Performance-Based Analysis and Design For Internal Stability of MSE Walls

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Georisk: Assessment and Management of Risk for

Engineered Systems and Geohazards

ISSN: 1749-9518 (Print) 1749-9526 (Online) Journal homepage: https://www.tandfonline.com/loi/ngrk20

Performance-based analysis and design for


internal stability of MSE walls

Richard J. Bathurst, Tony M. Allen, Yoshihisa Miyata, Sina Javankhoshdel &


Nezam Bozorgzadeh

To cite this article: Richard J. Bathurst, Tony M. Allen, Yoshihisa Miyata, Sina Javankhoshdel &
Nezam Bozorgzadeh (2019): Performance-based analysis and design for internal stability of MSE
walls, Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards

To link to this article: https://doi.org/10.1080/17499518.2019.1602879

Published online: 18 Apr 2019.

Submit your article to this journal

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=ngrk20
GEORISK
https://doi.org/10.1080/17499518.2019.1602879

Performance-based analysis and design for internal stability of MSE walls


Richard J. Bathursta, Tony M. Allenb, Yoshihisa Miyatac, Sina Javankhoshdeld and Nezam Bozorgzadeha
a
Department of Civil Engineering, GeoEngineering Center at Queen’s-RMC, Royal Military College of Canada, Kingston, ON, Canada;
b
Washington State Department of Transportation, State Materials Laboratory, Olympia, WA, USA; cDepartment of Civil and Environmental
Engineering, National Defense Academy, Yokosuka, Japan; dRocscience, Toronto, Ontario, Canada

ABSTRACT ARTICLE HISTORY


Mechanically stabilized earth (MSE) walls must be designed with adequate margins of safety against Received 22 October 2018
internal stability failure described by tensile strength and pullout limit states. Probabilities of failure Accepted 27 March 2019
(or reliability index) will vary strongly with the accuracy of the underlying models that appear in
KEYWORDS
limit state performance functions. In this paper, prior work by the authors and co-workers on this Mechanically stabilized earth
topic is reviewed. Relative performance is explored in the context of the reliability (or probability (MSE) walls; probabilistic
of failure) of steel and polymeric reinforcing elements in MSE walls using limit state performance analysis; reliability-based
functions with load and resistance models having different accuracy. design; tensile strength limit
state; pullout limit state
Abbreviations: AASHTO: American Association of State Highway and Transportation Officials; ASD:
allowable stress design; COV: coefficient of variation; CDF: cumulative distribution function; CSA:
Canadian Standards Association; FS: factor of safety; LRFD: load and resistance factor design; MC:
Monte Carlo (simulation); MSE: mechanically stabilized earth; OFS: operational factor of safety;
RBD: reliability-based design

1. Introduction
predictions for walls under operational conditions
Mechanically stabilized earth (MSE) walls constructed (e.g. BSI 2010; PWRC 2013; Allen and Bathurst 2015,
with steel or polymeric reinforcing elements have 2018; AASHTO 2017).
been in use for decades. These systems are essentially Models of varying accuracy have been used for the
gravity wall structures where the gravity mass is a com- internal stability analysis and design of MSE wall struc-
posite formed by soil backfill reinforced with horizontal tures using allowable stress design (ASD in USA termi-
layers of reinforcement and most often including a nology) where margins of safety are expressed by
hard concrete facing. The soil-reinforced mass must factors of safety, and using load and resistance factor
be internally stable over the life of the structure if the design (LRFD) in North American foundation engineer-
composite mass is to perform its primary function, ing practice (AASHTO 2017; CSA 2019). Both ASD and
which is to resist the external loads imposed on the LRFD for MSE walls do not provide an estimate of the
structure by the retained fill. probability that a limit state is satisfied. However, as geo-
The mechanisms of internal stability in the technical engineering moves to performance-based
reinforced soil zone are complex and difficult to formu- design, margins of safety expressed by probability of fail-
late if the objective is to capture the true mechanics of ure or reliability index have gained traction (e.g. Honjo
these systems because the component materials and and Kusakabe 2002; EN1997-1 2004; Honjo, Kikuchi,
their interactions are complicated. In order to make and Shirato 2010; Orr 2012; ISO 2394 2015; Phoon
analysis and design tractable for internal modes of fail- 2017). This is the motivation for the work by the writers
ure, such as tensile failure of the reinforcement layers and co-workers that is reviewed in this paper. This prior
and pullout of the reinforcement anchorage lengths, work is focused on the development and calibration of
simple concepts of classical earth pressure theory and more accurate models that can be used for perform-
interface shear resistance between dissimilar materials ance-based internal stability analysis and design of
have been adopted. However, in almost all cases, load MSE walls within a probabilistic framework, which is
and resistance model formulations have required also called reliability-based design (RBD) in the litera-
empirical adjustment to improve the accuracy of ture. The latter gives the designer the flexibility to

CONTACT Richard J. Bathurst bathurst-r@rmc.ca Department of Civil Engineering, GeoEngineering Center at Queen’s-RMC, Royal Military College of
Canada, Kingston, ON K7K 7B4, Canada
© 2019 Informa UK Limited, trading as Taylor & Francis Group
2 R. J. BATHURST ET AL.

compare between design options based on probabilistic laboratory test or from an in situ test performed in
estimates of reliability, and to adjust design options to the field. Nevertheless, for brevity in this paper, the
target levels of reliability (i.e. performance) selected by term bias is used.
the designer. Figure 1 shows the principal components of a MSE
wall and links them to the two limit states introduced
above. Details regarding the geometry of the partition
between the active soil zone and the passive zone in
2. Preliminaries
Figure 1 will vary depending on the type of reinforce-
Limit state performance functions used in this paper are ment (e.g. extensible polymeric type or inextensible
expressed as: steel). Similarly, there are a number of formulations to
Rm lR Rn compute the nominal value for the maximum reinforce-
g= −1= −1 (1) ment load in a reinforcement layer (Tmax) under oper-
Qm lQ Qn
ational conditions and the ultimate pullout capacity
Here Rm and Qm are observed (measured) resistance (Pc) of the same reinforcement material embedded in
and load values, respectively, and λR and λQ are the different soil materials.
corresponding bias values that are used to transform Formulations for Tmax for walls constructed with fric-
nominal values Rn and Qn to Rm and Qm. If nom- tional soils vary in complexity and accuracy (FHWA
inal and bias values are assumed to be random vari- 2009; Allen and Bathurst 2015, 2018; AASHTO 2017;
ables, then the probability Pf (g < 0) can be computed Bathurst and Yu 2018; Miyata and Bathurst 2019). All
using Monte Carlo simulation. If random values are formulations can be fully described by all or some of
lognormally distributed, a closed-form solution pro- the parameters in the following expression:
posed by Bathurst and Javankhoshdel (2017) is
available. Tmax = f (H, Ffs , Sv , z, f, sv , v, J) (2)
The structure of the limit state performance func-
tion introduced here is particularly attractive for Here H = height of wall, Φfs = facing stiffness, Sv = con-
internal stability limit state equations for MSE walls tributory area of the reinforcement layer, z = depth of
that are the focus of this paper. This is because the the layer below the top of the wall, f = friction angle
writers and co-workers have compiled large databases of the soil, and σv = vertical pressure acting at the
of measured reinforcement loads from instrumented reinforcement elevation due to soil self-weight (γ)
walls under operational (end of construction) con- and any surface distributed surcharge (q), ω = facing
ditions, and measured resistance values from labora- batter from the vertical, and J = reinforcement stiffness.
tory element tests used to quantify tensile strength The load values are expressed as force per unit length
and pullout capacity of a wide range of reinforcement of wall in the cross plane-strain direction. The reader
materials. Individual measured values can be paired is directed to the cited references for the details of
with predicted values using any of a large number the five different Tmax models used in this paper and
of load and resistance models that appear in design other models that are not included due to space
guidance documents and the journal literature. The constraints.
ratio of the measured value to the predicted (nom- For the calculation of pullout capacity (also in units of
inal) value is called method bias because the magni- force per length), equations having the following general
tude of a bias value is influenced by (a) the form are used:
intrinsic accuracy of the underlying deterministic
model to capture the mechanics of the limit state Pc = f (Nq , Le , sv , Rc ) (3)
load or resistance term, (b) the type of measurement
and its interpretation, and (c) the quality and quantity Here Le = anchorage length of the reinforcement layer in
of data (Allen, Nowak, and Bathurst 2005). In the the passive zone, Nq = dimensionless factor that depends
related literature the same parameter has been called on the reinforcement type, and Rc ≤ 1 is the reinforce-
(simply) bias or model factor (e.g. Phoon 2017). The ment layer coverage ratio. Reinforcement factor Nq is a
term method bias is preferred by the writers because function of soil friction angle, reinforcement type (exten-
the same model may be recommended for (say) pull- sible or inextensible) and reinforcement geometry. A
out of a particular class of soil reinforcement type, but number of pullout models can be found in design gui-
the measured value and hence the corresponding dance documents, design codes, and in the research lit-
bias value will vary depending on whether the erature (FHWA 2009; Huang and Bathurst 2009;
measured pullout value was gathered from a Miyata and Bathurst 2012; Yu and Bathurst 2015;
GEORISK: ASSESSMENT AND MANAGEMENT OF RISK FOR ENGINEERED SYSTEMS AND GEOHAZARDS 3

Figure 1. MSE wall components and nominal load and resistance terms.

AASHTO 2017; Miyata, Yu, and Bathurst 2018). In this following closed-form solution can be used to calculate
paper, only four pullout models are considered. reliability index (β) (Bathurst and Javankhoshdel 2017):
For the tensile strength limit state, the tensile strength
(Tal) of extensible polymeric reinforcement is expressed   
as: mlR mRn (1 + COV2Qn )(1 + COV2lQ )
ln
mlQ mQn (1 + COV2Rn )(1 + COV2lR )
Tult × Rc b=
Tal = (4)  ⎡ ln(1 + COV2 )(1 + COV2 )(1 + COV2 ) ⎤
RF  lQ
 ⎢ Qn Rn
 ⎢ ×(1 + COV2lR )(1 + rR COVRn COVlR )2 ⎥ ⎥
Where Tult = ultimate strength of the reinforcement and  ⎢ ⎥
 ⎢ ×(1 + rQ COVQn COVlQ )2 ⎥
RF = reduction factor that accounts for loss of strength ln⎢ ⎥

due to installation damage, creep and chemical durability  ⎢ (1 + rn COVRn COVQn )2 ⎥
 ⎢ ⎥
(e.g. FHWA 2009; AASHTO 2015; AASHTO 2017). For  ⎣ ⎦
steel reinforcement materials the corresponding
expression is: (6)
Ac Fy
Tal = (5)
b Here μRn and COVRn, μQn and COVQn, μλR and COVλR,
Here Ac = cross-section area, b = width of reinforce- and μλQ and COVλQ are mean and COV values for Rn,
ment element, and Fy = yield strength of the steel. This Qn, λR and λQ, respectively. Parameters ρR and ρQ are
calculation includes possible loss of section due to bias dependencies which are the Pearson correlation
corrosion. coefficients between Rn and λR and between Qn and λQ,
respectively. When ρR and ρQ are non-zero, the corre-
sponding resistance bias and load bias values vary with
the magnitude of the nominal values (Rn and Qn). Par-
3. Calculation of reliability index and
ameter ρn is the correlation between Rn and Qn and is
probability of failure
non-zero when equations for Rn and Qn share one or
The probability of failure [Pf = P(g < 0)] for the perform- more input parameters that are random variables. Here-
ance (limit state) function having the form shown in after, ρn is referred to as nominal correlation for simpli-
Equation (1) can be computed using Monte Carlo city. The conversion from β to Pf is made using the
(MC) simulation. For simple linear limit states with ran- approximation Pf = 1 − Φ(β) where Φ is the standard
dom variables described by lognormal distributions, the normal cumulative distribution function.
4 R. J. BATHURST ET AL.

The probability theory-based steps used to derive the 2016; CSA 2019). Bathurst, Javankhoshdel, and Allen
above equation for the single load case have been (2017) assigned values of COVRn and COVQn equal to
extended to the case of an additional permanent surface 0.10, 0.20 and 0.30 corresponding to high, typical and
traction representing a footing but with some additional low levels of understanding, respectively. These assign-
constraints. The interested reader is directed to the paper ments were necessary to quantify levels of understanding
by Javankhoshdel, Bathurst, and Cami (2018) for this for reliability theory-based LRFD calibration but are
case. equally applicable for reliability-based analysis and design
described in this paper. The exception to these values is
the assignment of COVRn = 0 for the nominal strength
4. Statistics of nominal values
of the geogrid or steel reinforcement that appears in the
Nominal values for resistance (Rn) and load (Qn) are tensile limit state function for these structures. This is
computed at time of design and are project specific. because there is no uncertainty in the calculation of the
They are computed using Equation (2) for load and nominal tensile strength at time of design which is
Equations (3)–(5) for resistance. A useful assumption taken as deterministic; rather, all uncertainty is captured
is that nominal values in each load and resistance term by the bias statistics for the tensile strength model.
in these equations are computed using mean values of
the input parameters that have some uncertainty (e.g.
5. Method bias statistics
soil unit weight (γ), soil friction angle (f), tensile
strength Fy and Tult). The mean and variance of these In this section we demonstrate example method bias
random variables can then be used to compute distri- statistics for load and resistance models that are judged
butions of Qn and Rn; from these distributions the to range from relatively poor to relatively good.
mean (μRn, μQn) and COV of nominal values can be
computed (COVRn, COVQn). Possible nominal corre-
5.1. Load
lations (ρn) between Qn and Rn for the pullout limit
state can be evaluated at the same time. They will vary Table 1 identifies five different load models examined in
between ±1 depending on the statistics for the random this paper. Figure 2 shows cumulative distribution func-
input variables that appear in the equations for Tmax tion (CDF) plots of bias values for one steel grid tensile
and Pc, and whether or not random values of component load model (LM2) and two different polymeric (geogrid)
material properties are taken from the same sample (Lin models (LM4 and LM5). The bias data were generated by
and Bathurst 2018). In the absence of evidence to the dividing each measured reinforcement load from instru-
contrary, the assumption of ρn = 0 simplifies the calcu- mented field walls by the corresponding predicted load
lation of Pf using MC simulation and the closed-form using the load model (i.e. load bias λQ = Qm/Qn). The
solution introduced earlier. In addition, taking ρn = 0 data have been plotted with a lognormal axis for bias.
has been recommended for practical LRFD calibration The data present as a straight line over the top half or
in the USA (Allen, Nowak, and Bathurst 2005). The two-thirds of the data indicating that the assumption
ratio of mean of nominal resistance (μRn) and mean of of lognormal distributed bias values is reasonable.
nominal load (μQn) is assumed as the conventional factor Approximations to the measured data are shown using
of safety used in ASD practice (i.e. FS = μRn/μQn). This the statistics for all data and visually fitting to the
assignment provides a link to the first step in design, upper tail of the distributions. The corresponding
which is typically to adjust the nominal resistance to mean and COV values that are required to match the
ensure that the design factor of safety for the limit visual fit are shown in the figures and in Table 1. It is
state is adequate based on past practice and experience. the upper portions of load bias distributions that strongly
A shortcoming of the approach described above to influence the magnitude of probability of failure of each
compute COVRn and COVQn is that it does not account limit state. Whether or not the approximations capture
for other sources of uncertainty beyond simply the varia- the lower tails is not important. Note that there are no
bility in soil and reinforcement properties. In LRFD foun- actual data corresponding to the fit to tail distribution
dation engineering practice in Canada, the notion of level that could be used to compute a correlation coefficient.
of understanding has been introduced to account for Hence, ρQ is taken as 0 when fit-to-upper tail bias stat-
uncertainty in the calculation of nominal values at time istics are used in reliability calculations.
of design. The motivation is to reward geotechnical design The plots in Figure 2 and the summary statistics in
engineers with larger resistance factors in limit state Table 1 show that load model LM2 underestimates
design equations as knowledge of project ground con- measured loads by about 22% on average (i.e. μλQ =
ditions and material properties improve (Fenton et al. 1.28) while load model LM4 overestimates measured
GEORISK: ASSESSMENT AND MANAGEMENT OF RISK FOR ENGINEERED SYSTEMS AND GEOHAZARDS 5

Table 1. Summary of bias statistics and bias dependency (cross-correlation) values for load models for MSE walls constructed with
granular soil.
Bias
Load Reinforcement Model Number of data Mean of bias COV of bias dependency Original measured data
model type reference points n mlQ COVlQ ρQ source
LM1 Steel grid AASHTO Simplified 81 all data 0.99 0.42 −0.18* Allen and Bathurst
Method fit to upper 1.15 0.26 ** (2015)
(FHWA 2009; AASHTO tail
2017)
LM2 Coherent Gravity 81 all data 1.28 0.45 −0.34 Allen and Bathurst
Method fit to upper 1.25 0.46 ** (2015)
(FHWA 2009; AASHTO tail
2017)
LM3 Simplified Stiffness 81 all data 0.99 0.34 −0.08 *
Allen and Bathurst
Method fit to upper 1.07 0.25 ** (2015)
(Allen and Bathurst tail
2015)
LM4 Polymeric AASHTO Simplified 95 all data 0.42 0.96 −0.40 Allen and Bathurst
(geogrid) Method fit to upper 0.42 1.10 ** (2015)
(FHWA 2009; AASHTO tail
2017)
LM5 Simplified Stiffness 95 all data 0.96 0.35 0.12* Allen and Bathurst
Method fit to upper 1.00 0.28 ** (2015)
(Allen and Bathurst tail
2015)
*ρQ = 0 at level of significance of 5%. **ρQ taken as 0 for the reason explained in the text.

loads by more than a factor of 2 on average (i.e. μλQ = 5.2. Pullout


0.42). The spread in bias values for load model LM2 cor-
Two pullout models for steel grids and two models for
responds to COVλQ = 0.45; the corresponding spread for
geogrid reinforcement products are identified in Table
bias values using load model LM4 is very much larger
2 together with computed bias values. Results of bias
(i.e. COVλQ = 0.96). The explanation for the large differ-
analyses of the two geogrid models (PM3 and PM4)
ences in COV of bias values can be found in the cited
are shown in Figure 3. The bias data for both models
sources. For example, load model LM5 considers the
are reasonably well approximated by a straight line in
important influence of reinforcement stiffness on the
the CDF plots indicating that they are lognormally dis-
tensile load generated in geogrid reinforcement layers
tributed from a practical point of view. Visual fits to
in walls under operational conditions, while load
the lower tail are also shown on the CDF plots and the
model LM4 does not.
corresponding mean and COV of bias values recorded.
Also shown in Figure 2 are plots of bias values versus
The motivation for fit-to-lower tail is that it is the distri-
predicted values of Tmax using the same models. Note
bution of bias resistance values at low bias values that
that the nominal values in these plots are deterministic
strongly influences reliability calculation outcomes. For
values computed at time of design and these single-valued
the same reason explained for the load bias data, ρR is
outputs do not have variability. Models LM2 and LM4
taken as 0 when fit-to-lower tail pullout bias statistics
show a statistically significant negative dependency
are used in reliability calculations.
between bias and calculated load values quantified by the
correlation coefficient (ρQ). This can be viewed as undesir-
able in ASD practice because it means that the accuracy of
5.3. Tensile strength
load predictions will vary with magnitude of predicted
load. Model LM5 has statistical bias characteristics that Tensile strength bias statistics are shown in Table 3. The
are judged to be better than load models LM2 and LM4. summary statistics for steel assume that the distribution
The mean of bias values for load model LM5 using all for yield strength is not truncated (e.g. fy = 450 MPa is
data is 0.96 (which is close to 1), there is less spread in the minimum specified yield strength recommended by
bias values (COVλQ = 0.35), and there is no bias depen- AASHTO (2017)). The initial zinc coating (galvaniza-
dency with predicted load Tmax. A further small improve- tion) is assumed to be thick enough to not be depleted
ment in bias statistics results if the visual fit to upper tail over the life of the structure. Hence, possible loss of sec-
is used (i.e. μλQ = 1.00, COVλQ = 0.28). Similar assessments tion due to corrosion over the design life is not con-
of the relative accuracy of the other two load models in this sidered in this study. These assumptions are made to
paper (LM1 and LM3) can be made by examining the bias simplify calculations and to allow the closed-form sol-
statistics summarized in Table 1. ution for β to be used. To consider the influence of the
6 R. J. BATHURST ET AL.

Figure 2. (a) Load model LM2 (steel grid), (c) Load model LM4 (geogrid) and (e) Load model LM5 (geogrid), Cumulative frequency
distribution (CDF) plots for load bias values and approximations to all bias data and fit to upper tail; (b) Load model LM2 (steel
grid), (d) Load model LM4 (geogrid) and (f) Load model LM5 (geogrid), Bias versus predicted (calculated) nominal load values (Tmax).
GEORISK: ASSESSMENT AND MANAGEMENT OF RISK FOR ENGINEERED SYSTEMS AND GEOHAZARDS 7

Table 2. Summary of bias statistics and bias dependency (cross-correlation) values for pullout models for MSE walls constructed with
granular soil.
Number of data Mean of Bias
Pullout Reinforcement points bias COV of bias dependency
model type Model reference n mlR COVlR ρR Original data source
PM1 Steel grid FHWA (2009) and 129 all data 1.33 0.44 −0.44 Miyata, Yu, and Bathurst (2018)
AASHTO (2017) fit to lower 0.83 0.18 **
tail
PM2 Yu and Bathurst (2015) all data 1.07 0.35 −0.10* Miyata, Yu, and Bathurst (2018)
PM3 Polymeric FHWA (2009) and 318 all data 2.23 0.55 −0.46 Huang and Bathurst (2009)
(geogrid) AASHTO (2017) fit to lower 1.10 0.27 **
tail
PM4 Huang and Bathurst 318 all data 1.07 0.36 0.03* Huang and Bathurst (2009)
(2009) fit to lower 0.95 0.36 **
tail
*ρR = 0 at level of significance of 5%. **ρR taken as 0 for the reason explained in the text.

two factors noted above on numerical outcomes requires for LRFD calibration of internal stability limit states for
MC simulation and a suitable corrosion model for the MSE walls (Allen, Nowak, and Bathurst 2005). The justifi-
zinc cover and steel (e.g. Allen, Bathurst, and Bozorgza- cation for what may appear to be a low margin of safety is
deh 2019). Regardless, the qualitative trends in this paper that these structures are very strength redundant with
are informative and quantitative outcomes are not respect to tensile and pullout failure. In other words, if
expected to be practically different from those using a one reinforcing element fails, the remaining elements
more sophisticated treatment for the steel grid cases. are present to compensate. An ideal outcome for
reliability-based design for the two limit states in this
paper is a nominal factor of safety that is acceptable
6. Results based on ASD and past practice (say, FS ∼ 1.5), and cor-
responds to a reliability index value β∼2.33. This outcome
6.1. General is difficult to achieve as shown in Figure 4 with the excep-
Computed results using Equation (6) with selected com- tion of two load and resistance model combinations
binations of load and resistance models are shown in (LM4-PM3 – Figure 4(f); LM4-TM2 – Figure 4(i)).
Figure 4 together with the assumption of ρn = 0 as dis-
cussed earlier. The results are presented as reliability
6.2. Influence of level of understanding
index β versus factor of safety FS and operational factor
of safety OFS. The former is the ratio of nominal resist- The plots in Figure 4 are presented for different coeffi-
ance to nominal load described earlier. The operational cients of variation of the nominal load and resistance
factor of safety is the factor of safety accounting for the values. For the pullout limit state examples, the plots
ratio of the average resistance bias to average load bias are presented with COVQn = COVRn = 0.10, 0.20 and
and appears as the first term in the numerator of 0.30. For the tensile strength limit state plots, COVRn
Equation (6), i.e.: = 0 was used in the calculations for the reasons explained
mlR m m earlier in the paper. For the same factor of safety or OFS
OFS = FS = lR Rn (7) in most plots, the reliability index increases with decreas-
mlQ mlQ mQn
ing COV of nominal values (e.g. Figures 4(a), (e), (g), (h)
If the nominal load and resistance models are perfectly and (i)). This outcome may be expected because
accurate on average (i.e. μλR = μλQ = 1), then FS = OFS. reliability is intuitively expected to increase with less
This is an unusual occurrence for soil-structure limit uncertainty in the choice of nominal values at the time
states. of design. However, there are exceptions such as Figure
Superimposed on the graphs in Figure 4 are values of 4(f) where the opposite trend appears. This is because
β = 2.33, 3.09 and 3.54, corresponding to probability of of the combined influence of the large negative bias
failure Pf = 1/100, 1/1000 and 1/5000, respectively. The dependencies and large COV of bias values for both
larger β values shown in the figures are typical target the load model (LM4 using all bias data) and the pullout
values used for LRFD calibration of structural elements model (PM3 using all bias data) (see Tables 1 and 2). The
in structural engineering design (β = 3.54) and for foot- same trend in the plots for Figure 4(f) can be seen in
ings in geotechnical engineering practice (β = 3.09) Figure 4(i) for the tensile strength limit state because of
(Allen, Nowak, and Bathurst 2005). A value of β = 2.33 the bias statistics for the LM4 load model. The decrease
has been recommended for reliability-based design and in β due to negative bias dependencies (i.e. negative
8 R. J. BATHURST ET AL.

Figure 3. (a) Pullout model PM3 and (c) Pullout model PM4, Cumulative frequency distribution (CDF) plots for pullout bias values and
approximations to all bias data and fit to lower tail; (b) Pullout model PM3 and (d) Pullout model PM4, Bias versus predicted (calculated)
nominal pullout capacity values (Pc).

values of cross-correlation coefficients ρQ and ρR) in 6.3. FS versus OFS


combination with increasing COVQn and COVRn can
Figure 4(f) also shows an example where OFS >> FS for
be appreciated by inspecting the denominator of
the same reliability index and all other parameters are
Equation (6) where the products of these parameters
the same. This is because the resistance model (PM3)
appear. The numerical outcomes shown in these plots
under-estimates reinforcement pullout capacity by a
are the same as those using MC simulations when the
large amount (on average) and thus is very conservative
same statistics are used.

Table 3. Summary of bias statistics and bias dependency values for tensile strength models for MSE walls constructed with granular soil.
Tensile Number of Mean of COV of Bias
strength Reinforcement Model data points bias bias dependency
model type reference n mlR COVlR ρR Original data source
TM1 Steel grid AASHTO N/A 1.10 0.07 0 Bathurst, Huang, and Allen (2011b), Nowak and
(2017) Szerszen (2003) and Nowak and Collins (2000)
TM2 Polymeric AASHTO N/A 1.10 0.10 0 Bathurst and Miyata (2015), Miyata and Bathurst
(geogrid) (2017) (2015), Miyata, Bathurst, and Allen (2014) and
Bathurst, Huang, and Allen (2011a, 2012)
GEORISK: ASSESSMENT AND MANAGEMENT OF RISK FOR ENGINEERED SYSTEMS AND GEOHAZARDS 9

Figure 4. Reliability index β versus FS and OFS for selected load and resistance models. (a) Models LM1-PM1 using all bias data; (b)
Models LM1-PM1 using fit to tails for model PM1 bias data; (c) Models LM2-PM1 using all bias data; (d) Models LM2-PM1 using fit to tails
for model PM1 bias data; (e) Models LM3-PM2 using all bias data; (f) Models LM4-PM3 using all bias data; (g) Models LM5-PM4 using all
bias data; (h) Models LM1-TM1 using all bias data; (i) Models LM4-TM2 using all bias data and (j) Models LM5-TM2 using all bias data.

(i.e. μλR = 2.23), and the load model (LM4) under-esti- nominal factor of safety. The OFS is 2.6 times the nom-
mates reinforcement tensile loads by a large margin inal factor of safety if the tensile limit state using the
(on average) (i.e. μλQ = 0.42); thus it is also very conser- same load model (LM4) is combined with the tensile
vative. The combined effect is that the actual factor of strength model for geogrid reinforcement (TM2), as
safety (i.e. OFS) is about five times larger than the shown in Figure 4(i). The opposite trend can occur
10 R. J. BATHURST ET AL.

when the ratio μλR/μλQ is less than one. For example, coefficient between nominal values is then computed in
Figures 4(b) and (d) show that the nominal FS is greater the usual manner.
than the OFS by a factor of 1.19 and 1.54, respectively
In some cases there is spread in the load and resist- 7. Conclusions
ance bias values, but the ratio μλR/μλQ is close to one
and thus OFS ∼ FS when all other parameters remain This paper examines performance assessment of MSE
the same; an example is shown in Figure 4(c). However, walls with respect to pullout and tensile strength internal
if the same limit state used to generate the curves in this stability limit states quantified by reliability index (or,
figure is considered but the pullout bias statistics for fit to probability of failure). The approach is demonstrated
tail are used, then plots for FS and OFS are noticeably using a closed-form solution to compute the reliability
different (Figure 4(d)). index for simple linear limit state (performance) func-
tions with one load term and one resistance term. The
case that is used in this study is the base case found in
6.4. Nominal correlation (ρn) design codes. It applies to typical MSE walls under typi-
Also included in the closed-form solution (Equation 6) is cal operational conditions which do not include extreme
parameter ρn, which is the nominal correlation between events and non-persistent simple surface loads. The for-
nominal load and nominal resistance term values. For mulation considers variability in method bias values that
example, if the load model and resistance model have quantify the accuracy of the load and resistance models
the same random variables that are sampled from the that appear in a limit state equation plus uncertainty in
same population of values, then a non-zero correlation the magnitude of the nominal values selected at the
is likely between random values of Rn and Qn. An time of design. Bias and nominal load and resistance
example is the combination of load model LM1 and pull- values are assumed to be lognormally distributed ran-
out model PM1; both are functions of soil friction dom variables. A novel feature of the approach used in
angle (f) and soil unit weight (γ) and are calculated this paper is the application of the notion of level of
using the same soil in the reinforced soil zone that understanding that is adopted in Canadian LRFD foun-
extends to distance L from the wall face in Figure 1. dation engineering practice.
For this combination of models, greater γ will increase Analysis results show that probabilities of failure (or
the reinforcement load and the pullout capacity, but reliability index) vary strongly with the accuracy of the
the opposite is true for increasing friction angle. If a pull- underlying models that appear in the limit state perform-
out model is independent of friction angle (e.g. pullout ance functions.
model PM2), then Qn and Rn will be positive correlated. The paper demonstrates that the interpretation of mar-
A larger positive nominal correlation coefficient will gin of safety based on conventional ASD practice and prob-
increase β. This can be understood by examining the abilistic notions of safety can be very different. For example,
last parenthetical term in the denominator of Equation a high factor of safety does not necessarily match an accep-
(6). Ignoring the positive nominal correlation will under- table minimum reliability index (or equivalent probability
estimate the margin of safety for that limit state which is of failure). Furthermore, the paper demonstrates that the
an error on the safe side for design. However, ignoring a margin of safety described by the operational safety of fac-
negative correlation will result in an error that is unsafe tor can be greater or smaller than the nominal factor of
for design. safety applied at the time of design.
For the tensile strength limit state that appears in this
study, ρn = 0 because the tensile capacity of a polymeric
(geogrid) reinforcement (e.g. tensile strength model Disclosure statement
TM1) and steel reinforcement elements (e.g. tensile No potential conflict of interest was reported by the authors.
strength model TM2) are independent of soil properties.
Regardless, the assumption of COVRn = 0 assumed earlier
removes the influence of ρn on numerical outcomes for Funding
this limit state as can be seen by inspecting Equation (6). The authors are grateful for financial support from the Natural
A strategy to compute ρn for a project-specific case is Sciences and Engineering Research Council of Canada
described by Lin and Bathurst (2018) for load and pull- (NSERC) [grant number 94344-2013], the Ministry of Trans-
portation of Ontario (MTO) Highway Infrastructure Inno-
out models that have common soil input parameters.
vations Funding Program [grant number 9017-R-0030] and
Pairs of nominal load (Qn) and nominal resistance (Rn) the Japan Ministry of Education, Culture, Sports, Science
are computed using the same values (e.g. f and γ) and Technology (Grant-in-Aid for Scientific Research (B)
from the same sample populations. The correlation No. 17H03309).
GEORISK: ASSESSMENT AND MANAGEMENT OF RISK FOR ENGINEERED SYSTEMS AND GEOHAZARDS 11

References EN1997-1. 2004. Eurocode7: Geotechnical Design – Part 1:


General Rules. Brussels: European Committee for
AASHTO. 2015. Standard Practice for Determination of Standardization (CEN).
Long-Term Strength for Geosynthetic Reinforcement, Fenton, G. A., F. Naghibi, D. Dundas, R. J. Bathurst, and D. V.
Standard Practice R 69-15. Washington, DC: American Griffiths. 2016. “Reliability-based Geotechnical Design in
Association of State Highway and Transportation the 2014 Canadian Highway Bridge Design Code.”
Officials (AASHTO). Canadian Geotechnical Journal 53 (2): 236–251.
AASHTO. 2017. LRFD Bridge Design Specifications. 8th ed. FHWA. 2009. Design of Mechanically Stabilized Earth
Washington, DC: American Association of State Highway Walls and Reinforced Slopes, No. FHWA-NHI-10-024
and Transportation Officials (AASHTO). Vol I and NHI-10-025 Vol II, (eds., Berg, R. R., B. R.
Allen, T. M., and R. J. Bathurst. 2015. “An Improved Simplified Christopher, and N. C. Samtani) Federal Highway
Method for Prediction of Loads in Reinforced Soil Walls.” Administration, 306 p (Vol I) and 378 p (Vol II),
Journal of Geotechnical and Geoenvironmental Engineering Washington, DC.
141 (11): 04015049. Honjo, Y., Y. Kikuchi, and M. Shirato. 2010. “Development of
Allen, T. M., and R. J. Bathurst. 2018. “Application of the the Design Codes Grounded on the Performance-Based
Simplified Stiffness Method to Design of Reinforced Soil Design Concept in Japan.” Soils and Foundations 50 (6):
Walls.” Journal of Geotechnical and Geoenvironmental 983–1000.
Engineering 144 (5): 04018024. Honjo, Y., and O. Kusakabe. 2002. “Proposal of a
Allen, T. M., R. J. Bathurst, and N Bozorgzadeh. 2019. Comprehensive Foundation Design Code: Geo-Code
“Probabilistic Tensile Strength Analysis of Steel Strips in 21 Version 2.” In Proceedings of the International
MSE Walls Considering Corrosion.” Journal of Workshop on Foundation Design Codes and Soil
Geotechnical and Geoenvironmental Engineering 145 (5): Investigation in View of International Harmonization
04019016. and Performance-Based Design, 95–103. Rotterdam:
Allen, T. M., A. S. Nowak, and R. J. Bathurst. 2005. Calibration A.A. Balkema.
to Determine Load and Resistance Factors for Geotechnical Huang, B., and R. J. Bathurst. 2009. “Evaluation of Soil-
and Structural Design. Washington, DC: Transportation Geogrid Pullout Models Using a Statistical Approach.”
Research Board Circular E-C079, 93 p. ASTM Geotechnical Testing Journal 32 (6): 489–504.
Bathurst, R. J., B. Huang, and T. M. Allen. 2011a. “Analysis of ISO 2394. 2015. General Principles on Reliability for
Installation Damage Tests for LRFD Calibration of Structures. Geneva: International Organization for
Reinforced Soil Structures.” Geotextiles and Standardization.
Geomembranes 29 (3): 323–334. Javankhoshdel, S., R. J. Bathurst, and B. Cami. 2018.
Bathurst, R. J., B. Huang, and T. M. Allen. 2011b. “Load and “Influence of Model Type, Bias and Input Parameter
Resistance Factor Design (LRFD) Calibration for Steel Variability on Reliability Analysis for Simple Limit States
Grid Reinforced Soil Walls.” Georisk 5 (3–4): 218–228. with two Load Terms.” Computers and Geotechnics 97:
Bathurst, R. J., B. Huang, and T. M. Allen. 2012. “LRFD 78–89.
Calibration of the Ultimate Pullout Limit State for Lin, P., and R. J. Bathurst. 2018. “Influence of Cross-
Geogrid Reinforced Soil Retaining Walls.” International Correlation Between Nominal Load and Resistance on
Journal of Geomechanics 12 (4): 399–413. Reliability-Based Design for Simple Linear Soil-Structure
Bathurst, R. J., and S. Javankhoshdel. 2017. “Influence of Limit States.” Canadian Geotechnical Journal 55 (2): 279–
Model Type, Bias and Input Parameter Variability on 295.
Reliability Analysis for Simple Limit States in Soil- Miyata, Y., and R. J. Bathurst. 2012. “Reliability Analysis of
Structure Interaction Problems.” Georisk 11 (1): 42–54. Soil-Geogrid Pullout Models in Japan.” Soils and
Bathurst, R. J., S. Javankhoshdel, and T. M. Allen. 2017. “LRFD Foundations 52 (4): 620–633.
Calibration of Simple Soil-Structure Limit States Miyata, Y., and R. J. Bathurst. 2015. “Reliability Analysis of
Considering Method Bias and Design Parameter Geogrid Installation Damage Test Data in Japan.” Soils
Variability.” Journal of Geotechnical and and Foundations 55 (2): 393–403.
Geoenvironmental Engineering 143 (9): 04017053. Miyata, Y., and R. J. Bathurst. 2019. “Statistical Assessment of
Bathurst, R. J., and Y. Miyata. 2015. “Reliability-based Analysis Load Model Accuracy for Steel Grid Reinforced Soil Walls.”
of Combined Installation Damage and Creep for the Tensile Acta Geotechnica 14: 57–70.
Rupture Limit State of Geogrid Reinforcement in Japan.” Miyata, Y., R. J. Bathurst, and T. M. Allen. 2014. “Reliability
Soils and Foundations 55 (2): 437–446. Analysis of Geogrid Creep Data in Japan.” Soils and
Bathurst, R. J., and Y. Yu. 2018. “Probabilistic Prediction of Foundations 54 (4): 608–620.
Reinforcement Loads for Steel MSE Walls Using a Miyata, Y., Y. Yu, and R. J. Bathurst. 2018. “Calibration of Soil-
Response Surface Method.” International Journal of Steel Grid Pullout Models Using a Statistical Approach.”
Geomechanics 18 (5): 04018027. Journal of Geotechnical and Geoenvironmental Engineering
BSI (British Standards Institution). 2010. BS8006-1:2010 144 (2): 04017106.
+A1:2016: Code of Practice for Strengthened/Reinforced Nowak, A. S., and K. R. Collins. 2000. Reliability of Structures.
Soil and Other Fills. Milton Keynes: British Standards New York: McGraw-Hill.
Institution. Nowak, A. S., and M. Szerszen. 2003. “Calibration of
CSA (Canadian Standards Association). 2019. Canadian Design Code for Buildings (AC 318): Part 1–Statistical
Highway Bridge Design Code. CAN/CSA-S6-14, Models for Resistance.” ACI Structural Journal 100 (3):
Mississauga, Ontario, Canada. 377–382.
12 R. J. BATHURST ET AL.

Orr, T. L. L. 2012. “How Eurocode 7 has Affected Geotechnical PWRC. 2013. Design and Construction Manual of
Design: A Review.” Proceedings of the Institution of Civil Geosynthetics Reinforced Soil (Second Revision). Tsukuba:
Engineers – Geotechnical Engineering 165 (6): 337–350. Public Works Research Center. (in Japanese)
Phoon, K. K. 2017. “Role of Reliability Calculations in Yu, Y., and R. J. Bathurst. 2015. “Analysis of Soil-Steel Bar Mat
Geotechnical Design.” Georisk: Assessment and Pullout Models Using a Statistical Approach.” Journal of
Management of Risk for Engineered Systems and Geotechnical and Geoenvironmental Engineering 141 (5):
Geohazards 11 (1): 4–21. 04015006.

You might also like