You are on page 1of 14

M O D E L UNCERTAINTY REPRESENTATION IN

GEOTECHNICAL RELIABILITY ANALYSES


By Knut O. Ronold 1 and Peter Bjerager, 2 Member ASCE

ABSTRACT: An adequate representation of model uncertainties is important for


Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

reliability analyses based on geotechnical analysis models of unknown credibility.


A procedure is outlined for calibration of the probability distribution function for
the model uncertainty associated with such a geotechnical model. The calibration
is based on available field test measurements of quantities, which can be predicted
by means of the actual model. The procedure is generally applicable, provided a
sufficient amount of field test measurements is available. The procedure is outlined
with specific reference to a prediction model for axial pile capacity and a set of
pile capacity measurements from field tests on full-scale piles. An example is pre-
sented to demonstrate the applicability of the procedure, based on the American
Petroleum Institutes's method for prediction of shear capacities of piles. A first-
order reliability analysis of an example pile is carried out and demonstrates the
significant importance of a best-possible representation of model uncertainty in
geotechnical reliability analyses.

INTRODUCTION

Conducting a limit state-based reliability analysis for a geotechnical en-


gineering problem involves specification of a limit state function. This func-
tion is usually based on a standard deterministic model for analysis of the
problem in question. The analysis model is associated with a model uncer-
tainty because of its imperfect representation of reality, e.g., due to sim-
plifications and idealizations that have been made, purposely or due to lack
of knowledge [see Ditlevsen (1981, 1982) and S0rensen (1984)].
Ditlevsen (1982) operates with subjective judgmental factors as one pos-
sibility for representation of such a model uncertainty. The model uncer-
tainty may be reflected in terms of a bias or a random variation in quantities,
such as capacities and deformations, which are predicted by the model when
applied to the considered problem. Recently, model uncertainty has been
dealt with by Tang (1989), who has approached the issue by statistical
analyses, and Der Kiureghian (1990), who has worked with consistent Bayes-
ian updating for assessment of model uncertainty.
As application of modern reliability methods to geotechnical engineering
problems becomes more frequent, there is an increasing need to know how
to properly represent model uncertainties in limit state functions where
uncertain engineering models are included. For geotechnical problems, the
models of interest are, first of all, models for prediction of capacities and
deformations for various types of foundations.
The model uncertainties can be represented by stochastic entities in the
reliability model. To establish a stochastic representation of the model un-
certainty associated with a particular analysis model, it is necessary to carry
out a series of predictions by the model (e.g., of a capacity, for a set of
^ e s . Fellow, Dept. of Civ. Engrg., Stanford Univ., Stanford, CA 94305-4020;
on leave, Prin. Res. Engr., Det norske Veritas, P.O. Box 300, N-1322 H0vik, Nor-
way.
2
Pres., Veritas Sesam Systems, P.O. Box 300, N-1322 H0vik, Norway.
Note. Discussion open until August 1,1992, To extend the closing date one month,
a written request must be filed with the ASCE Manager of Journals. The manuscript
for this paper was submitted for review and possible publication on Septemer 25,
1990. This paper is part of the Journal of Geotechnical Engineering, Vol. 118, No.
3, March, 1992. ASCE, ISSN 0733-9410/92/0003-0363/Sl.00 + $.15 per page. Paper
No. 569.
363

J. Geotech. Engrg. 1992.118:363-376.


foundations of the same basic shape or type), then compare these predictions
with known capacities for these foundations.
This paper demonstrates how the stochastic representation of model un-
certainty associated with a geotechnical analysis model can be established.
For reference, the prediction of axial pile capacities by means of one of
several available capacity prediction models is taken as an example. Such
a model is relevant for use in, for example, a reliability analysis for deter-
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

mination of the probability of pull-out for a tension pile.

THEORY FOR MODEL UNCERTAINTY AND ITS REPRESENTATION

The true, but unknown, pile capacity Z for a particular pile is sought.
The pile capacity is predicted by a selected prediction model to be X, a
stochastic variable that, first of all, depends on random soil strength prop-
erties. Once a realization of the soil strength properties is specified, an
outcome x of Xwill be achieved for the predicted pile capacity by the model.
x may deviate from the realized value z of the true capacity Z. This is due
to uncertainty and insufficiency in the model, e.g., due to simplifications
that have been made, such as neglecting field effects and discretizing by
neglecting some random quantities, e.g., geometry. The deviation between
the predicted and true capacities will be random, varying from case to case.
Based on this, the relationship between the true capacity Z and the predicted
capacity X can be written as

Z = / X (1)
where I = a random model uncertainty factor, describing the stochastic
functional dependence between the true and the predicted capacities.
/ is assumed to follow a probability distribution with mean value |x7 and
standard deviation 07. The mean value jx7 will be equal to 1.0 if the model
is expected to give an unbiased estimate of the true capacity Z. It will be
different from 1.0 if the model is biased and is expected not to give a correct
estimate for Zfor example, if the model is known to be conservative.
The mean value \L, can be a function of geometrical properties, such as the
pile length. This may be the case if the model is calibrated with measure-
ments from model tests rather than full-scale tests and scale effects are
known to prevail. The standard deviation 07 represents the variability in the
capacity predictions about the mean value |i,7. This is due to uncertainties
in the model, random variations in the predictions, etc., but may also reflect
scale effects, of which one has no control. It is of interest to establish realistic
statistics for /, i.e., estimates for |x7 and 07 that can be used together with
an assumed probability distribution type to represent / in a probabilistic
analysis. Capacity predictions Xk can be made for n piles, but the corre-
sponding true capacities Zk will be unknown, k = 1, . . . n. Information
about the true capacities Zk must be obtained by other means.
Full-scale measurements of pile capacities form a possible way to obtain
information about the true, but unknown, capacities Zk. This implies a
restriction on the number of piles and soil deposits that can be considered,
as one is limited to use available measurements from relatively few full-
scale tests performed on instrumented piles at given test sites. The use of
full-scale measurements in this context conforms to the fundamental sci-
entific principle, that mathematical models for strength and deformation
properties for structural elements are subjected to experimental tests in
364

J. Geotech. Engrg. 1992.118:363-376.


order to verify and support the empirical information value of the models
[see Ditlevsen and Madsen (1990)].
The measured capacity Yk for the /rth pile will be a stochastic variable
because it will be encumbered with uncertainties due to measurement errors
in load cells, etc. A realization yk of this measured capacity may deviate
from the true capacity Zk due to these measurement uncertainties. This may
be compensated for by multiplying the measured capacity Yk by a random
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

measurement uncertainty factor / , and the following relationship between


the true capacity Zk and the measured capacity Yk for the /cth pile can hence
be established
Zk = J Yk (2)
The probability distribution for / is unknown. This makes it difficult to
distinguish between distribution parameters for the model uncertainty factor
/ and those for J when pile capacity measurements and predictions are
compared to establish statistics for /. However, full-scale tests on instru-
mented piles are usually performed under well-controlled conditions with
no bias in the capacity measurements, and the measurements are usually
so accurate that the standard deviation of / can be assumed to be of a lower
order of magnitude than the standard deviation of /. This implies that it is
reasonable to neglect the uncertainties in the measured capacity Yk, re-
gardless of test type, and assume
Z~Yk (3)
For the /cth test pile, an outcome yk of the measured capacity Yk will be
observed, and an outcome xk of the predicted capacity Xk will result from
application of the selected prediction model for this pile. It is here assumed
that xk is obtained based on the mean properties of the soil strength as
estimated by statistical analysis of available test data. This assumption is
based on the fact that random field studies for pile capacity calculation
indicate that local variations of the soil strength about its mean will average
out over the exposed pile area and will not influence the resulting pile
capacity significantly [see Ronold (1990a)]. The uncertainty in the estimated
mean soil strength properties is usually much less than the variability in the
available soil strength observations. Provided that this uncertainty is of a
lower order of magnitude than that of /, use of the mean soil strength
properties to generate the outcome xk will therefore be accurate enough to
yield a ratio

(4)
'* = ?
xk
which can be considered as a realization of the model factor /for the capacity
prediction model.
Assume now that test data are available from n full-scale tests on instru-
mented piles, i.e., n outcomes yk of measured capacities, and n correspond-
ing outcomes xk of predicted capacities can be established by the capacity
prediction model. This will give n realizations of the model uncertainty
factor /. A statistical analysis of these n realizations will give information
about the probability distribution for /.
The mean value u,7 of / is estimated by
^--k--t^ (5)
n k=i n k=i xk
365

J. Geotech. Engrg. 1992.118:363-376.


The standard deviation 07 of / is estimated by

1
07 = (6)

There will hardly be sufficient data available to interpret the probability


Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

distribution type for /, but for the purpose of a probabilistic analysis based
on reliability methods it is necessary to assume a distribution type, and
normal and lognormal distributions are often the candidates for this. If a
normal distribution is assumed, the stochastic model uncertainty factor can
be expressed as
/ = M-/ Uu, (7)
where U = a standard normally distributed variable. This is based on the
assumption that the mean value \i, and the standard deviation 07 are fixed
numbers, u./ and a, are, however, estimated by statistical analysis of a limited
number of data, and statistical uncertainty in the estimates is present. It
can be shown that when the statistical uncertainty is included in the model
uncertainty factor, the model uncertainty factor can be expressed as

1
1= fi.,+ T 1 + (8)

where r_i is a Student's ^-distributed variable with n 1 degrees of


freedom [see Ditlevsen and Madsen (1990)].
When a lognormal distribution is assumed for /, the stochastic model
uncertainty factor can be expressed as
/ = exp((x; + U (T/) (9)
where

1
Wi = In ii, - - In (10a)

and

o-; = i In + 1 (10ft)

see Madsen et al. (1986). When statistical uncertainty is included in the


model uncertainty factor, the model uncertainty factor can be expressed as

1
/ = exp U ; + T^&ljl + (11)

where

|i; = In il, - - In 2 ) +1 (12a)


Pi,

and
366

J. Geotech. Engrg. 1992.118:363-376.


d; = \ In "'I + 1 (12b)

and r_! is a Student's ^-distributed variable with n 1 degrees of freedom,


as before.
These derivations indicate a need for measurement data from a series of
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

full-scale tests on piles, say at least four or five tests and preferably more,
and corresponding capacity predictions must be carried out for the respective
pile geometries and soil data by means of the considered capacity prediction
model, for which the model uncertainty factor / shall be established.
Scale effects will be present in the predicted capacities, when the true
capacity is known to depend to some extent on geometrical properties, such
as the pile length, and the prediction model does not account for this.
Without adequate knowledge of how scale effects influence the capacity of
a pile (and without ability to isolate these effects), scale effects will inevitably
become included in the model uncertainties, which are estimated based on
field test data obtained from measurements on piles of different lengths.
Ideally, the model uncertainty factor could be refined by letting the distri-
bution parameters |x and a be functions of the pile length, but sufficient
data to allow for such a refinement will very seldom be available.

IMPLEMENTATION IN RELIABILITY ANALYSES

The basis for the estimation of the mean value u.7 and standard deviation
07 of the model uncertainty factor / associated with a particular pile capacity
prediction method is formed by two parts, as described previously. The one
part consists of field measurements of the capacities of piles installed on
given test sites, whereas the other consists of predicted capacities for these
piles, calculated by means of the considered model and based on average
soil strength properties as estimated by regression analyses of available soil
strength observations from the respective test sites.
Once the statistics for the model uncertainty factor /have been established
for the considered capacity prediction model, and the input to this model
in terms of statistics for soil strengths, etc., has also been established, re-
liability analyses can be carried out by use of the model for one or more
pile geometries with specified environmental loadings at given sites.
Consider now just one pile at a given location. The soil profile at this
location is modeled with uncertainty in the soil strength properties. The
capacity of the pile can then be calculated by the model for any outcome
of these soil strength properties and can thereafter be multiplied by the
associated model uncertainty factor to account for the uncertainty involved
with the use of that particular model. If, as an example, the probability of
pull-out of the pile during its lifetime is sought, a simple limit state function
for this event can be established as
g(Z) = I-X- Qmax, (13a)
Z = (/, X, Qmax) (13ft)
This function fulfills the requirement that g(Z) > 0 for the safe set of Z,
and g(Z) < 0 for the failure set of Z. Z denotes the vector of stochastic
variables governing the loading and the capacity of the pile. It is noticed
that Z also includes the model uncertainty factor I. X = the capacity of the
pile, as estimated by the prediction model, and depends on one or more of
367

J. Geotech. Engrg. 1992.118:363-376.


the stochastic variables, including the soil strength variables. / = the model
uncertainty factor associated with the capacity prediction method, and <2max
= the maximum axial force at the pile head during the lifetime of the pile,
determined from some load model, which may be an environmental load
model if the pile is used offshore. X /, and QmdX are then all stochastic,
and / is represented by its estimator /, according to the expressions in (8)
or (11), depending on which distribution type is assumed for the model
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

uncertainty.
A reliability analysis can now be executed for the considered pile with
solution of the pull-out probability
PF = P[g (Z) < 0] (14)
by means of reliability methods, as described by Madsen et al. (1986) and
Bjerager (1989). These methods include first- and second-order reliability
methods as well as simulation methods, all of which are based on the use
of a limit state function as the one defined in (13).
It is of particular interest to note that the importance of the model un-
certainty factor for the applied capacity prediction model can be studied
from the results of a reliability analysis. Also, sensitivities in the analysis
results to changes in this model uncertainty factor and its underlying mean
and standard deviation can be assessed.
Since model uncertainty can be significant, and since it can form a con-
siderable contribution to the total sources of uncertainty associated with a
considered problem, the inclusion of a proper model uncertainty represen-
tation in a reliability analysis may be important. This is especially so if the
reliability analysis serves as a decision tool for structural design, in particular
for foundation design. On the other hand, problems may exist where the
model uncertainty does not play a dominating role at all, and the repre-
sentation of the model uncertainty in a reliability analysis may be left out
of consideration, when the insignificance has been verified through an anal-
ysis where the model uncertainty has been included and represented, e.g.,
as described earlier.

EXAMPLE

Example Pile Foundation


Axial loading of a vertical pile in a highly plastic, normally consolidated
clay is considered. The pile may be part of the foundation system for an
offshore tension leg platform, as indicated in Fig. 1. The length of the pile
is L = 35 m. The pile is closed-ended and tubular with diameter D =
0.219 m.
The loading at the pile head consists of some specified static pre-tension
force, which is superimposed by a cyclically varying force caused by wave
action in a storm. This loading is transferred by the pile to the supporting
soil by shear transfer at the surface of the pile. A possible pile tip resistance
is neglected. The maximum tensile force in the most severe storm during
the lifetime of the pile is considered to be the most critical load with respect
to pull-out and is therefore considered in the following.

Capacity Prediction Method


The American Petroleum Institute's method for prediction of shear ca-
pacities of piles is selected for reference [see "Recommended Practice"
(1984)]. This method is based on the assumption that the skin friction fs
368

J. Geotech. Engrg. 1992.118:363-376.


Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

FIG. 1. Axially Loaded Vertical Pile

between pile and soil can be represented as a fraction of the undrained


shear strength of the soil, as determined by vane tests. For the highly plastic,
normally consolidated clay considered here, this fraction is taken as / =
1.0, and hence the shear capacity of the pile is predicted by

Qf = irD Jo su dz (15)

where mean properties of the undrained shear strength su are to be inserted.


In principle, the total capacity of the pile against static pull-out consists
of the shear capacity, the weight of the pile, and a possible tip resistance.
Due to the closed end of the pile, there will be no soil plug, and the weight
of the pile will be counteracted by buoyancy and is therefore assumed not
to contribute to the tensile capacity. However, since the weight and buoy-
ancy are not expected to balance exactly, some uncertainty may be asso-
ciated with this assumption. This is neglected as minor in this context. A
tip resistance in terms of a possible suction is neglected as insignificant for
this very slender pile. This leaves the shear capacity as the only contributor
to the tensile capacity of the pile. When the shear capacity is predicted by
Qf, as calculated by (15), the tensile capacity of the pile becomes Z =
/ Qf, where the factor / represents the model uncertainty associated with
the prediction of Qf.

Reliability Method
The reliability of the example pile against pull-out is computed by a first-
order reliability method analysis. The input to the reliability analysis consists
of a limit state function in terms of a set of basic variables, comprising
stochastic variables as well as deterministic parameters. Furthermore, the
statistical distributions of the stochastic basic variables must be given and
the values of the deterministic parameters must be specified. The following
sections describe the stochastic basic variables, the deterministic parameters,
and the limit state function used.

Basic Variables
The uncertain variables involved in the analysis are described together
with their statistical distributions. Separate sections are devoted to envi-
369

J. Geotech. Engrg. 1992.118:363-376.


ronmental loading and response, soil strength properties, and model
uncertainty.

Environmental Loading and Response


Long-crested waves from one direction are considered and the Pierson-
Moskowitz spectral density is assumed for the wave energy
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

S = 4TT
m (2f exp
2TT
Tz.
w> 0 (16)

where Hs = the significant wave height in the most severe storm during
the lifetime of the pile; Tz = the corresponding mean zero-upcrossing
period; and w = the angular frequency. Tz relates to the spectral peak
period TP through

T (17)
"'ji
For a typical North Sea location, Hs follows an extreme value distribution,
and TP\HS is lognormally distributed
-j 1/1.385

H, = 0.5938 + { - T ^ T T In [1 - $(t/1)]r/r'ifc (18)


0.3174
TP = exp (U2 cr + |x) (19)
M21
(x = 1.187 + 0.8332HS (20)
a-2 = 0.1177 exp (-0.3123ff 5 ) + 0.005412 exp (-0.04122# 5 ) (21)
U1 and U2 = standard normally distributed variables; Tstorm = 6 hr is the
duration of the storm; and T]ite = 50 years is the design lifetime of the pile.
4>() = the standardized normal distribution function. Refer to Bitner-
Gregersen and Haver (1989).
The force transfer function HF(w) gives the ratio between the axial force
amplitude and the wave amplitude and is a function of the angular frequency
a) (see Fig. 2).
The y'th order moment of the cyclic axial force response is calculated as

-r Jo
ay' \HF(ca)\2 5^(w) dco

The force response process is assumed to be Gaussian and narrow-banded.


(22)

The maximum tensile force in the worst storm during the lifetime of the
pile is critical with respect to pull-out of the pile and can be found by standard
upcrossing considerations for a Gaussian process to be

2TT \
Qmax = i ~2kF0\n In *(/) (23)

see Madsen et al. (1986) and Ronold (1990b). FP = 150 kN = a permanent


static pre-tension force, which is considered fixed in the following with no
associated uncertainty, because it is usually well determined. The influence
of a possible change in FP is studied later. It is noticed that the expression
for the maximum force in (23) is based on linear wave theory, including
370

J. Geotech. Engrg. 1992.118:363-376.


H p ((u) (kN/m)
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

4.

_ s ^ <>

0 0.5 1.0 1.5 (rad/s)


FIG. 2. Transfer Function

only first-order wave loads in addition to the pretension. Wind, current,


and second-order wave loading are not included.

Soil Strength Properties


The undrained shear strength su of the soil is on an average well repre-
sented by a stepwise linear relationship with depth z
S
u al + a
2 Z (24a)
z = 0-10m (246)
su = b1 + b2 (z - 10) (25a)
z = 10 - 35 m (25b)
Linear regressions of 16 observations of su in the depth interval 0-10 m and
of 24 observations of su in the depth interval 10-35 m yield the following
estimates for mean values, standard deviations, and correlation coefficients
for a = (au a2)T and b = (br, b2)T

9.261
E[a] = (26a)
0.513

0.307
D[a] = (26b)
0.047

-0.9361 (26c)

E[b] = 14.33 (27a)


1.718
371

J. Geotech. Engrg. 1992.118:363-376.


0.279
D[6] = (27b)
0.029
Pii = -0.8202 (27c)
su is assumed to be in kPa, and z is given in meters.
In the reliability analysis the distributions of a and b are usedv Two-
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

dimensional lognormal distributions are assumed for both a and b. The


vectors a and b represent the average soil strength properties over the
available clay samples. If the samples are representative of the soil at the
pile, the average soil strength properties governing the pile capacity are
thus described. Effects of local strength variations from point to point along
the pile are assumed to average out over the length of the pile and are
therefore not included. Such effects could have been studied by introducing
a random field of the soil strength properties along the pile with inclusion
of a residual variance in the estimated strength properties [see Vanmarcke
(1977) and Ronold (1990a)].

Model Uncertainty Factor


Eleven static load tests on piles at the location of the example pile are
available and yield 11 measured capacities. The mean soil strength properties
at this location are given in (26) and (27) and are used to calculate 11
corresponding predicted capacities, according to (15). This gives 11 reali-
zations of the model uncertainty factor I associated with the API method.
A regression analysis of these 11 realizations leads to the following estimates
for the mean value and the standard deviation for /
A* = 0.801 (28a)
a, = 0.103 (28fo)
It is noted that these estimates correspond to a site-specific model uncer-
tainty factor /, because they are based on data from one particular location.
Some noise may be present in the estimated standard deviation due to scale
effects, because some of the test piles are considerably shorter than 35 m.
These effects have not been eliminated.
The mean value (L, = 0.801 indicates that the capacity prediction model
is biased when applied to piles on the considered location. The model
uncertainty factor I may be interpreted as the skin friction fraction / = fj
su, which should have been used rather than the prescribed value / = 1.0
when calculating pile capacities by the model for this location. Both (1/ and
o> are somewhat smaller than corresponding estimates reported by Tang
(1989). However, particularly for the standard deviation 67, this is not
unexpected, because Tang's data base is not site specific.
The model uncertainty factor / is assumed to be normally distributed.
This is based on a visual judgment of the available observations of /. How-
ever, the amount of data representing the tail of the distribution is very
small, if not negligible, so the assumption of the distribution type may still
be somewhat uncertain when it comes to the tail behavior and its importance
for the reliability calculations. With statistical uncertainty inherent in the
estimates for the distribution parameters a,, and 07, the model uncertainty
factor is represented in the reliability analysis by the estimator

/ = ( ! , + T^ &tJl+- (29)
n
372

J. Geotech. Engrg. 1992.118:363-376.


where r _ t is a Student's f-distributed variable with n \ degrees of
freedom and n = the number of observations of / used for estimation of
(1/ and 67. It is noted that the distribution of / is a transformed t distribution
with mean value ix, and a standard deviation which is considerably greater
than 67.
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

Limit State Function


The limit state function for pull-out of the pile when subjected to the
maximum tensile force in the lifetime of the pile can be defined as
g(Z) =I-Qf- Qmax (30)
where the predicted shear capacity Q / is calculated according to (15), the
maximum tensile force Qmax is calculated according Jo (23), and the model
uncertainty factor / is represented by its estimator /, is given by (29).

Results of Reliability Analysis


The probability of pull-out
PF = P[g(Z) < 0] (31)
is found by a first-order reliability method. The corresponding reliability
index (3 is related to the pull-out probability by
(3 = - * - 1 (JV) (32)
_1
where <I> () = the inverse of the standardized normal distribution function.
The results of the reliability analysis are presented in Table 1.
The design point u* = (3a gives the most likely values of the stochastic
basic variables at failure expressed in the transformed normal space [see
Madsen et al. (1986)]. The corresponding point in the space of basic variables
Z is shown in the fourth column of Table 1. The vector a is a unit vector
of so-called sensitivity factors a,, and a?, which is sometimes denoted as the

TABLE 1, Results of Reliability Analysis, Static Pull-Out of Example Pile

Design Sensitivity
Variable Symbol Distribution point z* factors a
(1) (2) (3) (4) (5)
(a) Loading
Significant wave height Hs Extreme value 14.09 m 0.0267
Peak period TP\HS Lognormal 15.94 s 0.0045
Inherent uncertainty in
max. force U0\HS, TP Normal 0.0870 0.0264
(b) Soil strength
Strength intercept i Lognormal 9.255 kPa -0.0011
Strength gradient a2k Lognormal 0.511 kPa/m -0.0010
Strength intercept 61 Lognormal 14.33 kPa 0.0010
Strength gradient b2\bt Lognormal 1.717 kPa/m -0.0065
(c) Capacity model
Model uncertainty factor I Transformed / 0.3045 -0.9993
Note: Reliability index ft = 3.300; probability of pull-out PF = 4.831 IP' 4 .

373

J. Geotech. Engrg. 1992.118:363-376.


ith importance factor, roughly gives the fraction of the total uncertainty
caused by the uncertainty in basic variable Z,.
The results in Table 1 show that the model uncertainty exclusively is the
most important source of uncertainty by a contribution to the total uncer-
tainty close to 100%. This clearly demonstrates how crucial model uncer-
tainty can be for a reliability analysis of this type, and how important it is
to represent it properly by a random factor in this analysis. In other words,
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

it is shown how little effort may be needed in establishing accurate models


of the uncertainty of the other variables in the limit state function.
The static pretension force FP is usually well-determined, with little or
no associated uncertainty. However, its value may change, mainly due to
changes in the deckload of the founded structure, and the influence of such
a change can be studied through the sensitivity

H = -0-0057 ...:..(33)

which is a byproduct of the reliability analysis. The value of this sensitivity


indicates that only moderate changes in the reliability index will take place
if the pre-tension should be subject to a change within an expected order
of magnitude, say 30%. This may reflect the little importance of modeling
accurately other variables than the model uncertainty factor.
Parametric sensitivities in the calculated reliability index (3 to changes in
the distribution parameters for the governing stochastic variables also are
a byproduct of the reliability analysis. The following sensitivities are pro-
duced to changes in the mean value and the standard deviation of the model
uncertainty factor

fr 3 - 98 w
S=~ 1 9 - 2 (**>
This indicates rather significant sensitivities in p to changes in the distri-
bution parameters for the model uncertainty factor, especially to changes
in the standard deviation.
Table 1 shows that the design point for the model uncertainty factor /
is a point on the tail of the distribution for / and hence has a rather low
value. An investigation shows that the reliability index (3 is very sensitive
to the choice of the distribution type for the model uncertainty factor, for
example if the underlying distribution of / is changed from a normal to a
lognormal distribution, as given by (9). This significance is known as the
tail sensitivity problem in structural reliability. The tail sensitivity problem
is generally believed to be less as regards the outcome of comparable studies
under consistent assumptions.
The importance of the model uncertainty seems to be almost indifferent
to the choice of the distribution type. Even if the significant statistical
uncertainty in / due to the very few observations of / is removed in the
analysis by using the normal distribution in (7) for / instead of the trans-
formed t distribution in (8), the importance remains very high.
The low tail value of the design point for / is a result of the exclusive
dominance of the model uncertainty among the uncertainty sources in com-
bination with its high standard deviation. It is demonstrated by the calcu-
374

J. Geotech. Engrg. 1992.118:363-376.


lation of the sensitivities in (3 to changes in the distribution parameters for
/ that much can be gained in terms of an improved reliability or a cheaper
foundation if a good control of the model uncertainty is achieved.
When the most important finding from the reliability analysis is that the
uncertainty in the model uncertainty factor is the overriding contributor to
the total uncertainty in the pull-out problem, one may conclude that the
uncertainties in the load and the soil strength properties could just as well
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

have been neglected in the analysis. That the importance of the model
uncertainty is so high is an expression of the fact that the physical under-
standing of the pull-out problem is very limited, and that the model, in this
case the API method, is crude and needs to be improved, at least for a
detailed reliability analysis.
No model uncertainty factor on the maximum pile force is included in
the analyses. The model uncertainty associated with this force can be sig-
nificant, for example because of the application of linear wave theory in
the load model. Also, model uncertainty in the scatter diagram underlying
the distribution of the significant wave height in (18) [see Winterstein et al.
(1989)], may be significant. The inclusion of a model uncertainty factor on
the maximum pile force in the analysis would most likely have turned the
reliability problem into a two-variable problem dominated by the two model
uncertainty factors for load and capacity, respectively, and the overriding
importance of the uncertainty in the capacity model would then have been
broken. However, the conclusion from the performed analyses with respect
to the dominating role of the uncertainty in the capacity model relative to
the uncertainties in the soil strength properties would still hold.

CONCLUSIONS

Model uncertainties can be associated with almost any engineering model


for prediction of capacities, stresses, deformations, etc., in structures or
structural elements, including foundations. An approach using a represen-
tation of such model uncertainties in terms of a stochastic model uncertainty
factor has been presented.
The distribution parameters for this stochastic model uncertainty factor
can be calibrated from sets of coherent measurement and prediction values
of quantities that can be predicted by means of a particular computational
or mechanical model. The calibration in the present study is based on the
assumption of no measurement uncertainty. The distribution parameters
often are encumbered with a significant statistical uncertainty. This can be
reduced by carrying out more measurements to form the background for
the calibration, if feasible.
For reference purposes, a model for prediction of axial pile capacities has
been considered, and the practical application of the approach has been
demonstrated by execution of an example with the pull-out of a vertical pile
in clay as the critical limit state. The example covered estimation of statistics
for a site-specific model uncertainty factor, and these statistics were then
used as part of the input to a subsequent reliability analysis of the considered
pile. The results of this reliability analysis showed that model uncertainty
by far is the most important source of uncertainty for this pull-out problem.
The general implication of this result is that model uncertainty can be a
very important source of uncertainty in a geotechnical engineering problem,
and that it is therefore important to include it in the reliability analysis of
such a problem. Then, in particular, it is important to represent the model
uncertainty by the best possible estimates of the distribution parameters.
375

J. Geotech. Engrg. 1992.118:363-376.


The presented approach to model uncertainty representation in a relia-
bility analysis is generally useful to account for the credibiility of the under-
lying model. However, to lead to applicable results it requires that the
applied model gives a good physical representation of the problem to be
analyzed, i.e., that model uncertainty is not completely dominant. The
approach may be useful for comparison of engineering analysis models and
Downloaded from ascelibrary.org by New York University on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

of designs by different models.

ACKNOWLEDGMENTS

This paper is based on work performed within the research program


Reliability of Marine Structures, supported by Det norske Veritas, Saga
Petroleum a.s., Statoil, and Conoco. This contribution is gratefully ac-
knowledged. The opinions expressed in the paper are those of the writers
and should not be construed as reflecting the views of the sponsoring
companies.

APPENDIX. REFERENCES

"Recommended practice for planning, designing, and constructing fixed offshore plat-
forms." (1984). APIRP2A, 15 Ed., American Petroleum Inst., Washington, D.C.
Bitner-Gregersen, E., and Haver, S. (1989). "Joint long term description of envi-
ronmental parameters for structural response calculation." Proc, Second Int.
Workshop on Wave Hindcasting and Forecasting, Vancouver, B.C., 1989.
Bjerager, P. (1989). "Probability computation methods in structural and mechanical
reliability." Computational mechanics ofprobabilistic and reliability analysis, W. K.
Liu and T. Belytschko, eds., Elme Press International, Lausanne, Switzerland.
Der Kiureghian, A. (1990). "Bayesian analysis of model uncertainty in structural
reliability." Proc, Third IFIP Working Conf. on Reliability and Optimization of
Structural Systems, Springer-Verlag, New York, N.Y.
Ditlevsen, O. (1981). Uncertainty modeling. McGraw-Hill, New York, N.Y.
Ditlevsen, O. (1982). "Model uncertainty in structural reliability." Structural safety,
1(1), 73-86.
Ditlevsen, O., and Madsen, H. O. (1990). "Baerende Konstruktioners Sikkerhed."
SBI Report No. 211, Statens Byggeforskningsinstitut, H0rsholm, Denmark (in
Danish).
Madsen, H. O., Krenk, S., and Lind, N. C. (1986). Methods of structural safety.
Prentice-Hall Inc., Englewood Cliffs, N.J.
Ronold, K. O. (1990a). "Random field modeling of foundation failure modes." /.
Geotech. Engrg., ASCE, 116(4), 554-570.
Ronold, K. O. (1990b). "Reliability analysis of a tension pile." /. Geotech. Engrg.,
ASCE, 116(5), 760-773.
S0rensen, J. D. (1984). "Paalidelighedsanalyse af statisk og dynamisk belastede
elasto-plastiske ramme- og gitterkonstruktioner." Structural Reliability Theory Pa-
per No. 4, Institute of Building Technology and Structural Engineering, University
of Aalborg, Denmark (in Danish).
Tang, W. H. (1989). "Uncertainties in offshore axial pile capacity." Foundation
engineering: Current principles and practices, Vol. II, F. Kulhawy, ed., ASCE,
New York, N.Y., 833-847.
Vanmarcke,E. (1977). "Probabilistic modeling of soil profiles."/. Geotech. Engrg.,
ASCE, 103(11), 1127-1246.
Winterstein, S. R., De, R. S., and Bjerager, P. (1989). "Correlated non-Gaussian
models in offshore structural reliability." Proc, Fifth Int. Conf. on Structural Safety
and Reliability, Vol. I, ASCE, New York, N.Y., 239-242.

376

J. Geotech. Engrg. 1992.118:363-376.

You might also like