Beruflich Dokumente
Kultur Dokumente
q
e
1
7
and
R
e
e
3
e
1
8
where e
3
and e
1
are the resilient radial and axial
strains, respectively. Resilient modulus and strain
ratio are determined as secant values in terms of
deviator stress. The denitions of resilient modulus
and strain ratio are similar to elastic or Youngs
modulus and Poisson ratio, respectively, which are
applicable to linear elastic response. It should be
noted that, at stress levels encountered during this
investigation, granular materials are commonly
inelastic. However, the term Poisson ratio will be
used interchangeably in following sections, where
reference is given to previously published results or
procedures.
2.4 Resilient models
A common approach in pavement design is to
consider the response of granular materials as linear
isotropic elastic, i.e. the materials can be character-
ized by two elastic parameters e.g. Youngs modulus
and Poisson ratio. However, the resilient response of
granular materials is unmistakably nonlinear and
depends on stress (or strain) level. To recognize this
circumstance, Youngs modulus is replaced by resil-
ient modulus to indicate typical inelastic resilient
behavior. The resilient modulus is expressed as a
function of stress level. This practice represents one
category of resilient models. Researchers have also
extended the resilient modulus framework by intro-
ducing degrees of anisotropy in the derivation of
elastic parameters required to describe the resilient
behavior [68]. In the case of cross-anisotropy
(transverse isotropy), ve elastic parameters are
Fig. 4 Stress paths used in
the triaxial tests (total
stress)
Materials and Structures (2008) 41:509525 513
required to characterize the behavior: resilient mod-
ulus and Poisson ratio in the vertical and horizontal
direction, respectively, and shear modulus. However,
it is not possible to uniquely derive the resilient
parameters from cylindrically conned triaxial tests;
either some of the parameters have to be assumed
constant or they are derived using numerical optimi-
zation. Furthermore, the techniques described by [6
8] would preferably require different testing sched-
ules and/or triaxial equipment than was used in this
investigation. Another category is based on decom-
posing the stresses and strains into shear and
volumetric components. In this work, the stress
dependent resilient modulus is used to model results
obtained from triaxial tests using constant conning
pressure while the shear and volumetric approach is
used mainly for measurements using cyclic conning
pressure. It should be noted that neither of these
model categories are restricted to one type of test
mode (constant or cyclic conning pressure).
2.4.1 Resilient modulus models
Probably the most well-known and commonly used
model is generally called the kh model. It assumes
that resilient modulus can be related to the sum of
principal stresses h according to:
M
c
k
1
h
k
2
9
where k
1
and k
2
are regression parameters. The origin
of the model dates back to at least 1960s. Brown and
Pell [9] plotted the modulus versus the rst stress
invariant (I
1
or h) in a loglog diagram. Seed et al.
[10] suggested a power law relationship relating
resilient modulus to sum of principal stresses, while
Hicks and Monismith [11] concluded that Eq. 9 was
veried in most of their test results.
From analyzing eld data, May and Witczak [12]
concluded that the resilient modulus is not solely
dependent on sum of principal stresses but also on
shear strain level. This shear effect was later recog-
nized by Uzan [13], who suggested the following
relationship:
M
c
k
1
h
k
2
r
k
4
d
10
where k
1
, k
2
and k
4
are regression parameters and r
d
is the deviatoric stress invariant q.
A disadvantage with these models is that they do
not cater for strains in the transversal direction.
Commonly, a constant Poisson ratio is used in this
connection. However, the models can also be
extended by use of stress dependent Poisson ratio
relationships to improve shear and volumetric strain
predictions [14]. Furthermore, In 1992 Uzan [15]
extended the resilient description expressed by Eq. 10
with a relationship expressing a nonlinear stress
dependent Poisson ratio as a function of sum of
principal stress (I
1
or h) and the second deviatoric
stress invariant (J
2
), together with the exponents of
Eq. 10, and additionally two regression constants. To
estimate the parameters, Uzan suggests that the
resilient modulus and Poisson ratio are simultaneously
tted using nonlinear regression. Basically, this inves-
tigation was not extended to include relationships
describing stress dependent Poisson ratio. However, as
an example of this category of models, results are also
tted using the extended Uzan model [15].
2.4.2 Shearvolumetric models
The second category of resilient models is based on
decomposing the principal stress and strain tensors,
respectively, into two tensors: a deviatoric (or shear)
and a volumetric. In this section, ve different
models evaluated in this investigation are briey
introduced.
In 1980, Boyce [16] presented general equations
for characterizing nonlinear behavior in terms of bulk
(K) and shear modulus (G), which are expressed as
stress dependent. Basic assumptions are elastic and
isotropic material behavior. In the case of granular
materials, bulk and shear modulus are expressed as
power functions of mean normal stress (p). By
imposing a constraint of no net loss of energy,
volumetric (e
v
) and shear (e
q
) strains are expressed as:
e
v
p
n
K
1
1 b
q
2
p
2
_ _
11
e
q
p
n
3G
1
q
p
12
b
1 n K
1
6G
1
13
514 Materials and Structures (2008) 41:509525
where n, K
1
, G
1
are regression parameters. The b
parameter is an effect of the thermodynamic con-
straint of no loss of energy and couples the volumet-
ric and shear behavior.
Brown and Pappin [17] developed a model com-
monly referred to as the contour model. The basic
idea is to mathematically describe observed strain
contours in terms of shear and normal stress state. For
this purpose, they derived the following equations:
e
v
p
A
_ _
m
1 B
q
p
_ _
n
_ _
14
e
q
Cq
p D
p
2
q
2
_
p=2
_ _
c
15
where A, m, B, n, C, D and r are regression
parameters. The term within parentheses in Eq. 15
depends on the length of the stress path. In contrast to
the Boyce model, the contour model does not couple
shear and volumetric behavior.
Based on the Boyce model, Jouve and Elhannani
[18] investigated the efciency of a number of
modications to the original model. By assuming
orthotropy and introducing an orthotropic parameter
(n), they proposed to express strains as:
e
v
p
1n
a
p
n
1
K
a
nn
q
p
b
K
a
q
2
p
2
_ _
16
e
q
p
1n
a
p
n
1
3G
a
q
p
n
_ _
17
where n, K
a
, n, b and G
a
are regression parameters,
and p
a
is a constant used to normalize the dimension
of the mean normal stress term. In the most general
formulation, as used in this investigation, b is an
independent parameter but could also be determined
by Eq. 13.
Another way of introducing anisotropy (orthotropy
or cross-anisotropy) to the Boyce model was
described by Hornych et al. [19]. In their approach,
the major principal strain is adjusted by an empirical
parameter, coefcient of anisotropy c, which leads to
the following equations:
e
v
p
n
p
n1
a
c 2
3K
a
n 1
18G
a
c 2
q
_ _
2
c 1
3G
a
q
_ _
18
e
q
2
3
p
n
p
n1
a
c 1
3K
a
n 1
18G
a
c 1
q
_ _
2
2c 1
6G
a
q
_ _
19
where n, c, K
a
and G
a
are regression parameters and
p
a
is a constant, mainly aimed at normalizing the
dimensions. The corresponding modied stress in-
variants become:
p
cr
1
2r
3
3
20
q
cr
1
r
3
21
In contrast to the models presented so far, Hoff
et al. [20] presented a different approach although
still using the concept of dividing the model into
volumetric and shear components. The derivation of
their relationship is based on hyperelasticity. For
general stress conditions, Hoff et al. dene the strain
energy (U) function as:
U
K I
e
1
_ _
2
2
DI
e
1
J
e
2
2GJ
e
2
22
where K is bulk modulus, G is shear modulus, D is a
dilatancy constant used in the coupling of the rst
strain invariant I
e
1
_ _
and the second deviatoric strain
invariant J
e
2
_ _
. The stressstrain relationship is
obtained by partial differentiation and for the triaxial
test stress state it becomes:
q
p
_ _
3G De
q
D
e
q
=
3
K
_ _
e
q
e
v
_ _
23
To comply with the previously described model, the
compliance matrix is required instead of the stiffness
matrix in Eq. 23. Unfortunately, due to the coupling of
strains in the strain energy function, the stiffness
matrix cannot be inverted. In this investigation, a
Materials and Structures (2008) 41:509525 515
different regression model was used to determine the
parameters of the Hoff model. In contrast to Eqs. 11
19, where the parameters were estimated by nonlinear
regression minimizing the squared error between
measured and estimated strain, the nonlinear regres-
sion of the Hoff model is based on minimizing the
squared error between actually applied stresses and
estimated stress levels required to reach the measured
strains. In other words, for a given strain state,
determined by measured volumetric and shear strain,
the nonlinear regression estimate the stresses required
to reach this strain state and minimize the difference to
actually applied stresses. Consequently, the goodness-
of-t statistics of the Hoff model cannot easily be
compared to the other models used. Although, results
from tting of the Hoff model are presented for
reference and visual assessment.
An important model feature, inuencing the sta-
tistical evaluation, is number of parameters to be
estimated. Table 2 summarizes the models and
number of parameters used in each model.
2.5 Statistical analysis
To compare the ability of the different models to
explain measured mechanical behavior, two different
statistical procedures were used: the extra sum of
squares F-test and the Akaike information criterion
(AIC
c
). The extra sum of squares is a measure of the
marginal increase in the regression sum of squares
gained by adding another parameter to the model,
after which hypothesis testing (signicance testing) is
performed to determine whether this increase could
be regarded as statistically signicant. This approach
is commonly used in multiple regression to determine
which variables should be used in a statistical model.
The basis of the Akaike information criterion is
different. In this case, the procedure is based on
maximum likelihood and information entropy and is
not readily comparable to hypothesis testing used in
the extra sum of squares procedure. The Akaike
information criterion provides an estimate of the
relative distance between the tted model and the true
mechanism. This distance can be used for comparison
of the different models. More detailed information on
these statistical techniques can be found in textbooks
(e.g. [21, 22]) although a short introduction of the
calculated statistics is given here.
The extra sum of squares test compares sum of
squares obtained from the full model (maximum
number of parameters) and reduced models (exclud-
ing parameters). The F-ratio is determined by:
F
SSE
R
SSE
F
df
R
df
F
_
SSE
F
df
F
24
where SSE
R
and SSE
F
is the sum of squared errors for
the reduced and full model, respectively, and df
R
and
df
F
are the corresponding degrees of freedom. The
extra sum of squares can be used when the compared
models are nested, i.e. the full model contains only an
addition to the reduced model. The null hypothesis is
that the additional parameter equals zero. In this
investigation, results obtained using the kh and Uzan
model tting of resilient modulus are tested using the
extra sum of squares test. The criterion for statistical
signicance is based on the P-value; if obtained
values are lower than 0.05, the null hypothesis is
rejected.
When comparing the shearvolumetric models, the
Akaike information criterion is used. This criterion is
dened as:
AIC
c
n ln SSE n ln n 2K
2K K 1
n K 1
25
where n is the number of data points, SSE is the
sum of squared error and K is the number of
parameters plus one. The model with the lowest
AIC
c
-value is estimated to be most close to the true
mechanism generating the data. From Eq. 25, it can
be seen that the rst term depends on sum of
squared error, which commonly decreases with
increased number of model parameters. The last
two terms depends on number of parameters and
these will increase AIC
c
as the number of estimated
parameters increase, i.e. these terms balance the
Table 2 Summary of resilient shearvolumetric models and
the number of regression parameters
Model Number of
parameters
Boyce [16] 3
Brown and Pappin [17] 7
Jouve and Elhannani [18] 5
Hornych et al. [19] 4
Hoff et al. [20] 3
516 Materials and Structures (2008) 41:509525
reduction of SSE reached by introducing additional
parameters. In a sense, there is a penalty in the
criterion of adding too many parameters, to avoid
overtting or overparameterized models. Using more
parameters in a model will commonly decrease the
bias of the model while the variance is increased.
Furthermore, the precision of estimated parameters
might be poor. If the model is too complex, compared
to empirical data available, it might include spurious
effects or anomalies in measured data; even if a
model is physically incorrect, it might provide a
reasonable t to measured data if the number of tted
parameters is high. When comparing models using
AIC
c
, the absolute magnitude is not important but
rather differences between model candidates. A
model with a lower value of AIC
c
is considered to
be a better model. The absolute value of AIC
c
depends on units used and number of observations. In
this investigation, results are given as DAIC
c
i.e. the
increase of AIC
c
for each model compared to the
model with the lowest value.
3 Results
3.1 Constant conning pressure
From measurements using constant conning pres-
sure, resilient modulus and strain ratio were calcu-
lated and are summarized in Fig. 5. For resilient
modulus, results are tted to a power law relationship
in similarity to the kh model. The strain ratio is
expressed in terms of the stress ratio q/r
3
. The scatter
indicated in both diagrams of Fig. 5 is mainly caused
by the fact that the stress dependence of neither of
these two resilient parameters can be related to a
single stress invariant.
Figure 5 gives an impression of the differences in
mechanical behavior between the various gradings
tested at maximum water content. The different
gradings range from 0.8 showing the stiffest response
to 0.3, which is signicantly softer. There is also a
difference in stress dependency. Grading 0.8 shows
the largest increase in modulus with increased mean
normal stress, whereas 0.3 is inuenced to a lesser
degree. Concerning strain ratio, all gradings except
0.8 undergoes dilation at higher deviator stress levels,
grading 0.3 showing the highest levels.
The efciency of the kh and Uzan models in
tting measured results are visualized in Fig. 6, where
tted resilient modulus is shown versus measured
modulus of all the gradings used. Also shown is the
line of equality and, consequently, a successful
regression of results should adhere closely to this
line. Fitted parameters are given in Table 3.
In Fig. 6, it can be seen that the Uzan model
regression yields a closer t to measured results, even
though the kh model also performs reasonably well.
In this context, it should be remembered that the
addition of another parameter in the model is
expected to reduce the error between measured and
tted results.
Table 4 summarizes goodness-of-t statistics and
the result obtained using the extra sum of squares F-
test for the various gradings. The extra sum of square
procedure tests whether the added parameter cause a
signicant increase in sum of squares explained by
the regression.
Overall, for both models, the degree of explained
variance (R
2
) is high; as expected, it is slightly higher
Fig. 5 Resilient modulus and strain ratio of the four different
gradings at maximum water content
Materials and Structures (2008) 41:509525 517
for the Uzan model. It can also be seen that for all
gradings, except 0.3, the improvement caused by
adding the deviator stress in the model is statistically
signicant. From Fig. 5 it seems that for grading 0.3,
the inuence of stress level on resilient modulus is
smaller compared to the other gradings and in
particular grading 0.8 (i.e. the slope of the curve is
smaller).
Even though the kh and Uzan models are mainly
intended for description of resilient modulus, shear
and volumetric strains can be predicted using an
estimated (constant) Poisson ratio. In this investiga-
tion, the average strain ratio value of the entire stress
range for each grading was used (cf. Fig. 5). It was
found that this measure caused the smallest predictive
errors. Used values ranged from 0.53 for the 0.3
grading to 0.24 for the 0.8 grading. For comparison,
measured shear and volumetric strains were also
tted using the Boyce model. Since the ratio q/p is
constant (3/1) for all stress paths using constant
conning pressure, q/p was calculated as Dq/p
max
,
where p
max
includes the conning pressure. Acquired
results for all gradings are shown in Fig. 7 using the
Uzan and Boyce models, respectively. The results
obtained using the kh model are omitted since
predicted strains are essentially showing the same
pattern as those obtained using the Uzan model. In
the lower part of Fig. 7 tted results using the
extended Uzan model is also shown for comparison.
From Fig. 7 it can be seen that, using the Uzan
model and a constant Poisson ratio, shear strain
predictions are closer to measured results compared
to volumetric strains. The prediction of volumetric
strain is poor. At higher strain levels, it seems as the
Uzan model underpredicts the shear strains. The
prediction of negative volumetric strains is an effect
of using the average measured Poisson ratio, which
for grading 0.3 was 0.53. It should be noted that this
value is inconsistent with an elastic framework and
isotropic behavior. Compared to the Uzan model, the
Boyce model yields an improved t for both shear
and volumetric strains. The Boyce model tends to
underestimate the volumetric strains during dilation.
By extending the resilient description with a stress
dependent Poisson ratio, as in the lower part of Fig. 7
(extended Uzan), the prediction of shear and volu-
metric strains is greatly improved. When comparing
the outcomes shown in Fig. 7 (the two upper
diagrams) and in Table 5, it should be remembered
that the parameters in the Boyce model are estimated
based on the shear and volumetric strains, while the
strains for the Uzan model are predictions obtained
by tting of resilient modulus and using a constant
Poisson ratio. However, the results and inferences
would be basically the same if the parameters of the
kh and Uzan models are derived based on regression
of shear and volumetric strains. In Table 5, statistical
comparisons between the applied models are given.
In terms of shear and volumetric strains, the kh and
Fig. 6 Fitted versus measured values of resilient modulus for
all gradings and corresponding line of equality
Table 3 Parameters obtained for the different models and
gradings using constant conning pressure
Grading kh
k
1
(kPa) k
2
08 16.6 0.69
05 27.2 0.53
04 27.2 0.49
03 33.2 0.41
Uzan
k
1
(kPa) k
2
k
4
08 21.0 0.87 0.22
05 32.4 0.66 0.16
04 31.3 0.58 0.11
03 39.6 0.48 0.10
Boyce, GK model
K
1
(kPa) G
1
(kPa) n
08 23.8 12.0 0.52
05 26.2 7.99 0.50
04 24.8 5.21 0.46
03 53.4 4.83 0.50
Stress input was in kPa
518 Materials and Structures (2008) 41:509525
Uzan models are compared using the extra sum of
squares F-test and the comparisons with the Boyce
model are made using the Akaike information
criterion. In all cases, the Boyce model gave the
lowest AIC
c
-value.
As indicated in Table 5, there is no statistically
signicant improvement (P > 0.05) of predicted
results by using the Uzan model compared to the
more simple kh model. Furthermore, DAIC
c
-values
are only marginally lower for the Uzan model
compared to the kh model. For all gradings, the
AIC
c
-value was at least 10 units lower for the Boyce
model compared to the other models, which is a
strong indication of higher likelihood that this is a
better description of measured strain data. By using
the extended Uzan model, the goodness-of-t statis-
tics becomes more comparable to the Boyce model;
in two cases the Akaike information criterion indi-
cates similar likelihood of the two models and in the
remaining two cases the Boyce models is more likely
to yield an improved description of empirical data.
Another way of comparison of the models is to use
the Boyce tting of volumetric and shear strains for
prediction of resilient modulus. Overall, the predicted
values from this calculation are comparable to results
obtained using the kh model, i.e. slightly less precise
than the Uzan t.
3.2 Cyclic conning pressure
Results obtained at cyclic conning pressure were
analyzed using the shearvolumetric approach. Fig-
ure 8 illustrates typical results of such tests for
grading 0.5; shear and volumetric strain are shown as
function of mean normal stress at various ratios of
deviator to normal stress (q/p). In Fig. 8, the
nonlinear regressions of the Boyce and Brown
Pappin models, respectively, are also shown. The
Boyce model is the one (among the described in this
paper) containing the lowest number of estimated
parameters and the BrownPappin model the one
containing the highest number.
Visually from Fig. 8 the pattern of stress depen-
dency is clearly seen. Both volumetric and shear
strain depend not solely on mean normal stress but
also on stress ratio. At lower levels of q/p, the
volumetric strain mainly depends on mean normal
stress, but as the stress ratio increases, this pattern is
changed. At the highest stress ratio used (2.5), the
volumetric strain is almost constant in the normal
stress range tested. The shear strain is also strongly
inuenced by stress ratio; the higher the strain ratio
the higher the strain for a given mean normal stress.
As can be expected the lowest levels of shear strain
was measured under isotropic loading. Considering
the nonlinear regressions shown, it seems as the
BrownPappin model provides a closer adherence to
measured data compared to the Boyce model. How-
ever, this is what would be expected considering the
higher number of parameters used.
The outcome of the nonlinear regressions is
summarized in Table 6, in which all estimated
parameters for the different models and gradings
are given. It should be noted that the dimension of the
models and parameters may be inconsistent. In the
JouveElhannani and Hornych relationships, this
inconsistence is remedied by introducing a constant
to render the stress term nondimensional. This
measure is more a technical solution and does not
change the bases or rationale of the models.
The nonlinear regression algorithm provides a
numerical optimization to determine the best t. It
should be noted that there might exist several local
minima of the squared error, of which the results in
Table 4 Statistical comparison of kh and Uzan models. Coefcient of determination, sum of squared errors and test of the extra
sum of squares (F-ratio and probability) are shown
Grading kh Uzan Extra sum of squares test
R
2
SSE R
2
SSE F P-value
08 0.96 0.032 0.99 0.0047 133 0.00*
05 0.96 0.019 0.99 0.0047 70 0.00*
04 0.97 0.012 0.99 0.0055 27 0.00*
03 0.93 0.015 0.94 0.012 4.1 0.06
*Statistically signicant (P < 0.05)
Materials and Structures (2008) 41:509525 519
Table 6 represent one minimum. Furthermore, esti-
mated parameters might depend on the initial values
chosen for the regression. Presented results were
obtained from nonlinear regressions using the GRG2-
algorithm (generalized reduced gradient) imple-
mented in Excel (Microsoft Corp.), by minimizing
the squared error between measured and tted strains.
The goodness-of-t for the nonlinear regression
tted using the different models are illustrated in
Fig. 9. The diagrams show tted volumetric and shear
strains versus measured strains as well as line of
equality.
The BrownPappin model shows the closest
adherence to the line of equality. However, in most
cases all models seem to generate reasonable esti-
mates of measured strains. Except for the Brown
Pappin model, negative volumetric strains seems to
be underestimated, while no obvious deviations of
shear strain can be distinguished. The goodness-of-t,
for the various models was also analyzed using the
Akaike information criterion statistic. Table 7 sum-
marizes the coefcient of determination (R
2
) and
differences in AIC
c
for the different models and
gradings, used.
The degree of explained variance (R
2
) is typically
fairly high. In all cases, the BrownPappin model
shows the highest R
2
-values. When comparing results
using AIC
c
, the model with lowest DAIC
c
is most
likely to be closest to the true mechanism i.e., the
larger the difference in AIC
c
the less empirical
support of a particular model compared to the best
performing model. Based on results presented in
Table 7, no distinct conclusions can be drawn; no
single model performs best for all materials studied.
To further explore the performance, the models are
ranked in order of increasing AIC
c
as shown in
Table 8.
To evaluate whether there is a statistically signif-
icant difference in the ranking, a Friedman test by
ranks was performed. In this case no statistically
signicant differences were indicated (P = 0.21).
As mentioned previously, results acquired using
the Hoff model could not be statistically compared to
the other models since the basis of the regression was
Fig. 7 Predicted strains using the Uzan (top), Boyce (mid) and
extended Uzan (bottom) models versus measured strains. Line
of equality is also shown
Table 5 Statistical comparison of estimation of shear and
volumetric strains using kh, Uzan and Boyce models. Coef-
cient of determination (R
2
), DAICc (compared to the Boyce
model) and P-value of the extra sum of squares F-test between
kh and Uzan models
Grading kh DAIC
c
Uzan DAIC
c
F-test (P-
value)
Boyce R
2
08 10 10 0.83
05 48 47 0.40 0.94
04 33 31 0.34 0.90
03 11 10 0.39 0.91
No reduction of SSE using the Uzan model
520 Materials and Structures (2008) 41:509525
different compared to the other models used. Never-
theless, the performance of the Hoff model can be
qualitatively assessed by visually comparing tted
results in Fig. 8, with the nonlinear regression
obtained using Eq. 23 as shown in Fig. 10.
It should be noted that the axes are reversed in
Fig. 10 compared to Fig. 8. This is because the
nonlinear regression using the Hoff model estimated
the stress levels required for a given strain level, in
contrast to the other models.
In similarity to the previous visual summaries of
model efciency (cf. Fig. 9), Fig. 11 shows tted
stresses using the Hoff model versus actually applied
stresses.
The individual R
2
-values for the different gradings
obtained by tting to the Hoff model were: 0.98
(grading 0.8), 0.92 (grading 0.5), 0.97 (grading 0.4)
and 0.73 (grading 0.3). It should be noted that the
nonlinear regression was performed using conditions
different to those used in for the other models thus
making a direct comparison less meaningful. How-
ever, the overall performance of the Hoff model is
considered fairly accurate. The scatter around the line
of equality is slightly smaller for the shear stress
compared to the mean normal stress.
Since equipment able to apply cyclic conning
pressure is considerably more complicated than
equipment using constant conning pressure, it would
be advantageous to be able to predict shear and
volumetric behavior when both stresses are cycled,
from triaxial tests using constant conning pressures.
The Boyce model parameters were estimated by
tting results obtained using constant conning
pressure. These parameters were subsequently used
to predict the behavior using cyclic conning
pressure. In Fig. 12, a comparison between predicted
and measured values is shown.
Overall, predicted values show poor agreement
with measured values, especially for the volumetric
strain. The result seems to be in agreement with
Karasahin et al. [23] who concluded, from perform-
ing a similar prediction, that model parameters
derived from constant conning pressure can only
approximately predict behavior when both stresses
are cycled.
4 Discussion and conclusions
Even for the comparably simple stress conditions
applied at cylindrically conned triaxial tests, it is
difcult to describe the resilient stressstrain rela-
tionship in a more general way. The comparably great
number of, more or less empirical, models found in
the literature indicates this complexity. In this
context, it should be remembered that the resilient
framework itself is a simplication of the complex
inelastic response of granular materials, including
e.g., nonlinearity and stress history dependence.
Furthermore, laboratory testing of coarse granular
Fig. 8 Comparison of
measured and tted strains
for different stress paths
(grading 05)
Materials and Structures (2008) 41:509525 521
materials is inherently difcult in respect to mini-
mizing the inuence of the highly particulate nature
of the larger aggregates. There is a practical conict
between minimizing the sample size and maintaining
approximate continuum conditions. Generally, on-
sample instrumentation is preferred to avoid end
effects close to the platens but buried anchors might
be more susceptible to particle effects compared to
the averaging effect plate-to-plate measurements
show. By determining radial strains from measure-
ments of circumferential length change, an averaging
effect is achieved. In addition, particle effects are
probably also inuenced by stress path, especially
when using cyclic conning pressure as the aggregate
matrix may deform in different ways under different
stress paths. Altogether, this poses difculties when
analyzing triaxial testing of granular materials, both
of a theoretical and technical nature.
An important feature of the models, based on shear
and volumetric components is that nonlinear regres-
sion is required to estimate the parameters. This
makes the outcome dependent on the algorithms used
and the initial parameter values needed for optimi-
zation. The nal solution might represent a local
maximum of the goodness-of-t. Sometimes, the
regression might even fail to provide a solution. In
Table 6 Parameters obtained for the different models and gradings
Grading Boyce, GK model
K
1
(kPa) G
1
(kPa) n
08 14.7 15.2 0.516
05 9.13 9.06 0.443
04 14.1 12.9 0.547
03 3.84 2.62 0.263
Brown and Pappin, contour model
A (MPa) B m n C D (kPa) r
08 3.70 0.0615 0.688 2.81 1.70 10
4
0.296 1.08
05 7.45 0.128 0.635 1.95 2.61 10
4
0.105 0.539
04 5.78 0.0290 0.646 3.72 5.56 10
4
0.187 0.330
03 90.0 0.0527 0.522 3.68 3.89 10
4
0.097 0.445
JouveElhannani
K
a
(kPa) G
a
(kPa) n b n p
a
(kPa)
08 130 155 0.667 0.108 2.09 10
7
100
05 116 122 0.538 0.111 3.52 10
7
100
04 109 120 0.648 0.121 1.23 10
7
100
03 110 91.7 0.468 0.213 1.88 10
7
100
Hornych et al.
K
a
(kPa) G
a
(kPa) n c p
a
(kPa)
08 118 103 0.622 0.805 100
05 108 94.1 0.538 0.845 100
04 101 82.4 0.618 0.849 100
03 104 63.1 0.378 0.828 100
Hoff et al.
K (MPa) G (MPa) D (GPa)
08 180 97.8 475
05 176 73.9 476
04 157 82.0 238
03 156 100.0 210
Stress input was in kPa
522 Materials and Structures (2008) 41:509525
contrast, parameters in the relationships used to
model resilient modulus (kh and Uzan, respectively)
can be estimated by linear regression.
Since the main objective of this investigation was
to evaluate the efciency of various resilient models,
no direct comparison between the different gradings
is performed. However, in Fig. 5 it can be observed
that testing the various gradings, at in principal equal
levels of matric suction, gave a variety of mechanic
response ranging from a fairly stiff material showing
a comparably low strain ratio to a considerably softer
and dilative material. In other words, the different
models were used to describe materials showing
markedly different resilient behavior.
Based on results and analyses presented in this
study the following conclusions can be drawn:
The Uzan model provided a statistically signi-
cant improvement to modeling resilient modulus
compared to the kh model, in three out of four
cases. However, if a constant Poisson ratio is used
to predict shear and volumetric strains, this
improvement is lost; the difference in regression
sum of squares is no longer statistically signi-
cant. In particular, the prediction of volumetric
strain is inadequate when using these models. The
goodness-of-t of tting shear and volumetric
strains is greatly improved by using a stress
dependent Poisson ratio.
Fig. 9 Comparison of
measured and tted strains
using the different models
Table 7 Statistical parameters for the various models and gradings
Grading Boyce BrownPappin JouveElhannani Hornych et al.
R
2
DAIC
c
R
2
DAIC
c
R
2
DAIC
c
R
2
DAIC
c
08 0.88 3 0.96 5 0.93 2 0.92 0
05 0.91 3 0.97 6 0.95 4 0.94 0
04 0.91 6 0.99 0 0.95 6 0.94 5
03 0.86 0 0.96 16 0.91 6 0.89 2
Table 8 Ranking of model performance for the different
gradings
Grading Boyce Brown
Pappin
Jouve
Elhannani
Hornych
et al.
08 3 4 2 1
05 2 4 3 1
04 3 1 4 2
03 1 4 3 2
Mean
rank
2.25 3.25 3.00 1.50
Materials and Structures (2008) 41:509525 523
When modeling shear and volumetric strains from
the cyclic conning pressure triaxial tests, no
single model was most likely to be correct, for all
the gradings tested. Although not statistically
signicant, the model proposed by Hornych et al.
showed the highest mean ranking.
A nonlinear regression method for estimating the
parameters of the model presented by Hoff et al.
[20] was proposed. However, this measure inval-
idates a direct statistical comparison of the
efciency of this model to the others.
When using parameters estimated by triaxial tests
at constant conning pressure for prediction of
mechanical behavior under cyclic conning
pressure, predicted and measured results showed
poor agreement.
References
1. Fuller WB (1905) Proportioning concrete. In: Taylor FW,
Thompson SE (eds) A treatise on concrete plain and
reinforced. John Wiley, pp 183215
2. ASTM D4253 (1996) Standard test methods for maximum
index density and unit weight of soils using a vibratory
table. ASTM, Philadelphia PA
3. Ekblad J, Isacsson U (2006) Inuence of water on resilient
properties of coarse granular materials. Road Mater
Pavement Des 7(3):369404
4. Andrei D, Witczak MW, Schwartz CW, Uzan J (2004)
Harmonized resilient test method for unbound pavement
materials. Transport Res Rec 1874:2937
5. Raad L, Figueroa JL (1980) Load response of transporta-
tion support systems. Transport Eng J ASCE
106(TE1):111128
6. Tutumluer E, Thompson MR (1997) Anisotropic modeling
of granular bases in exible pavements. Transport Res Rec
1577:1826
7. Tutumluer E, Seyhan U (1999) Laboratory determination
of anisotropic aggregate resilient moduli using an innova-
tive test device. Transport Res Rec 1687:1321
8. Adu-Osei A, Little DN, Lytton RL (2001) Cross-aniso-
tropic characterization of unbound granular materials.
Transport Res Rec 1757:8291
Fig. 10 Measured and tted results using the Hoff model
(grading 05)
Fig. 11 Comparison of applied and tted stress for the Hoff
model (all gradings)
Fig. 12 Predicted strain obtained by tting of the Boyce model
to constant conning pressure data versus measured strains
obtained using cyclic conning pressure stress paths
524 Materials and Structures (2008) 41:509525
9. Brown SF, Pell PS (1967) An experimental investigation of
the stresses, strains and deections in a layered pavement
structure subjected to dynamic loads. In: Proceedings,
second international conference structural design of asphalt
pavements, Ann Arbor, USA, pp 487504
10. Seed HB, Mitry FG, Monismith CL, Chan CK (1967)
Prediction of exible pavement deections from laboratory
repeated load test. NHCRP Report No. 35, National
Cooperative Highway Research Program
11. Hicks RG, Monismith CL (1971) Factors inuencing the
resilient properties of granular materials. Highway Res Rec
345:1531
12. May RW, Witczak MW (1981) Effective granular modulus
to model pavement responses. Transport Res Rec 810:19
13. Uzan J (1985) Characterization of granular material.
Transport Res Rec 1022:5259
14. Sweere GTH (1990) Unbound granular bases for roads.
Dissertation, University of Delft
15. Uzan J (1992) Resilient characterization of pavement
materials. Int J Numer Anal Meth Geomech 16:453459
16. Boyce HR (1980) A non-linear model for the elastic
behaviour of granular materials under repeated loading. In:
International symposium on soils under cyclic and tran-
sient loading, Swansea 711, January 1980, pp 285294
17. Brown SF, Pappin JW (1981) Analysis of pavements with
granular bases. Transport Res Rec 810:1723
18. Jouve P, Elhannani M (1994) Application des mode`les non
lineaires au calcul des chaussees souples. Bull liaison Labo
P et Ch 190:3955
19. Hornych P, Kazai A, Piau J-M (1998) Study of the resilient
behaviour of unbound granular materials. In: Nordal RS,
Refsdal G (eds) Proceedings of BCRA98, vol 3. Trond-
heim, pp 12771287
20. Hoff I, Nordal S, Nordal RS (1999) Constitutive model for
unbound granular materials based on hyperelasticity. In:
Proceedings of an international workshop on modelling
and advanced testing for unbound granular materials,
Lisbon, 2122 January 1999, pp 187196
21. Kutner MH, Nachtsheim CJ, Neter J, Li W (2005) Applied
linear statistical models. McGraw-Hill
22. Burnham KP, Anderson DR (2002) Model selection and
multi-model inference: a practical information-theoretic
approach. Springer-Verlag, New York
23. Karasahin M, Dawson AR, Holden JT (1993) Applicability
of resilient constitutive models of granular material for
unbound base layers. Transport Res Rec 1406:98107
Materials and Structures (2008) 41:509525 525