Sunteți pe pagina 1din 15

625

Evaluation of geotechnical property variability


Kok-Kwang Phoon and Fred H. Kulhawy

Abstract: To evaluate geotechnical variability on a general basis that will facilitate the use of reliability-based design
procedures, it is necessary to assess inherent soil variability, measurement error, and transformation uncertainty
separately. The inherent variability and measurement error are addressed in a companion paper, and transformation
uncertainty is addressed herein. A second-moment probabilistic approach is applied to combine these uncertainties
consistently based on the manner in which the design soil property is derived. The design properties considered in this
paper are undrained shear strength, effective stress friction angle, in situ horizontal stress coefficient, and Young’s
modulus. This paper concludes with specific guidelines on the typical coefficients of variation for these common design
soil properties as a function of the test type and the type of correlation used.

Key words: transformation uncertainty, undrained shear strength, friction angle, in situ horizontal stress coefficient,
Young’s modulus, geotechnical variability.
Résumé : Pour évaluer la variabilité géotechnique d’une manière générale permettant d’utiliser des procédures de
calcul basée sur la fiabilité, il est nécessaire de prendre en compte séparément la variabilité intrinsèque du sol,
l’erreur de mesure et l’incertitude de transformation. On a traité de la variabilité intrinsèque et de l’erreur de mesure
dans un article conjoint et l’incertitude de transformation fait l’objet de la présente communication. Puis on applique
une approche probabiliste d’ordre deux pour combiner ces incertitudes, en accord avec la façon dont on acquiert la
propriété du sol en question. Les propriétés étudiées dans cet article sont : (i) la résistance au cisaillement non drainé,
(ii) l’angle de frottement effectif, (iii) le coefficient horizontal de poussée des terres en place et (iv) le module
d’Young. On conclut par des recommandations sur les coefficients de variation typiques à employer pour des
paramètres courants de calcul, en fonction du type d’essai et de corrélation utilisés.
Mots clés : incertitude de transformation, résistance au cisaillement non drainé, angle de frottement, coefficient horizon-
tal de poussée des terres en place, module d’Young, variabilité géotechnique.
[Traduit par la Rédaction] Phoon and Kulhawy 639

Introduction A number of the soil property statistics reported in the


geotechnical literature have been determined from total vari-
Since the mid-1970s, numerous reliability-based design ability analyses that implicitly assume a uniform source of
(RBD) codes have been put into practice for routine struc- uncertainty. Clearly, these lumped statistics are only applica-
tural design (e.g., ACI 1983; BSI 1972; CSA 1974; NKB ble to the specific set of circumstances (site conditions, mea-
1978). However, the geotechnical design community has surement techniques, correlation models) for which the
been slow in assimilating this new design methodology. design soil properties were derived. To evaluate geotechnical
Part of the reason lies in the difficulty of assessing the vari- variability on a more general basis that will facilitate the use
ability of design soil properties that are needed for these of RBD procedures, it is necessary to assess the three pri-
new RBD procedures. Unlike the variability of manufac- mary component uncertainties separately and to develop a
tured materials used in structures, geotechnical variability systematic approach that can combine these uncertainties
is a complex attribute that results from many disparate consistently, based on the manner in which the design soil
sources of uncertainties. The three primary sources are in- property is derived. Statistics on inherent soil variability and
herent soil variability, measurement error, and transforma- measurement error were reviewed and summarized in the
tion uncertainty. The degree of uncertainty arising from companion paper of this two-part series (Phoon and
these sources generally depends on factors such as the vari- Kulhawy 1999). This paper first describes the evaluation of
ability of the soil profile at the site, the degree of equip- transformation uncertainty. Then the uncertainties in the de-
ment and procedural control maintained during testing, and sign soil properties are evaluated rationally by combining
the precision of the correlation model used to transform the the appropriate component uncertainties using a second-
test result measurement into the desired soil property. moment probabilistic approach. The paper concludes with

Received March 11, 1998. Accepted February 16, 1999.


K.-K. Phoon. Department of Civil Engineering, National University of Singapore, 10 Kent Ridge Crescent, Singapore 119260.
F.H. Kulhawy.1 School of Civil and Environmental Engineering, Hollister Hall, Cornell University, Ithaca,
NY 14853-3501, U.S.A.
1
Author to whom all correspondence should be addressed.

Can. Geotech. J. 36: 625–639 (1999) © 1999 NRC Canada


626 Can. Geotech. J. Vol. 36, 1999

Fig. 1. Probabilistic characterization of transformation model.

specific guidelines on the typical coefficients of variation Table 1. Second-moment statistics for DK (source: Kulhawy et
(COVs) for some common design soil properties as a func- al. 1992, p. 5-24).
tion of the type of test and the type of correlation used.
su test typea Mean DKb COV of DK (%)
CIUC 0.0789 35
Transformation uncertainty UU 0.0512 29
VST 0.0906 40
The direct measurement from a geotechnical test typically a
CIUC, consolidated isotropic undrained triaxial compression test; UU,
is not directly applicable to design. Instead, a transformation unconsolidated–undrained triaxial compression test; VST, vane shear test.
model is needed to relate the test measurement to an appro- b
su/ σ vo = DK(qT – σvo)/ σ vo.
priate design property. Some degree of uncertainty will be
introduced, because most transformation models are ob- in which qc is the cone tip resistance, a is the net area ratio,
tained by empirical data fitting. Transformation uncertainty and ubt is the pore-water stress behind the cone tip. Ideally,
still would be present even for theoretical relationships be- the transformation uncertainty should be evaluated so that it
cause of idealizations and simplifications in the theory. The does not include extraneous variabilities, such as inherent
data scatter about the transformation model can be quanti- soil variability and measurement error. This result can be
fied using probabilistic methods, as illustrated in Fig. 1. In achieved approximately in practice by using high-quality
this approach, the transformation model typically is evalu- data that satisfy some basic requirements, such as (i) keep-
ated using regression analyses. The spread of data about the ing the distances between corresponding pairs of in situ
regression curve is modeled as a zero-mean random variable property (qT) and soil property (su) measurements at a mini-
(ε). The standard deviation of ε (SDε) is an indicator of the mum to reduce inherent soil variability; and (ii) using the
magnitude of transformation uncertainty, as shown in Fig. 1. same cone type, method of obtaining qT, and test to measure
An example of a probabilistic transformation model is su, where possible, to reduce systematic measurement errors.
given below: In a comprehensive study of this type (Kulhawy et al. 1992),
it was found that considerable uncertainty still can be attrib-
su (q − σvo) (q − σvo)
[1] = DK T = (m DK + ε) T uted to the model slope (DK), despite efforts to minimize the
σvo σvo σvo influence of extraneous variabilities, as shown in Table 1.
This study demonstrated that transformation uncertainty is a
in which su is the undrained shear strength (design property); significant and independent source of geotechnical variabil-
qT is the corrected cone tip resistance (test measurement); ity that cannot simply be explained away as being the result
σvo and σvo are the effective and total overburden stresses, of inherent soil variability and measurement error.
respectively; DK is the uncertain model slope; mDK is the Transformation models for other design properties and
mean of DK; and ε is the zero-mean random variable repre- test measurements can be found in the geotechnical litera-
senting transformation uncertainty. The corrected cone tip ture, and many have been summarized by Kulhawy and
resistance, qT, is defined as Mayne (1990). Tables 2 and 3 summarize the availability of
these models in the literature for a range of soil properties
[2] qT = qc + (1 – a)ubt developed from theory, laboratory test measurements, or

© 1999 NRC Canada


Phoon and Kulhawy 627

Table 2. Available correlations for cohesive soils (source: Kulhawy and Mayne 1990, pp. J2–J3).

Laboratory or Field test correlation


Property category Soil property theory correlation SPT CPT CPTU PMT DMT VST
Basic characterization Classification × — × × — × —
Unit weight × — — — — — —
Consistency — × × — — — —
In situ stress Preconsolidation stress × × × × × × ×
Overconsolidation ratio × × — × — × ×
Coefficient of horizontal soil stress × × × × × × —
Strength Effective stress friction angle × — — — — — —
Undrained shear strength × × × × × × ×
Deformability Poisson’s ratio × — — — — — —
Young’s modulus × — — — × — —
Compression indices × — — — — — —
Constrained modulus × × × — — × —
Coefficient of consolidation × — — × — × —
Coefficient of secondary compression × — — — — — —
Permeability Hydraulic conductivity × — — — — — —
Note: CPT, cone penetration test; CPTU, piezocone test; DMT, dilatometer test; PMT, pressuremeter test; SPT, standard penetration test; VST, vane
shear test.

Table 3. Available correlations for cohesionless soils (source: Kulhawy and Mayne 1990, pp. J4–J5).

Laboratory or Field test correlation


Property category Soil property theory correlation SPT CPT CPTU PMT DMT
Basic characterization Classification × — × × — ×
Unit weight × — — — — —
Relative density — × × × — ×
In situ stress Coefficient of horizontal soil stress × — × — × ×
Strength Effective stress friction angle × × × — × ×
Deformability Poisson’s ratio × — — — — —
Young’s modulus × × — — × ×
Compression index × — — — — —
Constrained modulus × — × — — —
Subgrade modulus × — — — — —
Permeability Hydraulic conductivity × — — — — —
Liquefaction resistance Cyclic stress ratio — × × — — ×

field test measurements. However, the transformation uncer- drained strength, modulus, and liquefaction resistance. Al-
tainties associated with these models are seldom analyzed though these characteristics undoubtedly influence N indi-
with the same degree of rigor as that shown in eq. [1]. The rectly, it is too much to expect that they all (singly or
majority are empirical and do not contain sufficient informa- collectively) can be predicted reliably without incurring sig-
tion for probabilistic evaluation. Although the magnitudes of nificant uncertainties.
uncertainty in these empirical models are unknown, they are
likely to be as large as those indicated in Table 1, particu- Variability of design soil properties
larly for the case of empirical models where two (or more)
items are being linked together that are not directly related. From the above observations, it is clear that the uncer-
A good example is the standard penetration test (SPT) N tainty in a design soil property is a function of inherent soil
value noted in Tables 2 and 3. The N value is the dynamic variability, measurement error, and transformation uncer-
driving resistance for a particular type of sampler, yet it has tainty. These components can be combined consistently us-
been correlated with the soil consistency, relative density, ing the simple second-moment probabilistic approach
vertical and horizontal soil stress state, drained and un- described below (Phoon et al. 1995).

© 1999 NRC Canada


628 Can. Geotech. J. Vol. 36, 1999

• Assume that the design property (ξ d) is predicted from a [9a] Γ2(L) = 1 for L = δv
test measurement (ξ m) using the following generic probabil-
istic transformation model: δv
[9b] Γ 2 (L) = for L > δv
L
[3] ξ d = T(ξ m, ε)
in which T(C) is the transformation function, which might be in which δv is the vertical scale of fluctuation. Equation [9]
nonlinear; and ε is the transformation uncertainty. states that the variance reduction function decreases as the
length of the averaging interval increases. Therefore, the ef-
• Introduce inherent soil variability (w) and measurement
fect of averaging is to reduce the uncertainty associated with
error (e) into eq. [3] by substituting ξ m = t + w + e (refer to
inherent soil variability (SD2w ), as shown in eq. [8]. For the
companion paper for details):
case in which t and ∂T/∂w are general functions of depth, no
[4] ξ d = T(t + w + e, ε) simple results are available. However, it can be noted that
the requirements of constant t and ∂T/∂w can be satisfied in
in which t is the deterministic trend function. Note that the most cases because (i) t is approximately constant if the av-
mean of w, e, and ε is zero. eraging interval is not too large, and (ii) ∂T/∂w is a constant
• Linearize eq. [4] about the mean of (w, e, ε) using a first- for linear transformation models, which are commonly as-
order Taylor-series expansion (Benjamin and Cornell 1970): sumed in foundation design.
∂T ∂T ∂T The above general probabilistic approach can be applied
[5] ξ d ≈ T (t,0) + w +e +ε to determine the typical range of variability for some com-
∂w ( t , 0) ∂e ( t , 0) ∂ε ( t, 0) mon design soil properties. The design properties considered
herein are undrained shear strength, effective stress friction
• Estimate the mean and variance of ξ d by applying second- angle, in situ horizontal stress coefficient, and Young’s
moment probabilistic techniques to eq. [5] (Benjamin and modulus. The compilation of statistical data on inherent soil
Cornell 1970): variability and measurement error presented in the compan-
[6a] m ξ d ≈ T(t,0) ion paper serves as realistic inputs for the evaluation of
eqs. [6] and [8]. However, eq. [8] only can be approximately
2 2 2
 ∂T   ∂T   ∂T  evaluated for many cases, because the vertical scale of fluc-
[6b] SDξ2d ≈   SDw2 +   SDe2 +   SD2ε tuation of most test measurements is not available. In addi-
 ∂w   ∂e   ∂ε  tion, rigorous statistics on transformation uncertainty
in which mξd and SD2ξ d are the mean and variance of ξ d, re- generally are not available, as noted previously. However, a
spectively; SD2w is the variance of inherent soil variability; first-order estimate of SDε can be obtained by noting that
SD2e is the variance of measurement error; and SD2ε is the about two thirds of the data typically fall within ± one stan-
variance of transformation uncertainty. dard deviation of the transformation model, because ε is ap-
The approximate mean and variance given in eq. [6] con- proximately normally distributed (Fig. 1). Unless the data
stitute the second-moment statistics of the design property at population has been screened carefully to remove extraneous
a point in the soil mass. For foundation design, it is not un- variabilities as noted above, significant improvement on this
common to evaluate the spatial average of the design prop- first-order estimate is unlikely, even if more sophisticated
erty over some depth interval, rather than use the value of statistical techniques were employed. Even with this simple
the design property at a point. The spatial average of ξ d is technique of evaluating transformation uncertainty, only a
defined as limited number of correlations could be examined because
most are presented without the supporting data, thereby
1 eliminating the possibility of assessing the transformation
L ∫L
[7] ξa= ξ d (z) dz uncertainty.

in which ξ a is the spatial average, L is the averaging length, Undrained shear strength
and z is the depth.
In principle, the mean and variance of ξ a can be computed For direct laboratory measurement of the undrained shear
by substituting eq. [5] into eq. [7] and applying second- strength (su), eq. [4] simplifies to
moment probabilistic techniques to the resulting equation. If
t and ∂T/∂w (or ∂ξ d /∂w) are constants, it can be shown that [10] ξ d = su = t + w + e
the mean of the spatial average is the same as that given in The mean and variance for this design property (su) are de-
eq. [6a], and the variance of the spatial average (SD2ξ a ) is termined from eq. [6] as follows:
given by (Vanmarcke 1983)
2 2 2 [11a] m ξd = t
 ∂T  2  ∂T   ∂T 
[8] SD 2ξ a ≈   Γ ( L) SD 2w +   SD 2e +   SD 2ε
 ∂w   ∂e   ∂ε  [11b] SD2ξ d = SD2w + SD2e

in which Γ2(C) is the variance reduction function, which de- Therefore, the COV of ξ d, which is defined by SDξd/mξd, is
pends on the length of the averaging interval, L. The follow- given by
ing approximate variance reduction function has been
proposed for practical applications (Vanmarcke 1983): [12] COVξ2d = COVw2 + COVe2

© 1999 NRC Canada


Phoon and Kulhawy 629

in which COVw is the COV of inherent variability = SDw/t, (eq. [9]). By substituting this variance reduction into
and COVe is the COV of measurement error = SDe/t. The eq. [17], the corresponding COV for the spatial average
COV of the spatial average (ξa) can be obtained from eq. [8] (COVξa) would be between 14 and 47%.
in a similar manner:
Correlation with corrected cone tip resistance
[13] COVξ2a = Γ2(L) COVw2 + COVe2 The correlation between the undrained shear strength and
The COVs of inherent variability (COVw) for the uncon- the corrected cone tip resistance was given in eq. [1]. This
fined compression test (UC), unconsolidated–undrained correlation model can be expressed in the form of eq. [4] as
triaxial compression test (UU), and consolidated isotropic follows:
undrained triaxial compression test (CIUC) were estimated [18] ξ d = su = (mDK + ε)(t + w + e – σvo)
in the companion paper to be between 20 and 55%, 10 and
30%, and 20 and 40%, respectively. The COV of measure- in which (t + w + e) = qT. The mean and variance for this de-
ment error (COVe) for undrained strength tests was esti- sign property are determined from eq. [6] as follows:
mated to be between 5 and 15%. The typical vertical scale
[19a] mξd ≈ mDK (t – σvo)
of fluctuation for su was judged to be on the order of 1–2 m.
Using these numerical data, the total COVs (COVξd) for di- [19b] SD2ξ d ≈ m DK
2
(SD2w + SD2e ) + (t – σvo)2 SD2ε
rect measurement of su using UC, UU, and CIUC tests are
21–57%, 11–34%, and 21–43%, respectively (eq. [12]). For The COV of ξ d is given by
an averaging length of 5 m, the amount of variance reduc- (COVw2 + COVe2 )
tion [Γ2(L)] would vary from 0.2 to 0.4 (eq. [9]). By substi- [20] COVξ2d ≈ 2
+ COVε2
tuting this variance reduction into eq. [13], the  σvo 
1 − 
corresponding COVs for the spatial averages (COVξa) of the  t 
three tests would be 10–38%, 7–24%, and 10–29%.
in which COVε is the COV of transformation uncertainty =
Correction from vane shear test SDε/mDK. The COV of the spatial average can be obtained
An undrained shear strength also can be estimated using from eq. [8] in a similar manner:
the in situ vane shear test (VST). However, a correction fac- [ Γ 2 (L)COVw2 + COVe2 ]
tor is needed to account for strain-rate effects and soil aniso- [21] COVξ2a ≈ 2
+ COVε2
tropy (e.g., Kulhawy and Mayne 1990). For this case, eq. [4]  σvm 
1 − 
can be expressed as follows (e.g., Baecher and Ladd 1985):  t 
[14] ξ d = (mµ + ε) (t + w + e) in which σvm is the average total overburden stress over L.
in which ξd is the design property, which is the corrected su The COVs of transformation uncertainty (COVε) for
from the vane shear test; mµ is the mean correction factor; ε CIUC and UU reference tests were found to be 35 and 29%,
is the uncertainty in the correction factor; and (t + w + e) = respectively (Table 1). The COV of inherent variability
su(VST). The mean and variance for this design property are (COVw) for the corrected cone tip resistance (qT) in clay was
determined from eq. [6] as follows: estimated in the companion paper to be less than 20%. The
measurement error (COVe) for qT likely would be compara-
[15a] m ξd ≈ mµ t ble to that produced by the electric cone penetration test,
which was estimated to be between 5 and 15%. The vertical
[15b] SD2ξ d ≈ mµ2 (SD2w + SD2e ) + t2SD2ε scale of fluctuation for qT in clay was found to be less than
The COV of ξd is given by 0.5 m. To evaluate eqs. [20] and [21], it also is necessary to
determine the ratios σvo/t and σvm/t. Note that t is equal to
[16] COVξ2d = COVw2 + COVe2 + COVε2 the mean value of qT, because the mean of w and e is zero.
From Fig. A1a in the companion paper, the mean value of
in which COVε is the COV of transformation uncertainty = qT (= t) in clay is on the order of 2 MN/m2. For a typical to-
SDε /mµ . The COV of the spatial average can be obtained tal soil unit weight of 20 kN/m3 and a depth of 10 m, the to-
from eq. [8] in a similar manner: tal overburden stress (σvo or σvm) is 200 kN/m2. Therefore,
[17] COVξ2a = Γ2(L) COVw2 + COVe2 + COVε2 the ratios σvo/t and σvm/t are on the order of 0.1. Using these
numerical data, the total COVs (COVξd) for su(CIUC) and
The COV of transformation uncertainty (COVε) for soft su(UU) predicted from qT would be between 35 and 47%
clays was estimated to be between 7.5 and 15% (Baecher and 30 and 40%, respectively (eq. [20]). For an averaging
and Ladd 1985). The COV of inherent variability (COVw) length of 5 m, the amount of variance reduction [Γ2(L)]
and measurement error (COVe) for the vane shear test were would be less than 0.1 (eq. [9]). By substituting this variance
estimated in the companion paper to be between 10 and 40% reduction into eq. [21], the corresponding COVs for the spa-
and 10 and 20%, respectively. The vertical scale of fluctua- tial averages (COVξa) of the two reference tests would be
tion for su(VST) was judged to vary between 2 and 6 m. between 35 and 39% and 30 and 34%.
Using these numerical data, the total COV (COVξd) for the
corrected su from the vane shear test is between 16 and 47% Correlation with SPT N value
(eq. [16]). For an averaging length of 5 m, the amount of Reasonable correlation between the undrained shear
variance reduction [Γ2(L)] would vary from 0.4 to 1.0 strength and the SPT N value can be achieved, provided the

© 1999 NRC Canada


630 Can. Geotech. J. Vol. 36, 1999

Fig. 2. Relationship between su and SPT N value (source: Hara et al. 1974, p. 9). OCR, overconsolidation ratio; r, correlation coefficient.

correlation is restricted to one type of geology (Kulhawy [27] COVξ2a ≈ 0.722 [Γ2(L) COVw2 + COVe2 ]
and Mayne 1990). An example of such a local correlation is
illustrated in Fig. 2, which is applicable to alluvial clays in + (loge10)2 SD2ε
Japan. The probabilistic equation for this correlation is
The standard deviation of the transformation uncertainty
[22] log10(su/pa) = log10(0.29) + 0.72 log10(N) + ε (SDε) was not reported, but about two thirds of the data fall
within ±0.15 of the transformation equation shown in Fig. 2.
in which su is the undrained shear strength from UU tests,
Therefore, a first-order estimate of SDε is 0.15. The COV of
and pa is the atmospheric pressure. Equation [22] can be re-
inherent variability (COVw) and measurement error (COVe)
written in an equivalent exponential form:
for N were estimated in the companion paper to be between
[23] su/pa = 0.29N 0.7210 ε 25 and 50% and 15 and 45%, respectively. The vertical scale
of fluctuation for N in clays is not available. However, a
Equation [23] then is expressed in the form of eq. [4] as fol- comparison between the inherent variability of N in sands
lows: and clays showed that the effect of soil type was not signifi-
cant (Fig. A1c in the companion paper). Therefore, it is pos-
[24] ξ d = su = 0.29pa(t + w + e)0.7210 ε sible that the vertical scale of fluctuation for N is similar for
both soil types. The vertical scale of fluctuation for N in
in which (t + w + e) = N. The mean and variance for this de- sandy soils was reported as 2.4 m in one study (Vanmarcke
sign property are determined from eq. [6] as follows: 1977). Using these numerical data, the total COV (COVξa)
[25a] mξd ≈ 0.29t 0.72 for su(UU) predicted from N is between 40 and 60%
(eq. [26]). For an averaging length of 5 m, the amount of
variance reduction [Γ2(L)] would be about 0.5 (eq. [9]). By
[25b] SD2ξ d ≈ m ξd
2
[0.722 (COVw2 + COVe2 ) substituting this variance reduction into eq. [27], the corre-
+ (loge10)2 SD2ε ] sponding COV for the spatial average (COVξa) would be be-
tween 38 and 54%.
The COV of ξ d is given by
[26] COVξ2d ≈ 0.722 (COVw2 + COVe2 ) + (loge10)2 SD2ε Correlation with DMT horizontal stress index
The undrained shear strength also can be correlated to the
The COV of the spatial average (ξa) can be obtained from dilatometer test (DMT) horizontal stress index, KD. One ex-
eq. [8] in a similar manner: ample of such a correlation for Italian clays is illustrated in

© 1999 NRC Canada


Phoon and Kulhawy 631

Fig. 3. Relationship between su and KD from DMT (source: clays, which were estimated to lie between 10 and 35%. The
Marchetti 1980, p. 317). ID, dilatometer material index. measurement error (COVe) for KD could be assumed to be
the same as that produced by the dilatometer test, which was
estimated to be between 5 and 15%. Using these numerical
data, the total COV (COVξd) for su determined from the
dilatometer test would be between 29 and 53% (eq. [30]).
The vertical scale of fluctuation for KD is not available, but
it is likely to lie between 2 and 6 m, as discussed in the
companion paper. For an averaging length of 5 m, the
amount of variance reduction [Γ2(L)] would vary from 0.4 to
1.0 (eq. [9]). By substituting this variance reduction into
eq. [31], the corresponding COV for the spatial average
(COVξa) would be between 27 and 53%.

Correlation with plasticity index


The undrained shear strength determined from the vane
shear test can be correlated to the plasticity index (PI) as fol-
lows (e.g., Chandler 1988):
su (VST)
[32] = 0.11 + 0.0037 PI
σp

in which σp is the preconsolidation stress. The transforma-


tion uncertainty was not given, but the author noted that the
accuracy of eq. [32] would be on the order of ±25%. For de-
sign, a correction factor should be applied to su(VST), as
Fig. 3. The recommended correlation equation is (Marchetti noted previously. The COV of the correction factor also was
1980) noted as lying between 7.5 and 15%. In addition, the
preconsolidation stress is an uncertain parameter. The total
 su  variability of σp was found to lie between 12 and 39% in
[28]   = 0.22(0.5KD)1.25
σ  one study (Vanmarcke and Fuleihan 1975). A simple proba-
 vo  DMT bilistic model for the above correlation is
Note that the strength data have not been corrected to the [33] ξ d = mσp (1 + ε) [0.11 + 0.0037 (t + w + e)]
same reference test. Also, the mean line and the standard de-
viation about the mean were not reported by the author. For in which ξ d is the design property, which is the corrected su
subsequent probabilistic analysis, the following model was from the vane shear test; mσp is the mean preconsolidation
determined using linear regression analysis: stress; (t + w + e) = PI; and ε is a random variable that ac-
counts for the inaccuracy in eq. [32], uncertainty in the VST
 s 
[29] log10  u  = log10 (0.12) + 1.24 log10 KD + ε correction factor, and uncertainty in the preconsolidation
 σvo  DMT stress. The mean and variance for this design property are
determined from eq. [6] as follows:
The standard deviation of the transformation uncertainty [34a] mξd ≈ mσp (0.11 + 0.0037t)
(SDε ) was found to be 0.11. Equations [29] and [22] are
similar, and therefore the COV for this case can be obtained
by direct comparison with eq. [26]: [34b] SD2ξ d ≈ m ξ2d 
(
 SD2 + SD2
w e
+
)
SD 2

ε
 (29.73 + t) 2 
 
[30] COVξ2d ≈ 1.242 (COVw2 + COVe2 ) + (loge10)2 SD2ε
The COV of ξ d is given by
Similarly, the COV of the spatial average (ξa) can be ob-
tained by direct comparison with eq. [27]:
[35] COVξ2d ≈
(SD 2
w + SDe2 ) + SD 2
ε
[31] COVξ2a ≈ 1.242 [Γ2(L) COVw2 + COVe2 ] (29.73 + t) 2

+ (loge10)2 SD2ε
From eq. [33], it can be seen that ∂ξ d/∂w = 0.0037mσp,
The COV of inherent variability (COVw) for KD in sands which is not a constant, because the mean preconsolidation
was found to lie between 20 and 60%. However, a compari- stress (mσp) generally varies with depth. As mentioned pre-
son between the COVs of dilatometer A and B readings re- viously, there are no simple solutions for the COV of the
vealed that the effect of soil type is significant (Fig. A1d in spatial average (COVξa) in such cases.
the companion paper). In the absence of relevant statistics, The standard deviation of inherent variability (SDw) and
the COV of inherent variability for KD in clays could be as- measurement error (SDe) of PI were estimated in the com-
sumed to be comparable to those for the A and B readings in panion paper to be between 3 and 12% and 2 and 6%, re-

© 1999 NRC Canada


632 Can. Geotech. J. Vol. 36, 1999

spectively. The lower and upper bounds on SDε can be  t/ p 


estimated from the uncertainty in eq. [32] (≈ 0.25), the un- [40a] m ξ d ≈ 17.6 + 11.0 log10  a 
 σvo/ pa 
certainty in the VST correction factor (≈ 0.08–0.15), and the  
uncertainty in σp (≈ 0.10–0.40), as follows:
[40b] SD2ξ d ≈ 22.8 (COVw2 + COVe2 ) + SD2ε
[36a] SDε ≥ 0.25 + 0.08 + 0.10 = 0.28
2 2 2
The COV of ξd is given by
[36b] SDε ≤ 0.252 + 0.152 + 0.402 = 0.50 22.8 (COVw2 + COVe2 ) + SDε2
[41] COVξ2d ≈
m ξ2d
Using these numerical data and assuming a typical mean PI
(= t) of 30%, the total COV (COVξd) for the corrected su The COV of the spatial average (ξ a) can be obtained from
from the vane shear test is between 29 and 55% (eq. [35]). eq. [8] in a similar manner:
22.8[Γ 2(L)COVw2 + COVe2 ] + SD2ε
Effective stress friction angle [42] COVξ2a ≈
m ξ2d
For direct laboratory measurement of the effective stress
friction angle (φ), eq. [4] simplifies to The COV of inherent variability (COVw) for qc in sand
was estimated in the companion paper to be between 20 and
[37] ξd = φ = t + w + e 60%. The measurement error (COVe) for qc can be assumed
to be comparable to that produced by the electric cone pene-
Equation [37] is the same as eq. [10]. Therefore, the mean, tration test, which was estimated to be between 5 and 15%.
variance, and COV of the design property (φ) can be deter- The vertical scale of fluctuation for qc was found to be less
mined from eqs. [11a], [11b], and [12], respectively. Simi- than 1 m. The mean friction angle (mξd) for most soils is be-
larly, the COV of the spatial average can be determined tween 20° and 40°. A typical value for mξd is 30°. Using
using eq. [13]. these numerical data, the total COV (COVξd) of φ (TC) de-
The COV of inherent variability (COVw) and measure- termined from qc is between 10 and 14% (eq. [41]). For an
ment error (COVe) for typical mean values of φ were both averaging length of 5 m, the amount of variance reduction
estimated in the companion paper to be between 5 and 15%. [Γ2(L)] would be less than 0.2 (eq. [9]). By substituting this
Using these numerical data, the total COV (COVξd) for di- variance reduction into eq. [42], the corresponding COV for
rect determination of φ is between 7 and 21% (eq. [12]). The the spatial average (COVξa) would be between 9 and 11%.
vertical scale of fluctuation for φ is not available, but it is
likely to lie between 2 and 6 m, as discussed in the compan- Correlation with plasticity index
ion paper. For an averaging length of 5 m, the amount of The constant-volume effective stress friction angle (φcv) of
variance reduction [Γ2(L)] would vary from 0.4 to 1.0 normally consolidated clays can be correlated to the plastic-
(eq. [9]). By substituting this variance reduction into ity index (PI) as follows (e.g., Mitchell 1976):
eq. [13], the corresponding COV for the spatial average
(COVξa) would be between 6 and 21%. [43] sin φcv = 0.8 – 0.094 logePI
This correlation is illustrated in Fig. 4. A simple probabilis-
Correlation with cone tip resistance tic equation that accounts for the spread of data about this
A number of correlations between the triaxial compres- correlation is as follows:
sion effective stress friction angle [φ (TC)] and the cone tip
resistance (qc) have been summarized (Kulhawy and Mayne [44] sin φcv = 0.8 – 0.094 logePI + ε
1990). The following correlation for sands is selected for
study, because the transformation uncertainty is available for Equation [44] can be expressed in the form of eq. [4] as fol-
subsequent probabilistic analysis: lows:

 q /p  [45] ξ d = φcv = sin–1[0.8 – 0.094 loge(t + w + e) + ε]


[38] φ (TC) = 17.6 + 11.0 log10  c a +ε
 σvo/ pa  (degrees)
 
in which (t + w + e) = PI. The mean and variance for this de-
The standard deviation of ε (SDε) for this correlation is 2.8° sign property (φcv) can be determined from eq. [6] as fol-
(Kulhawy and Mayne 1990). Equation [38] can be expressed lows:
in the form of eq. [4] as follows:
[46a] mξd ≈ sin–1(0.8 – 0.094 loget)
 (t + w + e)/ p 
[39] ξ d = φ (TC) = 17.6 + 11.0 log10  a+ε (degrees)
 σ / p 
 vo a 
 180   0.0942 (COVw2 + COVe2 ) + SDε2 
2
[46b] SDξ2 d ≈    
in which (t + w + e) = qc. The mean and variance for this  π   1 − (0.8 − 0.094 loget) 2 
design property [φ (TC)] can be determined from eq. [6] as
follows: (degrees)2

© 1999 NRC Canada


Phoon and Kulhawy 633

Fig. 4. Relationship between φcv and PI for normally consolidated clays (source: Mitchell 1976, p. 284).

The COV of ξ d simply is given by SDξ d/m ξ d. The COV of larly, the COV of the spatial average can be determined us-
the spatial average (ξ a) can be obtained from eq. [8] in a ing eq. [13].
similar manner: The COV of inherent variability (COVw) for the in situ
2
horizontal stress is not available, but it can be assumed to be
180   0.0942 [ Γ 2 ( L )COVw2 + COVe2 ] + SD2ε  comparable to the corresponding COV for the pressuremeter
[47] COVξ2a ≈    
 π   m ξ2d [1 − (0.8 − 0.094 log e t)2 ]  limit stress, which has been estimated in the companion pa-
per to be between 10 and 35% for clay and between 20 and
The standard deviation of the transformation uncertainty 50% for sand. The measurement error (COVe) for Ko can be
(SDε) was not reported, but it can be seen that about two assumed to be comparable to that produced by the self-
thirds of the data fall within ±0.07 of the transformation boring pressuremeter test, which was estimated to be be-
equation shown in Fig. 4. Therefore, a first-order estimate of tween 15 and 25%. Using these numerical data, the total
SDε is 0.07. The standard deviation of inherent variability COVs (COVξd) for direct determination of Ko using the self-
(SDw) and measurement error (SDe) for PI were previously boring pressuremeter in clay and sand are between 18 and
noted to be between 3 and 12% and 2 and 6%, respectively. 43% and 25 and 56%, respectively (eq. [12]). The vertical
Using these numerical data and assuming a typical mean PI scale of fluctuation for Ko is not available, but it is likely to
(= t) of 30%, the total COV (COVξd) for φcv determined lie between 2 and 6 m, as discussed in the companion paper.
from PI is between 16 and 19% (eq. [46]). The vertical scale For an averaging length of 5 m, the amount of variance re-
of fluctuation for PI is not available, but it is likely to lie be- duction [Γ2(L)] would vary from 0.4 to 1.0 (eq. [9]). By sub-
tween 2 and 6 m, as discussed in the companion paper. For stituting this variance reduction into eq. [13], the
an averaging length of 5 m, the amount of variance reduc- corresponding COVs for the spatial average (COVξa) in clay
tion [Γ2(L)] would vary from 0.4 to 1.0 (eq. [9]). By substi- and sand would be between 16 and 43% and 20 and 56%,
tuting this variance reduction into eq. [47], the respectively.
corresponding COV for the spatial average (COVξa) would
be between 16 and 19%. Correlation with dilatometer horizontal stress index
The dilatometer test provides an indirect measurement of
Ko because of disturbance caused by the insertion of the
In situ horizontal stress coefficient
blade into the ground. Correlations between Ko from the
The self-boring pressuremeter test (SBPMT) provides a SBPMT and the dilatometer horizontal stress index (KD) are
direct measurement of the in situ horizontal stress. No corre- available for sand and clay. However, only the following cor-
lation is needed because the stress is measured directly, tak- relation for clay is selected for study, because the transfor-
ing into account equipment calibrations. The in situ mation uncertainty is available for subsequent probabilistic
horizontal stress is expressed commonly in terms of the co- analysis (Kulhawy and Mayne 1990):
efficient Ko, which is defined as the ratio of the effective in
situ horizontal stress to the corresponding vertical effective [49] Ko = 0.27KD + ε
stress. For this direct field measurement, eq. [4] simplifies to
The standard deviation of ε (SDε) for this correlation is 0.48
[48] ξ d = Ko = t + w + e (Kulhawy and Mayne 1990). Equation [49] can be expressed
Equation [48] is the same as eq. [10]. Therefore, the mean, in the form of eq. [4] as follows:
variance, and COV of the design property (Ko) can be deter-
mined from eqs. [11a], [11b], and [12], respectively. Simi- [50] ξ d = Ko = 0.27 (t + w + e) + ε

© 1999 NRC Canada


634 Can. Geotech. J. Vol. 36, 1999

in which (t + w + e) = KD. The mean and variance for this The COV of ξd is given by
design property (Ko) can be determined from eq. [6] as fol- 2
lows:  SD 
[57] COVξ2d ≈ COVw2 + COVe2 + ε
[51a] mξd ≈ 0.27t  m ξd 
 
[51b] SD2ξ d ≈ 0.272(SD2w + SD2e ) + SD2ε From eq. [55], it can be seen that ∂ξd/∂w = 0.073pa/ σvo,
which is not a constant, because the effective overburden
The COV of ξd is given by stress ( σvo) varies linearly with depth. As mentioned previ-
2 ously, there are no simple solutions for the COV of the spa-
 SD 
[52] COVξ2d ≈ COVw2 + COVe2 + ε tial average (COVξa) in such cases.
 m ξd  The COVs of inherent variability (COVw) and measure-
 
ment error (COVe) for N were estimated previously to be be-
The COV of the spatial average (ξa) can be obtained from tween 25 and 50% and 15 and 45%, respectively. Using
eq. [8] in a similar manner: these numerical data and assuming a typical mean Ko (mξd)
of 1.5, the total COV (COVξd) predicted from N is between
2
 SD  41 and 73% (eq. [57]).
[53] COVξ2a ≈Γ 2
(L)COVw2 + COVe2 + ε
 m ξd 
  Young’s modulus
The COV of inherent variability (COVw) for KD in clays is The pressuremeter test (PMT) provides a direct measure-
not available. As discussed previously, this COV might be ment of the horizontal modulus of soils. This modulus
comparable to those for the A and B readings in clay, which (EPMT) often is presumed to be roughly equivalent to the
were estimated to lie between 10 and 35%. The measure- Young’s modulus (E). For this direct field measurement of
ment error (COVe) for KD could be assumed to be the same the soil modulus, eq. [4] simplifies to
as that produced by the dilatometer test, which was esti-
mated to be between 5 and 15%. Using these numerical data [58] ξ d = EPMT = t + w + e
and assuming a typical mean Ko (mξd) of 1.5, the total COV Equation [58] is the same as eq. [10]. Therefore, the mean,
(COVξd) for Ko determined from a dilatometer test would be variance, and COV of the design property (EPMT) can be de-
between 34 and 50% (eq. [52]). As noted previously, the termined from eqs. [11a], [11b], and [12], respectively. Sim-
vertical scale of fluctuation for KD is not available, but it is ilarly, the COV of the spatial average can be determined
likely to lie between 2 and 6 m. For an averaging length of using eq. [13].
5 m, the amount of variance reduction [Γ2(L)] would vary The COV of inherent variability (COVw) for EPMT in sand
from 0.4 to 1.0 (eq. [9]). By substituting this variance reduc- was estimated in the companion paper to be between 15 and
tion into eq. [53], the corresponding COV for the spatial av- 65%. The measurement error (COVe) for EPMT can be as-
erage (COVξa) would be between 33 and 50%. sumed to be comparable to that produced by the
pressuremeter test, which was estimated to be between 10
Correlation with SPT N value and 20%. Using these numerical data, the total COV
The standard penetration test N value also is related indi- (COVξd) for direct determination of EPMT in sand using the
rectly to Ko, because it is a measurement of vertical penetra- pressuremeter is between 18 and 68% (eq. [12]). The verti-
tion. An example of a correlation between Ko and N for clay cal scale of fluctuation for EPMT is not available, but it is
is given below (Kulhawy and Mayne 1990): likely to lie between 2 and 6 m, as discussed in the compan-
0.073Npa ion paper. For an averaging length of 5 m, the amount of
[54] Ko = +ε variance reduction [Γ2(L)] would vary from 0.4 to 1.0
σvo
(eq. [9]). By substituting this variance reduction into
The standard deviation of ε (SDε) for this correlation is 0.43 eq. [13], the corresponding COV for the spatial average
(Kulhawy and Mayne 1990). Equation [54] can be expressed (COVξa) would be between 14 and 68%.
in the form of eq. [4] as follows:
Correlation with dilatometer modulus
0.073(t + w + e) pa The dilatometer test (DMT) also provides a direct mea-
[55] ξ d = Ko = +ε
σvo surement of soil modulus. The dilatometer modulus (ED) is
related to the Young’s modulus as follows:
in which (t + w + e) = N. The mean and variance for this de-
E
sign property (Ko) can be determined from eq. [6] as fol- [59] ED =
lows: (1 − v2 )
0.073t pa in which ν is Poisson’s ratio. For this direct field measure-
[56a] m ξd =
σvo ment of the soil modulus, eq. [4] simplifies to

2 [60] ξ d = ED = t + w + e
 0.073 pa 
[56b] SDξ2 d ≈   (SDw2 + SDe2 ) + SD2ε
 Equation [60] is the same as eq. [10]. Therefore, the mean,
 σvo  variance, and COV of the design property (ED) can be deter-

© 1999 NRC Canada


Phoon and Kulhawy 635

Fig. 5. Relationship between EPMT of clays and SPT N value (source: Ohya et al. 1982, p. 129).

mined from eqs. [11a], [11b], and [12], respectively. Simi- Similarly, the COV of the spatial average (ξa) can be ob-
larly, the COV of the spatial average can be determined tained by direct comparison with eq. [27]:
using eq. [13].
The COV of inherent variability (COVw) for ED in sand [63] COVξ2a ≈ 0.632 [Γ2(L) COVw2 + COVe2 ]
was estimated in the companion paper to be between 15 and
65%. The measurement error (COVe) for ED can be assumed + (loge10)2 SD2ε
to be comparable to that produced by the dilatometer test,
which was estimated to be between 5 and 15%. Using these The COV of inherent variability (COVw) and measure-
numerical data, the total COV (COVξd) for direct determina- ment error (COVe) for N were previously noted to be be-
tion of ED using the dilatometer test in sand is between 16 tween 25 and 50% and 15 and 45%, respectively. The
and 67% (eq. [12]). The vertical scale of fluctuation for ED vertical scale of fluctuation for N in sandy soils also was
is not available, but it is likely to lie between 2 and 6 m, as noted above as equal to 2.4 m. Using these numerical data,
discussed in the companion paper. For an averaging length the total COV (COVξd) for EPMT predicted from N is be-
of 5 m, the amount of variance reduction [Γ2(L)] would vary tween 87 and 95% (eq. [62]). For an averaging length of
from 0.4 to 1.0 (eq. [9]). By substituting this variance reduc- 5 m, and assuming that the vertical scales of fluctuation in
tion into eq. [13], the corresponding COV for the spatial av- clay and sand are comparable as before, the amount of vari-
erage (COVξa) would be between 11 and 67%. ance reduction [Γ2(L)] would be about 0.5 (eq. [9]). By sub-
stituting this variance reduction into eq. [63], the
Correlation with SPT N value corresponding COV for the spatial average (COVξa) would
Correlations between soil modulus and N are fairly weak, be between 86 and 93%. The large COV provides a quantita-
as shown in the example between EPMT and N in clays given tive measure of the significant inaccuracy in the correlation
in Fig. 5. The appropriate probabilistic equation for this cor- model given above.
relation is
Importance of local correlations for Young’s modulus
E 
[61] log10  PMT  = log10(19.3) + 0.63 log10N + ε Local correlations that are developed within a specific
 Pa  geologic setting generally are preferable to generalized
global correlations because they are significantly more accu-
From the data scatter in Fig. 5, a first-order estimate of the rate. An example of such a correlation between the
transformation uncertainty (SD ε) is 0.37. Equations [61] and dilatometer modulus (ED) and SPT N value is given in
[22] are similar. Therefore, the COV for this case can be ob- Fig. 6. The data in Fig. 6 were obtained in sandy silts of the
tained by direct comparison with eq. [26]: Piedmont geologic province, in and around the Washington,
D.C., area. A simple probabilistic equation for this correla-
[62] COVξ2d ≈ 0.632 (COVw2 + COVe2 ) + (loge10)2 SD2ε tion is

© 1999 NRC Canada


636 Can. Geotech. J. Vol. 36, 1999

Fig. 6. Relationship between ED and SPT N value of Piedmont sandy silts (source: Mayne and Frost 1989, p. 22).

E  Table 4. Typical load and resistance statistics for structures (data


[64] log10  D  = log10(22) + 0.82 log10N + ε from Ellingwood et al. 1980).
 Pa 
COV (%)
From the data scatter in Fig. 6, a first-order estimate of the
a
transformation uncertainty (SDε) is 0.12. Equations [64] and Load
[22] are similar. Therefore, the COV for this case can be ob- Dead 10
tained by direct comparison with eq. [26]: Live 25
Wind 37
[65] COVξ2d ≈ 0.822 (COVw2 + COVe2 ) + (loge10)2 SD2ε Snow 26
Similarly, the COV of the spatial average (ξa) can be ob- Earthquake 138
tained by direct comparison with eq. [27]: Resistances
Concrete
[66] COVξ2a ≈ 0.822 [Γ2(L) COVw2 + COVe2 ] Flexure beams 8–14
+ (loge10)2 SD2ε Short columns 12–16
Slender columns 12–17
The numerical data on inherent variability (COVw), measure- Shear beams 17–21
ment error (COVe), and vertical scale of fluctuation for N Steel
were noted above. The total COV (COVξd) for ED predicted Tension members 11
from N is between 37 and 62% (eq. [65]). For a variance re- Compact beams, uniform moment 13
duction [Γ2(L)] of 0.5, the COV for the spatial average Axially loaded columns 14
(COVξa) would be between 34 and 54% (eq. [66]). The sig- Beam-columns 15
nificant reduction in the COVs, in contrast to those obtained Aluminum
in the preceding section, clearly demonstrates that reason- Tension members 8
able correlations with N can be achieved, provided the corre- Beams 8–13
lation is restricted to one type of geology. Columns 8–14
Glue-laminated timber beams 18
Geotechnical versus structural variability a
Fifty-year maximum load effects.
The results presented in these two companion papers am-
ply illustrate the variable nature of the uncertainties in contrast, the uncertainties in structural resistances typically
geotechnical design properties. For example, the COV of the fall within a narrow range of 10–20% for a wide range of
undrained shear strength of clays can vary from 10 to 60%, materials (e.g., concrete, steel, aluminum) and resistance
depending on whether it is measured directly or correlated models (e.g., tension, flexure, shear), as shown in Table 4.
empirically with certain field measurements. It is overly sim- Note that the uncertainties in structural material properties
plistic to assign “typical” COVs for geotechnical design are even lower because the uncertainties in structural
properties without proper reference to some of the important resistances shown in Table 4 also include uncertainties aris-
factors discussed in this paper, such as the soil type, the ing from fabrication and modeling errors. The uncertainties
measurement process, and the transformation model used. In in structural loads generally depend on the source of the

© 1999 NRC Canada


Phoon and Kulhawy 637

Table 5. Approximate guidelines for design soil property variability (source: Phoon et al. 1995, p. 4-51).
Design Spatial average Correlation
propertya Testb Soil type Point COV (%) COVc (%) equation
su(UC) Direct (lab) Clay 20–55 10–40 —
su(UU) Direct (lab) Clay 10–35 7–25 —
su(CIUC) Direct (lab) Clay 20–45 10–30 —
su(field) VST Clay 15–50 15–50 14
su(UU) qT Clay 30–40d 30–35d 18
su(CIUC) qT Clay 35–50d 35–40d 18
su(UU) N Clay 40–60 40–55 23
sue KD Clay 30–55 30–55 29
su(field) PI Clay 30–55d — 32
φ Direct (lab) Clay, sand 7–20 6–20 —
φ (TC) qT Sand 10–15d 10d 38
φcv PI Clay 15–20d 15–20d 43
Ko Direct (SBPMT) Clay 20–45 15–45 —
Ko Direct (SBPMT) Sand 25–55 20–55 —
Ko KD Clay 35–50d 35–50d 49
Ko N Clay 40–75d — 54
EPMT Direct (PMT) Sand 20–70 15–70 —
ED Direct (DMT) Sand 15–70 10–70 —
EPMT N Clay 85–95 85–95 61
ED N Silt 40–60 35–55 64
a
ED, dilatometer modulus; EPMT, pressuremeter modulus; Ko, in situ horizontal stress coefficient; su, undrained shear strength; su(field), corrected su from
vane shear test; φ, effective stress friction angle; φcv, constant-volume φ ; TC, triaxial compression; UC, unconfined compression test.
b
KD, dilatometer horizontal stress index; N, standard penetration test blow count; PI, plasticity index; qT, corrected cone tip resistance.
c
Averaging over 5 m.
d
COV is a function of the mean; refer to COV equations in the text for details.
e
Mixture of su from UU, UC, and VST.

loadings (e.g., dead, live, wind). Typical COVs can be as- Summary and conclusions
signed to each loading type as shown in Table 4. The COVs
are approximately 30%, with the exception of the nearly de- Rigorous statistics on transformation uncertainty generally
terministic dead loads (COV = 10%) and the highly variable are not available. A first-order estimate of the transformation
earthquake loads (COV = 138%). uncertainty can be obtained by noting that about two thirds
The differences between geotechnical and structural of the data typically fall within one standard deviation of the
variabilities have a significant impact on the development of transformation model. Even with this simple technique, only
reliability-based design procedures for geotechnical engi- a limited number of models could be examined, because
neering. As noted above, most of the COVs in structural de- most models have been presented without their supporting
sign are fairly small. Therefore, structural loads and data.
resistances usually can be modeled adequately as lumped Second-moment probabilistic techniques were applied to
parameters in the reliability calibration process. In addition, combine inherent soil variability, measurement error, and
the COVs in structural design either fall within narrow transformation uncertainty in a consistent manner. A sum-
ranges (e.g., structural resistances) or can be categorized mary of the variability of some design properties as a func-
readily into a few cases (e.g., loading type). It is therefore tion of the test measurement, correlation equation, and soil
relatively easy to obtain structural load and resistance factors type is presented in Table 5. The variability of the design
that are not dependent on COVs from the reliability calibra- properties is given in terms of the COV at a point and the
tion process. The situation in geotechnical engineering is COV for a spatial average. These ranges of COV are based
more complex. As discussed herein, there are many existing on representative statistics of inherent variability and mea-
means of evaluating the same design soil property. A practi- surement error, as described herein. More accurate COVs
cal reliability-based geotechnical design procedure will have can be calculated by substituting site-specific data on inher-
to account for the wide range of COVs resulting from the ent variability and measurement error into the closed-form
different evaluation methods and possibly high COVs from COV equations given in this paper. For example, the COV
highly variable site conditions, poor equipment and proce- for the undrained shear strength determined from the
dural control, and (or) low-quality correlations. It may no dilatometer test is given by eqs. [30] and [31]. However, it
longer be realistic to model geotechnical capacities as must be emphasized that these COV equations only are ap-
lumped parameters or to calibrate geotechnical resistance plicable to the specific correlations considered herein.
factors over the entire range of COVs. A more complete dis- The COVs of the undrained shear strength determined by
cussion of these important considerations is given elsewhere several different methods were found to be in the range of
(Phoon et al. 1995). 10–60%. For the undrained shear strength predicted from N,

© 1999 NRC Canada


638 Can. Geotech. J. Vol. 36, 1999

higher COVs develop when “universal” relationships are Kulhawy, F.H., Birgisson, B., and Grigoriu, M.D. 1992. Reliabil-
used that are not calibrated to a specific geology. The proba- ity-based foundation design for transmission line structures:
ble range of COV for the undrained shear strength is esti- transformation models for in situ tests. Electric Power Research
mated to be between 10 and 70%. The COV of the friction Institute, Palo Alto, Calif., Report EL-5507(4).
angle for sand and clay was found to be between 5 and 20%. Marchetti, S. 1980. In-situ tests by flat dilatometer. Journal of
For the in situ horizontal stress coefficient (Ko), the COV Geotechnical Engineering, ASCE, 106(GT3): 299–321.
was found to be in the range of 20–80% for clay, depending Mayne, P.W., and Frost, D.D. 1989. Dilatometer experience in
on the method of evaluation. The corresponding range of Washington, D.C., and vicinity. Research Record 1169, Trans-
portation Research Board, Washington, D.C., 16-23.
COV for sand, which was found to be in the range of 25–
Mitchell, J.K. 1976. Fundamentals of soil behavior. Wiley, New
55%, only would be applicable to the direct determination of
York.
Ko. The COV for indirect methods of evaluation could not
NKB. 1978. Recommendations for loading and safety regulations
evaluated because probabilistic transformation models were for structural design. Report 36, Nordic Committee on Building
not available. By comparing the range of COV for sand and Regulations (NKB), Copenhagen.
clay in the case of direct determination, the COV for sand Ohya, S., Imai, T., and Matsubara, M. 1982. Relationships between
generally is higher. It is possible that the COV for indirect N value by SPT and LLT pressuremeter results. In Proceedings
determination of Ko in sand also is higher than that for clay. of the 2nd European Symposium on Penetration Testing, Am-
Based on this consideration and the range of 20–80% for sterdam, Vol. 1, pp. 125–130.
clay as a reference, an overall COV of 30–90% is estimated Phoon, K.-K., and Kulhawy, F.H. 1999. Characterization of geo-
to be appropriate for both soil types. The COV of soil modu- technical variability. Canadian Geotechnical Journal, 36: 612–624.
lus was found to be highest. Even for direct methods of Phoon, K.-K., Kulhawy, F.H., and Grigoriu, M.D. 1995. Reliabil-
evaluation, the COV was found to be in the range of 20– ity-based design of foundations for transmission line structures.
70%. Higher COVs were obtained for correlations with N, Electric Power Research Institute, Palo Alto, Calif., Report TR-
particularly if the correlation is not restricted to a specific 105000.
geology. The probable range of COV for soil modulus also Vanmarcke, E.H. 1977. Probabilistic modeling of soil profiles.
is estimated to be on the order of 30–90%. Journal of the Geotechnical Engineering Division, ASCE,
103(GT11): 1227–1246.
Vanmarcke, E.H. 1983. Random fields: analysis and synthesis.
Acknowledgments MIT Press, Cambridge.
Vanmarcke, E.H., and Fuleihan, N.F. 1975. Probabilistic prediction
This research was supported, in part, by the Electric of levee settlements. In Proceedings of the 2nd International
Power Research Institute (EPRI) under RP1493. The EPRI Conference on Applications of Statistics and Probability in Soil
project manager was A. Hirany. and Structural Engineering, Aachen, Vol. 2, pp. 175–190.

References
ACI. 1983. Building code requirements for reinforced concrete. A: dilatometer A reading
ACI 318-83, American Concrete Institute (ACI), Detroit.
B: dilatometer B reading
Baecher, G.B., and Ladd, C.C. 1985. Reliability analysis of stabil-
COV: coefficient of variation
ity of embankments of soft clays. Special Summer Course 1.60s,
Massachusetts Institute of Technology, Cambridge, Mass.
COVe: coefficient of variation of measurement error
Benjamin, J.R., and Cornell, C.A. 1970. Probability, statistics, and COVw: coefficient of variation of inherent variability
decision for civil engineers. McGraw-Hill, New York. COVε: coefficient of variation of transformation uncertainty
BSI. 1972. Code of practice for structural use of concrete. CP110 COVξa: coefficient of variation of spatial average
(Pt. 1), British Standards Institution (BSI), London. COVξd: coefficient of variation of design soil property
Chandler, R.J. 1988. In-situ measurement of undrained shear DK: model slope for su and qT correlation
strength of clays using field vane. In Vane shear strength testing E: Young’s modulus
in soils: field and laboratory studies. American Society for ED: dilatometer modulus
Testing and Materials, Special Technical Publication 1014, EPMT: pressuremeter modulus
ASTM, Philadelphia, pp. 13–44. ID: dilatometer material index
CSA. 1974. Cold-formed steel structural members. Standard S136, KD: dilatometer horizontal stress index
Canadian Standards Association (CSA), Rexdale, Ont. Ko: in situ coefficient of horizontal soil stress
Ellingwood, B., Galambos, T.V., MacGregor, J.G., and Cornell, L: averaging length
C.A. 1980. Development of a probability-based load criterion N: standard penetration test value
for American National Standard A58 — Building code require-
OCR: overconsolidation ratio
ments for minimum design loads in buildings and other struc-
tures. Special Publication 577, National Bureau of Standards,
PI: plasticity index
Washington, D.C. SDe: standard deviation of measurement error
Hara, A., Ohta, T., Niwa, M., Tanaka, S., and Banno, T. 1974. SDw: standard deviation of inherent variability
Shear modulus and shear strength of cohesive soils. Soils and SDε: standard deviation of transformation uncertainty
Foundations, 14(3): 1–12. SDξa: standard deviation of spatial average
Kulhawy, F.H., and Mayne, P.W. 1990. Manual on estimating soil SDξd: standard deviation of design soil property
properties for foundation design. Electric Power Research Insti- T(C): transformation function
tute, Palo Alto, Calif., Report EL-6800. a: net area ratio

© 1999 NRC Canada


Phoon and Kulhawy 639

e: measurement error z: depth


mDK: mean of DK Γ2(C): variance reduction function
mµ : mean correction factor for vane shear test δv: vertical scale of fluctuation
mξd: mean design soil property ε: transformation uncertainty
mσp: mean preconsolidation stress ν: Poisson’s ratio
pa: atmospheric pressure ξa: spatial average
ξd: design soil property
qc: cone tip resistance
ξm: measured soil property
qT: corrected cone tip resistance φ: effective stress friction angle
r: correlation coefficient φcv: constant-volume effective stress friction angle
su: undrained shear strength σp: preconsolidation stress
t(C): trend function σvm: average total overburden stress over L
ubt: pore-water stress behind the cone tip σvo: total overburden stress
w: inherent soil variability σvo: effective overburden stress

© 1999 NRC Canada

S-ar putea să vă placă și