = [M, k, j, e
1
, m] gener-
ated after discretizing the search space dened in columns
3 and 4 of Table 1 by means of the grid dened in column
5. So, each numerical test has associated 87,846 (11 11
11 11 6) vectors e
aM
= [e
aM,1
, . . . , e
aM,n
], which, when
compared with e
aT
, allowed us to obtain a systematic sam-
pling (SyS) of the RMSE. A graph of the variation in the
RMSE was made by drawing the lowest RMSE values
which are found as one advance along the grid, starting
0.001
0.01
0.1
1
1 6 11 16 21 26 31 36 41 46
R
M
S
E

/

R
M
S
E
M
0.001
0.01
0.1
1
R
M
S
E

/

R
M
S
E
M
0.01
0.1
1
R
M
S
E

/

R
M
S
E
M
D-OC D-NC U-OC U-NC
combined
a
b
c
Fig. 1. Curves of minimums: variation of the RMSE in the grids under
study (see Table 1). A graph was drawn of the lowest RMSE values which
are found as you advance along each one of the grids, starting at the point
where the parameters exhibit minimum values, passing through the point
at which the RMSE reaches its minimum and ending at the point where all
the parameters show their maximum values. (a) Weald Clay, (b) Kaolin
and (c) London Clay from Table 1. The 12 curves have been made
dimensionless by dividing them by the maximum of the root of the mean
square error, RMSE
M
, recorded in each one. The numerical tests are
identied in the gure legend (see Table 2).
Table 1
Description of the examples used to illustrate the roughness in the error
Value Search space GD
Max Min
Weald Clay
M 0.95 1.2 0.8 10
k 0.093 0.2 0.04 10
j 0.035 0.04 0.01 10
e
1
(p
0
= 1 kPa) 1.100 1.2 0.8 10
m
0
0.3 0.45 0.2 5
Kaolin
M 1.02 1.2 0.8 10
k 0.26 0.3 0.1 10
j 0.05 0.06 0.03 10
e
1
(p
0
= 1 kPa) 2.913 2.9 2.5 10
m
0
0.3 0.4 0.2 5
London Clay
M 0.888 1 0.6 10
k 0.161 0.3 0.04 10
j 0.062 0.05 0.01 10
e
1
(p
0
= 1 kPa) 1.828 2 1.6 10
m
0
0.3 0.4 0.2 5
Column 2 shows the parameters used for the three soils analyzed. Col-
umns 3 and 4 indicate the space where error topology is described. Col-
umn 5 denes the number of grid divisions employed to characterize the
above topology.
Table 2
Denition of the tests simulated to illustrate error topology
p
0
CSD
(kPa) p
0
O
(kPa) Drained e
O
W K L
D-OC 400 100 Yes 0.592 1.424 0.949
D-NC 400 400 Yes 0.543 1.355 0.863
U-OC 400 100 No 0.592 1.424 0.949
U-NC 400 400 No 0.543 1.355 0.863
The identication of the tests is shown in column 1. They were all con-
ventional triaxial tests: after isotropic consolidation, the deviatoric stress
was increased with the chamber pressure remaining constant. It was
assumed that the eective preconsolidation of the soil, p
0
CSD
, was always
equal to 400 kPa. After consolidation, the mean eective stress, p
0
o
was
assumed to be equal to 100 kPa (overconsolidated samples), and 400 kPa
(normally consolidated samples). Both drained and undrained tests were
simulated. Columns 57 indicate the void ratio, e
O
, after consolidation for
the three soils under consideration (W, Weald Clay; K, Kaolin; L, London
Clay; see Table 1).
V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531 525
at the point where the parameters exhibit minimum values,
passing through the point at which the RMSE reaches its
minimum, and ending at the point where all the parameters
show their maximum values. This resulted in the curve of
minimums shown in Fig. 1, which illustrates that dierent
local minimum values do exist. The local minimums can
also be seen in Fig. 2. In this gure, for the D-NC test (Ta-
ble 2) on Weald Clay (see Table 1), the variation in the
RMSE has been represented with M and k, keeping con-
stants j, e
1
and m at the values at which the RMSE is the
lowest. Fig. 2 also allows us to corroborate the irregular
or rough shape of the RMSE.
Moreover both Figs. 1 and 2 reveal that parameter iden-
tication is often an ill-posed problem in the sense of
Hadamard [9], as indicated by the plateaus that can be
seen in the area around the minimum values in both g-
ures. To assess the extent to which these plateaus are
important, it is useful to obtain a histogram, like the one
portrayed in Fig. 3, for each identication process. Here,
the range of the RMSEs recorded in the simulation of test
D-NC (Table 2) on Weald Clay (Table 1) has been divided
into 10 intervals of exponentially increasing amplitude. As
can be seen, a relatively substantial number of points (191)
on the grid in which the search space has been discretized,
dene RMSE values which are similar to the minimum.
Consequently, some degree of uncertainty arises when
dening the optimum value of x.
Kirby et al. [6] made a proposal to reduce uncertainty
by conducting a multiobjective identication. This was
done as follows. In drained tests, in addition to entering
the data relative to the axial or deviatoric deformation,
we used data related to the evolution of the void ratio.
Merit function F
D
= RMSE
e
RMSE
e
was dened where
the root mean square error associated with the axial
or deviatoric strain (RMSE
e
) is equal to the RMSE
dened up to now, while the RMSE
e
was calculated
Fig. 2. Variation in the RMSE with M and k, maintaining constants j, e
1
and m at the values where the RMSE is the lowest, for test D-NC (Table 2) on
Weald Clay (see Table 1).
191
100
199
1265
2560
3222
3659
8196
957
1510
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
7
.
5
E
-
0
3
1
.
1
E
-
0
2
1
.
5
E
-
0
2
2
.
2
E
-
0
2
3
.
1
E
-
0
2
4
.
6
E
-
0
2
6
.
7
E
-
0
2
9
.
7
E
-
0
2
1
.
4
E
-
0
1
2
.
1
E
-
0
1
RMSE
f
r
e
c
u
e
n
c
y
Fig. 3. Histogram of the RMSE (its range has been divided into 10
intervals of exponentially increasing amplitude) related to test D-NC
(Table 2) on Weald Clay (Table 1).
0.001
0.01
0.1
1
1 6 11 16 21 26 31
grid steps
R
M
S
E
RSME
FD
Fig. 4. A comparison of the evolution of the RMSE and F
D
associated
with test D-NC (Table 2) on Weald Clay (Table 1).
526 V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531
by replacing e with e. In undrained cases, merit function
F
U
= RMSE
e
RMSE
q
was used. Here the RMSE
q
error
associated with the stress path was dened as:
RMSE
q

1
n
X
n
i1
q
T;i
q
M;i
q
T;AV
!
2
0
@
1
A
1=2
2
where subscripts T and M dene, again, Test and Model
values, and q
T,AV
is the average value of q throughout
the test. In Fig. 4, we have compared the evolution of the
RMSE and F
D
associated with the numerical test in
Fig. 2. Similar to what occurred when performing the same
exercise with other tests and soils, the uncertainty has been
reduced by very little. This would seem to indicate that we
have not succeeding in introducing very much independent
information by using the multiobjective analysis.
Therefore, an analysis was carried out to determine the
eectiveness of combining data from several tests. Fig. 5
presents a comparison of the curves of minimums of Weald
Clay (Fig. 1a), with the minimum values obtained after
jointly analyzing the four tests on the same clay described
in Table 2. As was also demonstrated in the other soils
under study, although the plateaus seem to be reduced,
the roughness remains. While in practice, the joint analysis
of several triaxial tests should allow us to clarify the reli-
ability of experimental data [7], in ideally reliable numer-
ical tests the results show no signicant improvement.
The basic question probably lies in analyzing the extent
to which a conventional triaxial test is actually able to con-
tribute sucient information to carry out the identication
proposed here. This analysis, however, goes beyond the
scope of this paper. Here we will focus on putting forth a
method to support the interpretation of the triaxial tests
that are usually carried out in Soil Mechanics laboratories.
0.001
0.01
0.1
1
1 6 11 16 21 26 31 36 41 46
grid steps
R
M
S
E
/

R
M
S
E
M
D-OC D-NC U-OC U-NC combined
Fig. 5. A comparison of the four curves of the minimum values of the
RMSE related to Weald Clay (Fig. 1a) with the minimums obtained by
jointly analyzing the four tests on the same clay (see Table 2) associated
with the four curves.
0
100
200
300
400
500
600
350 375 400 425 450 475 500 525 550 575 600
p

' (kPa)
q
(
k
P
a
)
test
SyS
local
0
100
200
300
400
500
600
0 0.04 0.08 0.12 0.16 0.2 0.24 0.28 0.32 0.36
a
q
(
k
P
a
)
test
SyS
local
0.380
0.400
0.420
0.440
0.460
0.480
0.500
0.520
350 375 400 425 450 475 500 525 550 575 600
p' (kPa)
v
o
i
d

r
a
t
i
o
test
SyS
local
Fig. 6. Trajectories followed in spaces p
0
q, e
a
q, and p
0
e when test D-NC (Table 2) was simulated on Weald Clay (Table 1) by using the parameters
identied after performing the SyS and the grid-search (SyS + local identication).
V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531 527
3. Proposed identication process
We might not have been aware of the uncertainty asso-
ciated with the determination of the optimal if we had only
performed a t-by-eye (expression adapted from Press
et al. [10]) by means of curves p
0
q, qe
a
, ye
a
Du (where
Du is the increase in pore water pressure). The good t seen
in Fig. 6 for the simulation of the test associated with Fig. 2
(numerical test D-NC, Table 2, on Weald Clay, Table 1)
does not give us any indication of the uncertainty revealed
in Figs. 24. Moreover, as can be seen in Fig. 6, the t is
even better if we use the optimal value found with the
SyS as the initial value for local minimization using gradi-
ent-based techniques (see [11,12,8]). Fig. 7 conrms the
identication of parameters having errors of generally less
than 1% in the cases analyzed with this double global/local
strategy. If, however, in the problem associated with Fig. 2,
we carry out a simple sensitivity analysis based on the opti-
mal value obtained after local identication, we will once
again be able to see that the topology of the error is consid-
erably irregular (see Fig. 8). Therefore, although in the
exercises carried out on the soils shown in Table 1 the iden-
tication was satisfactory, we cannot be sure that by apply-
ing the same methodology, the results will be as successful
in other cases.
The methodology proposed here, i.e., SyS/global identi-
cation + local identication, is commonly referred to as a
grid-search. From a numerical standpoint, it is probably
the simplest complete method for bound constrained prob-
lems [13]. To identify the minimum, it would be more e-
cient to use one of the global search algorithms currently
available (see, for example Horst and Pardalos [14], or
Pinter [15]). In the problems analyzed here, however, it
took us no more than 4 min on a laptop computer with a
1500 MHz Intel Pentium processor to obtain the 87,846
error values associated with each test (each of them having
121 data points). Therefore, given the fact that the CPU
time spent is acceptable, and that with this it is possible
to obtain a good characterization of the roughness of the
RMSE, we recommend the use of the SyS to analyze triax-
ial tests. However, grid-search is not appropriate for
another constitutive model with more parameters than
Cam clay. In this case, from a computational point of view,
it may be excessively expensive. Nevertheless, parameter
identication should not focus only on determining the
optimal value. It should also provide us with information
on the existence of other combinations of parameters that
are capable of reproducing, with comparable quality, the
experimental behaviors observed. It is not a question of
mere minimization, but also of oering criteria to be able
to evaluate the likelihood of the identication. Therefore,
the application of a global identication procedure which
can oer us an indication of the error topology is strongly
recommended.
To assist with this task we have designed a Microsoft
Excel spreadsheet which uses a macro to automatically
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
10
100
M e1
r
e
l
a
t
i
v
e

e
r
r
o
r

(
%
)
r
e
l
a
t
i
v
e

e
r
r
o
r

(
%
)
r
e
l
a
t
i
v
e

e
r
r
o
r

(
%
)
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
10
100
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
10
100
D-OCG D-OCL D-NCG D-NCL
U-OCG U-OCL U-NCG U-NCL
Fig. 7. Relative error in each parameter of the model. The error was
dened for each parameter and each type of material ((a) Weald Clay; (b)
Kaolin; (c) London Clay) using the values dened in column 2 of Table 1
as a reference. The numerical tests are identied in the gure legend (see
Table 2). Also indicated is whether the identication was based on the SyS
alone (identied by the letter G), or if a local search was also used
(L).
-

e(p'=1)
kappa
lambda
MM

e(p=1)
0.0E+00
5.0E-02
1.0E-01
1.5E-01
2.0E-01
2.5E-01
3.0E-01
3.5E-01
4.0E-01
-10% -8% -6% -4% -2% 0% 2% 4% 6% 8% 10%
variation

e(p'=1)
MM

e(p=1)
R
M
S
E
Fig. 8. Sensitivity analysis of the error around the optimal value (column
2 of Table 4) resulting from the grid-search associated with test D-NC
(Table 2) on Weald Clay (see Table 1). The percentage of variation
corresponding to each parameter is depicted on the x-axis.
528 V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531
carry out both the calculations associated with the SyS as
well as the illustration of the graphs shown in Figs. 13
and 6. The spreadsheet includes modules which, assuming
that the samples analyzed have a perfectly homogeneous
behavior, allow us to characterize their behavior for tri-
axials tests in which the ratio Dp/Dq is constant (where
p is the mean total stress). For undrained tests, since
the eective stress path is a monotonic function con-
trolled by the initial void ratio, it was relatively easy to
integrate the basic equations of the MCC model
(described, for example, in Schoeld and Wroth [16], or
Wood [17]) to obtain closed-form expressions of q and
e
a
(equal to e
S
for undrained conditions) as functions of
p
0
. For drained tests, the eective stress path may not
be a monotonic function, as occurs with the softening
of an over-consolidated clay. So, a numerical integration
(second order RungeKutta) is carried out to compute
both the void ratio and the deviatoric strain. Both types
of modules, undrained and drained, have been duly ver-
ied. Although the systematic sampling needs no user
intervention once the analysis has been started, the user
must dene the search space to be considered, as well
as the density of the grid on which it is discretized. In
short, the user must direct the search. The local iden-
tication was carried out with the help of the minimiza-
tion algorithms included in the Solver utility on the
spreadsheet. The analysis of the sensitivity of the solution
found through local identication (Fig. 8) was also done
automatically with a macro. The spreadsheet can be
obtained by submitting a request to the rst author.
4. Examples of application
This section includes two examples of the application of
the method. First of all, we analyzed the tests performed by
Bishop and Henkel [18] on Weald Clay, using the results
reported by Atkinson and Bransby [19] as a reference. As
can be seen in Table 3, only tests on normally consolidated
samples were analyzed. By simulating the tests on overcon-
solidated samples (tests with an overconsolidation ratio of
24 were studied) it was found that the MCC model signif-
icantly overestimates yield stresses. This is a well known
consequence that occurs if yielding takes place to the left
of the maximum of the Cam clay ellipse (dry or super-
critical side) [3]. In this article, we will not analyze this
question. An excellent study for consultation can be found
in Gens and Potts [3]. Here, we have limited ourselves to
adopting an MCC model, which, despite its shortcomings
on the supercritical side, remains the most widely used crit-
ical state model (Borja and Andrade [20]).
Both combined and individual analyses of data from
tests on normally consolidated samples have allowed us
to obtain, by means of the SyS, a good grouping of the
parameters identied with regard to the parameters pro-
posed by Schoeld and Wroth [16] for Weald Clay (see
Tables 1 and 4). However, as can be seen in Fig. 9, error
roughness is not negligible. Moreover, the RMSE value
obtained after simulating both tests using the parameters
of Schoeld and Wroth [16] is higher than the value result-
ing from the parameters identied as optimal (Table 4).
What this does is to emphasize the uncertainty associated
with the estimation of optimal parameters.
To nish the analysis of this example, it is interesting to
note that, while joint identication has not eliminated error
roughness (local minimums still exist), it has, in fact,
reduced the plateau around the minimum (see Fig. 9).
It has gone from 100 (drained case) and 67 (undrained
case) vectors of parameters that dene an RMSE similar
to that of the minimum, to a mere 20.
Secondly, included here is the application of the proce-
dure proposed to obtain the parameters of Spestone Kao-
lin analysed by Wood et al. [7]. The control maintained
in both the making of this soil and the performance of
the tests makes the use of these experiments, as an exer-
Table 4
Optimal values found for tests D-NC (column 2) and U-NC (column 3) carried out by Bishop and Henkel [18] on Weald Clay (see Table 3), in addition to
the RMSE of both grid-search procedures
O D-NC O U-NC O (U&D)-NC Max Min GD D-NC S&W U-NC S&W
M 0.88 0.92 0.92 1.4 0.8 10 0.95
k 0.095 0.095 0.106 0.15 0.04 10 0.093
j 0.03 0.035 0.03 0.06 0.01 10 0.035
e
1
(p
0
= 1 kPa) 1.139 1.139 1.200 1.2 0.8 10 1.100
m
0
0.28 0.45 0.4 0.45 0.2 5 0.3
RMSE 2.57 10
3
1.39 10
2
1.25 10
2
3.19 10
2
2.56 10
2
Column 4 shows the parameters and the RMSE associated with the joint identication of the parameters, using the results of the two tests, D-NC and
U-NC simultaneously. The search space dened in columns 5 and 6, as well as the number of grid divisions in column 7, were employed in the three
identications. The last two cells indicate the RMSE obtained with each test provided that it is simulated following the parameters proposed by Schoeld
and Wroth [16], assuming that m = 0.3.
Table 3
Tests on normally consolidated samples performed by Bishop and Henkel
[18] on Weald Clay
p
0
CSD
(kPa) p
0
O
(kPa) Drained e
O
D-NC 207 207 Yes 0.632
U-NC 207 207 No 0.632
As in Table 2, the p
0
CSD
indicates the eective preconsolidation, and p
0
O
the
mean eective stress after consolidation.
V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531 529
cise in validation, particularly advantageous. Similar to
the work carried out by Wood et al. [7] for optimization,
among the dierent tests done, we analyzed the drained
cycle of constant p
0
= 150 kPa loading from q = 0 to
q = 100 kPa of test L1. Although this soil was tested in
a true triaxial apparatus, the whole test was performed
with r
2
= r
3
[7], and can be interpreted as a conventional
triaxial test.
The parameters were identied by using the search
space and grid dened in Table 5. After conducting the
SyS, it was found that the minimum was close to the val-
ues assigned in column x
G
of Table 5. The same table
shows the parameters identied after the local search,
x
L
. In Fig. 10 it is possible to see that the t obtained
after the local search appears to improve upon the one
reported by Wood et al. [7]. Thus, the RMSE values
resulting from the parameters proposed by Wood et al.
[7] when analyzing both the whole cycle of loading, as
well as only the initial part of the shearing, are higher
(see Table 5). Since the local search applied in our paper
is not of better quality than the one applied by Wood
et al. [7], the results would seem to indicate that the
SyS made it possible to start at a more consistent initial
value than the initial value used by these authors,
thereby improving the end result of the search. It is
interesting to take into account that, in our model, we
have introduced the eective stress path as known action,
whereas Wood et al. [7] have used the observed strain
path as input control. Provided that the mean stress
paths are very sensitive to the erratic changes experi-
enced by the experimental strain paths, lower variations
of the mean stress are obtained as higher values of j
and k are considered. In consequence, an extremely high
value of j was identied by Wood et al. [7] (see Table 5,
and Holtz and Kovacs [21], for instance, to look up typ-
ical values typical values of j). Hence, for the shear
modulus considered by Wood et al. [7] (4090 kPa and
5000 kPa, for W1 and W2, respectively), negative values
of the Poissons ratio were obtained. In this sense, the
parameters identied by Wood et al. [7] are perhaps less
consistent than those that we have obtained.
Table 5
Search space (columns 2 and 3) and grid (column 4) used to identify the parameters of Spestone Kaolin tested by Wood et al. [7] (p
0
CSD
150 kPa,
p
0
O
150 kPa, e
O
= 1.479)
Max Min GD x
G
x
L
W
1
W
2
M 0.9 0.7 10 0.78 0.73 0.82 0.78
k 0.3 0.1 10 0.22 0.18 0.62 0.58
j 0.03 0.01 10 0.018 0.043 0.350 0.430
e
1
(p
0
= 1 kPa) 2.9 2.5 10 2.62 2.404
m
0
0.45 0.20 5 0.35 0.34 0.42 0.42
RMSE 2.860 10
3
2.475 10
3
Column x
G
indicates the parameters identied after carrying out the SyS. The parameters identied after the local search are given in column x
L
. Columns
6 and 7 contain the parameters proposed by Wood et al. [7], after analyzing both the complete cycle of loading, W
1
, as well as only the initial part of the
shearing (q/p
0
6 0.5), W
2
.
0
25
50
75
100
125
0 0.02 0.04 0.06 0.08 0.1 0.12

q
(
k
P
a
)
test
grid-search
Wood et al. (1992)
s
Fig. 10. Results in space qe
S
of the test carried out by Wood et al. [7] on
Spestone Kaolin. The nal drained-shearing cycle of constant
p
0
= 150 kPa loading from q = 0 to q = 100 kPa was analyzed. The
experimental results are compared with those obtained using both
parameters resulting from the grid-search (column 6 of Table 5), as well
as the optimal parameters identied by Wood et al. [7] after analyzing the
complete loading cycle (column W1 of Table 5).
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
1 5 9 13 17 21 25 29 33
grid steps
R
M
S
E
combined
D-NC
U-NC
Fig. 9. A comparison of the curves of minimums of the RMSE obtained
after performing the grid-search using data from the tests carried out by
Schoeld and Wroth [16] on normally consolidated samples. A graph is
drawn of the minimum values obtained by using data from a drained test
(D-NC), an undrained test (U-NC), and by the identication carried out
joining together the data from both tests (curve labeled as combined).
530 V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531
5. Conclusions
The application of the grid-search to the estimation of
modied Cam clay parameters from triaxial tests has pro-
duced good quality results at a lower computational cost.
Moreover, the systematic sampling of the error has made
it possible to characterize its roughness, which is a reec-
tion of the uncertainty associated with the location of the
optimal value and a way to measure the quality of the
identication carried out. To make its application more
user-friendly, a free-access spreadsheet was developed.
However, the tool is not the one trying to locate the opti-
mal parameters. Although it is of great help, it is the user
and his or her criteria, who actually solve the identication
problem.
Acknowledgements
This research was nanced in part by a Research Grant
awarded to the third and fourth authors by the Education
and Research Department of the Castilla-La Mancha Re-
gional Government and the European Social Fund within
the framework of the Integrated Operative Programme
for Castilla-La Mancha 20002006, approved by Commis-
sion Decision C(2001) 525/1.
References
[1] Roscoe KH, Schoeld AN. Mechanical behaviour of an idealised
wet clay. In: Proceedings second European conference on soil
mechanics and foundation engineering, Wiesbaden, vol. 1; 1963, p.
4754.
[2] Roscoe KH, Burland JB. On the generalised stressstrain behaviour
of wet clay. In: Heyman J, Leckie FA, editors. Engineering
plasticity. Cambridge University Press; 1968. p. 535609.
[3] Gens A, Potts DM. Critical state models in computational geome-
chanics. Eng Comput 1988;5:17897.
[4] Potts DM, Zdravkovic L. Finite element analysis in geotechnical
engineering: theory. London: Thomas; 1999.
[5] Liu MD, Carter JP. A structured Cam clay model. Can Geotech J
2002;39(6):131332.
[6] Kirby JM, OSullivan MF, Wood JT. Estimating critical state soil
mechanics parameters from constant cell volume triaxial cell. Eur J
Soil Sci 1998;49(1):8593.
[7] Wood DM, Mackenzie NL, Chan AHC. Selection of parameters
for numerical predictions. In: Houlsby GT, Schoeld AN,
editors. Predictive soil mechanics proceedings of the Wroth
memorial symposium. London: Thomas Telford; 1992. p.
496512.
[8] Calvello M, Finno RJ. Selecting parameters to optimize in model
calibration by inverse analysis. Comput Geotech 2004;31(5):
41125.
[9] Hadamard J. Sur les proble`mes aux derivees partielles et leur
signication physique. Princeton Univ Bull 1902;13:4952.
[10] Press WH, Teulowsky SA, Vetterling WT, Flannery BP, editors.
Numerical recipes in C++: the art of scientic computing. Cam-
bridge: Cambridge University Press; 2002.
[11] Ledesma A, Gens A, Alonso EE. Parameter and variance estimation
in geotechnical backanalysis using prior information. Int J Numer
Anal Methods Geomech 1996;20(2):11941.
[12] Yang Z, Elgamal A. Application of unconstrained optimization and
sensitivity analysis to calibration of a soil constitutive model. Int J
Numer Anal Methods Geomech 2003;27(15):127797.
[13] Neumaier A. Complete search in continuous global optimization and
constraint satisfaction. In: Iserles A, editor. Acta Numerica 2000.
Cambridge University Press; 2004. p. 271369.
[14] Horst R, Pardalos PM. Handbook of global optimization. Dordrecht:
Kluwer Academic Publishers; 1995.
[15] Pinter JD. Global optimization in action: continuous and Lipschitz
optimization: algorithms, implementations and applications (noncon-
vex optimization and its applications). Dordrecht: Kluwer Academic
Publishers; 1996.
[16] Schoeld AN, Wroth CP. Critical state soil mechanics. London:
McGraw-Hill; 1968.
[17] Wood DM. Soil behaviour and critical state soil mechanics.
Cambridge: Cambridge University Press; 1990.
[18] Bishop AW, Henkel DJ. The measurement of soil properties in the
triaxial test. 2nd ed. London: Edward Arnold; 1962.
[19] Atkinson JH, Bransby PL. The mechanics of soils: an introduction to
critical state soil mechanics. London: McGraw-Hill; 1978.
[20] Borja RI, Andrade JE. Plastic deformation and instability in discrete
granular materials with spatially varying density. In: VIII Interna-
tional Conference on Computational Plasticity, 57 September 2005,
Barcelona, Spain. Conference extended abstracts. http://con-
gress.cimne.upc.es/complas05/frontal/papers.asp, 4 pp.
[21] Holtz RD, Kovacs WD. An introduction to geotechnical engineering.
New Jersey: Prentice-Hall; 1981.
V. Navarro et al. / Computers and Geotechnics 34 (2007) 524531 531

You might also like