Professional Documents
Culture Documents
Influence of Input Motion and Site Property Variabilities On Seismic Site Response Analysis
Influence of Input Motion and Site Property Variabilities On Seismic Site Response Analysis
net/publication/245294971
CITATIONS READS
204 2,898
3 authors, including:
Some of the authors of this publication are also working on these related projects:
NSF-funded research "Improving our Understanding of Liquefaction-Induced Displacements: Integration of Remote Sensing Data and Field Data from the 2010/2011 New
Zealand Earthquakes"; E. Rathje, J. Bachhuber, G. DePascale. View project
All content following this page was uploaded by Ellen Rathje on 04 August 2017.
Abstract: Seismic site response analysis evaluates the influence of local soil conditions on earthquake ground shaking. There are
multiple sources of potential uncertainty in this analysis; the most significant pertaining to the specification of the input motions and to the
characterization of the soil properties. The influence of the selection of input ground motions on equivalent-linear site response analysis
is evaluated through analyses performed with multiple suites of input motions selected to fit the same target acceleration response
spectrum. The results indicate that a stable median surface response spectrum 共i.e., within ⫾20% of any other suite兲 can be obtained with
as few as five motions, if the motions fit the input target spectrum well. The stability of the median is improved to ⫾5 to 10% when 10
or 20 input motions are used. If the standard deviation of the surface response spectra is required, at least 10 motions 共and preferably 20兲
are required to adequately model the standard deviation. The influence of soil characterization uncertainty is assessed through Monte
Carlo simulations, where variations in the shear-wave velocity profile and nonlinear soil properties are considered. Modeling shear-wave
velocity variability generally reduces the predicted median surface motions and amplification factors, most significantly at periods less
than the site period. Modeling the variability in nonlinear properties has a similar, although slightly smaller, effect. Finally, including the
variability in soil properties significantly increases the standard deviation of the amplification factors but has a lesser effect on the standard
deviation of the surface motions.
DOI: 10.1061/共ASCE兲GT.1943-5606.0000255
CE Database subject headings: Soil properties; Uncertainty principles; Seismic effects; Earthquakes.
Author keywords: Site response analysis; Input motions; Soil properties; Uncertainty.
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 1. Sources of uncertainty in seismic site response analysis
Appropriate nonlinear properties can be evaluated via laboratory transfer functions, and random vibration theory 共Schneider et al.
testing on “undisturbed ” specimens or through empirical corre- 1991兲. The study considered the relative influence of uncertainties
lations that relate these curves to soil type, stress conditions, etc. in the model parameters for the earthquake source, source-to-site
Various methods of analysis are available for simulating the path, and site properties. The results showed that site property
seismic response of a soil deposit. These methods encompass dif- variability caused a decrease in the median surface response spec-
ferent models for the stress strain constitutive response of the soil tra, as compared with analyses incorporating source and path un-
共i.e., equivalent linear or nonlinear兲 and may simulate one-, two-, certainties. Site property variability also caused an increase in the
or three-dimensional wave propagation. In practice, one- standard deviation of the surface motions.
dimensional analysis of vertically propagating horizontal shear Bazzurro and Cornell 共2004a兲 performed Monte Carlo simula-
waves is most often performed, and the nonlinear response of the tions of nonlinear site response analysis to evaluate the amplifi-
soil is most commonly modeled through the equivalent-linear cation factors for sites with uncertain soil properties. The input
共EQL兲 approximation, which uses an iterative technique to select parameters for the nonlinear models were varied statistically, and
strain-compatible soil properties for a linear-elastic analysis. The a large suite of 78 input motions was used. These input motions
EQL approach has several shortcomings related to overdamping were selected from a large range of earthquakes and represented a
of high frequencies and overamplification at the site period par- significant source of variability with peak ground accelerations
ticularly during strong shaking 共Kramer 1996兲. Nonetheless, this ranging from 0.01 to 1.0 g. Bazzurro and Cornell 共2004a兲 dem-
technique is computationally efficient and has yielded acceptable
onstrated that including soil property variability slightly de-
site response results over a range of applications.
creased the median amplification factor and slightly increased its
Much of the previous work on site response uncertainties has
standard deviation; but they concluded that the variability intro-
focused on the influence of site property uncertainty. Field and
duced by the input motions was more important than the variabil-
Jacob 共1993兲 performed Monte Carlo simulations of the linear
ity introduced by soil property uncertainties.
elastic response of the Turkey Flat, California site subjected to
weak motions. Their results indicated that uncertainties in the Andrade and Borja 共2006兲 performed Monte Carlo simulations
shear-wave velocity profile and small strain damping ratio signifi- of EQL and nonlinear site response analyses to evaluate the sen-
cantly affected the amplification predictions. Kwok et al. 共2008兲 sitivity of these analytical procedures to their input parameters.
and Stewart and Kwok 共2008兲 investigated the contributions of The study focused on the ground motion parameters Arias inten-
shear-wave velocity and nonlinear curve variability on nonlinear sity and acceleration spectrum intensity and compared the cumu-
site response results for Turkey Flat, Calif. and three other sites. lative distribution functions of these parameters for single
These studies indicated that, depending on the site and the spec- earthquake motions. In this way, the variability introduced by
tral period, shear-wave velocity variability or nonlinear curve differences in input motions was not included in the analysis. The
variability may be the dominant contributor to the variability in cumulative distribution functions showed similar slopes for the
the computed response. nonlinear and EQL analyses, and for each type of analysis the
Roblee et al. 共1996兲 performed Monte Carlo simulations of median value was similar to the value computed for the best-
EQL site response analysis to evaluate the influence of various estimate soil properties. Andrade and Borja 共2006兲 concluded that
uncertainties on computed surface response spectra. Input mo- nonlinear and EQL analyses demonstrate similar sensitivities to
tions were prescribed using the stochastic finite-fault model, their input parameters. However, it should be noted that the EQL
which uses a seismological model of the earthquake source and assumption breaks down for conditions, in which large strains are
path to predict the input Fourier amplitude spectrum 共FAS兲. Sur- generated 共i.e., large intensity and/or soft soil conditions兲, and in
face response spectra were computed using the input FAS, EQL these cases nonlinear analysis should be performed.
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
This paper focuses on the influence of input motion selection
and soil property variability on the results from site response
analysis. The numerical simulations of site response are per-
formed using one-dimensional EQL analysis because this method
of analysis is most common in practice. All analyses are per-
formed for a deep alluvium site based on the characterization of
the Sylmar County Hospital 共SCH兲 in Southern California 共Chang
1996兲. The influence of the selection of input ground motions on
EQL site response analysis is evaluated through analyses per-
formed with multiple suites of input motions selected to fit the
same target acceleration response spectrum. In this effort, we in-
vestigate how the number of motions and their goodness of fit
influence the computed response. The influence of soil property
uncertainty is assessed through Monte Carlo simulations, where
variations in the shear-wave velocity profile and nonlinear soil
properties are considered. While the previous studies on soil prop-
erty uncertainty demonstrated that this source of uncertainty is
important, in this study we quantify how modeling soil property
variability influences both the median site response and its stan-
dard deviation.
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Soil Property Variability Table 1. Parameters for Toro 共1995兲 Model for Shear-Wave Velocity
Randomization
The most critical soil property data used in seismic site response Vs30
analysis are the shear-wave velocity profile and the nonlinear
modulus reduction and damping curves. There are multiple rea-
sons to incorporate soil property variability in seismic site re- Parameter ⬎750 m/s 360–750 m/s 180–360 m/s ⬍180 m/s
sponse analysis. Soil properties vary spatially across the footprint ln Vs 0.36 0.27 0.31 0.37
of a site due to natural variability induced by geologic processes. 200 0.42 1.00 0.98 0.50
This aleatory variability is random in nature and cannot be con- do 0.0 0.0 0.0 0.0
trolled or reduced, although the level of aleatory variability will b 0.063 0.293 0.344 0.744
vary from site to site. The level of aleatory variability at a site can 0 0.95 0.97 0.99 0.00
be quantified by measuring multiple shear-wave velocity profiles ⌬ 3.4 3.8 3.9 5.0
across a site and is often related to the type of depositional envi-
ronment 共e.g., less variability at sites within deep alluvial basins,
more variability at sites within active river valleys兲. ditions 共i.e., perfect correlation versus statistical independence兲.
Another reason to incorporate soil property variability is Toro 共1995兲 developed a model of the interlayer correlation of Zi
epistemic uncertainty in the measurement of the soil properties. based on an analysis of over 500 shear-wave velocity profiles.
Uncertainties in the soil properties are caused by measurement The standard normal variable for the surface layer 共Z1 , i = 1兲 is
errors, sample disturbance, or the use of generic soil type/site independent of all other layers and defined as
class data. As opposed to aleatory variability, epistemic uncer-
Z1 = 1 共2兲
tainty can be reduced through the collection of more and/or better
data. In practice, it is difficult to differentiate between aleatory where 1 = random normal variable with zero mean and unit stan-
variability and epistemic uncertainty as they pertain to soil char- dard deviation. Zi is correlated with the layer above it using 共Toro
acterization, and we will not attempt to do so in this study. Thus, 1995兲
these sources of uncertainty/variability will be combined and sim-
ply considered total soil property variability. Zi = IL · Zi−1 + i · 冑1 − IL
2
共3兲
The variability in soil properties can be incorporated in site where Zi−1 = standard normal variable of the previous layer; i
response analysis through a Monte Carlo simulation. A Monte = new normal random variable with zero mean and unit standard
Carlo simulation estimates the statistical response of a system by deviation; and IL = interlayer correlation coefficient. Toro 共1995兲
computing the system response for different values of the input modeled the interlayer correlation as depth 共d兲 and layer thickness
parameters. These input parameters are generated from specified 共t兲 dependent
probability distributions, and the simulation results for each set of
input parameters are used to estimate the mean, standard devia- IL共d,t兲 = 关1 − d共d兲兴t共t兲 + d共d兲 共4兲
tion, or full probability distribution of the response. For the site where d共d兲 = depth dependent correlation coefficient and t共t兲
response problem, the relevant input parameters are the shear- = thickness dependent correlation coefficient. These correlation
wave velocity profile and nonlinear soil property curves, and the coefficients are defined as 共Toro 1995兲
冋 册
relevant response parameters are surface response spectral values
冦 冧
b
and spectral amplification factors. d + do
200 d 艋 200m
d共d兲 = 200 + do 共5兲
Shear-Wave Velocity Profile 200 d ⬎ 200m
Randomized shear-wave velocity profiles generated for Monte
Carlo simulations are based on a baseline shear-wave velocity
profile, a measure of variability, and an interlayer correlation. The
t共t兲 = o exp 冉 冊 −t
⌬
共6兲
profiles generated in this study are based on the statistical models where 200, do, b, o, and ⌬ are model parameters. Generally, this
of Toro 共1995兲 and assume that shear-wave velocity is lognor- model incorporates a larger interlayer correlation for deeper lay-
mally distributed at any given depth. Based on this assumption, ers and thinner layers. If one wants to model constant interlayer
the shear-wave velocity of layer i 关Vs共i兲兴 can be derived from the correlation, the Toro 共1995兲 model parameters can be judiciously
baseline value for that layer 关Vs,o共i兲, assumed to represent the selected to achieve this goal 共i.e., 200 = IL; b = o = 0.0兲.
mean in logarithmic space兴, the standard deviation of the natural The parameters required to develop randomized shear-wave
logarithm of Vs 共ln Vs兲, and a random standard normal variable velocity profiles from the Toro 共1995兲 models are the baseline
for layer i 共Zi兲 using 共Toro 1995兲 shear-wave velocity profile, the standard deviation of the natural
Vs共i兲 = exp兵ln关Vs,o共i兲兴 + Zi · ln Vs其 共1兲 log of the shear-wave velocity 共ln Vs兲, and the interlayer correla-
tion parameters 共200, do, b, o, and ⌬兲. Toro 共1995兲 developed
where Zi represents the number of standard deviations from the these model parameters 共Table 1兲, including baseline shear-wave
mean value of ln关Vs,o共i兲兴 in log space. velocity profiles, for different generalized site classes 共i.e. average
Various assumptions have been applied to assign the layer val- 30 m shear-wave velocity Vs30兲 using a total of 557 shear-wave
ues of Zi. McGuire et al. 共1989兲 and Toro et al. 共1992兲 assumed velocity profiles. For site-specific applications, the baseline shear-
that the Zi values were perfectly correlated and, thus, a single wave velocity profile is generally evaluated from site-specific in
value of Z was assigned to all layers in the profile. Costantino et situ testing, while the other parameters can be estimated based on
al. 共1991兲 assumed that the Zi values were independent and as- site class and Table 1. However, the ln Vs values in Table 1 rep-
signed independent random values to each layer. Neither of these resent variability between velocity profiles across an entire site
assumptions is rigorously correct and they represent extreme con- class and thus would typically be too large for most site-specific
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 3. Examples of randomized 共a兲 shear-wave velocity profiles; 共b兲 nonlinear property curves
studies. In these cases, ln Vs should be estimated from multiple resents the most comprehensive investigation regarding the
measured profiles across the site or from knowledge of the variability of nonlinear property curves. This model predicts base-
geology/variability across the site. Toro 共1995兲 also developed a line modulus reduction and damping curves based on plasticity
model for randomizing layer thicknesses, but that model was not index, confining pressure, overconsolidation ratio, loading fre-
used in this study. quency, and number of loading cycles and also provides an esti-
Fig. 3共a兲 displays examples of randomized shear-wave veloc- mate of the standard deviation associated with these curves for a
ity profiles for the deep alluvium SCH site used in this study. The given set of generic soil conditions. The standard deviation de-
profiles were developed using the baseline shear-wave velocity rived from laboratory tests on site-specific specimens would gen-
profile from Chang 共1996兲 and values of ln Vs equal to 0.1 and erally be smaller.
0.3. To visually demonstrate the relative ranges of velocity pro- The Darendeli and Stokoe 共2001兲 model assumes that the pa-
files generated for ln Vs = 0.1 and 0.3, also shown are the ⫾2ln rameters G / Gmax and D are normally distributed at a given strain
profiles 共note that these profiles do not represent the limits of the level and that the standard deviation for each is a function of the
randomization兲. The one realization shown for ln Vs = 0.1 was magnitude of G / Gmax and D, respectively. The standard deviation
generated for IL = 1.0, and thus the velocity profile simply repre- of the normalized shear modulus 共NG兲 is given as
sents a shift of the baseline profile. The one realization shown for
ln Vs = 0.3 was generated for IL = 0.5. For IL ⬍ 1.0, the character NG = 0.015 + 0.16冑0.25 − 共G/Gmax − 0.5兲2 共7兲
of the baseline profile is not maintained. For the realization This function produces the smallest standard deviation 共NG
shown, some layers are above the baseline curve, while others are ⬃ 0.015兲 when G / Gmax is close to 0.0 or 1.0, and the largest
below. A velocity inversion, in which the velocity decreases with standard deviation 共NG = 0.095兲 when G / Gmax is equal to 0.5.
depth, is observed between 20 and 40 m. While not all of the The standard deviation of the damping ratio 共D共%兲兲 is given as
IL = 0.5 realizations displayed velocity inversions, they are more
common as IL is reduced. Fig. 3共a兲 demonstrates that while ln Vs D共%兲 = 0.0067 + 0.78冑D共%兲 共8兲
affects the velocity profiles generated by the Monte Carlo simu-
lation, IL also has a profound effect. This function produces a standard deviation that increases with
increasing damping ratio 共D共%兲 ⬃ 0.75 at D = 1% and D共%兲
⬃ 3.0 at D = 15%兲.
Nonlinear Soil Properties
G / Gmax and D curves are not independent of each other be-
The majority of studies investigating nonlinear soil property cause a modulus curve that is more linear 共i.e., shifted up兲 is
curves 共i.e., modulus reduction G / Gmax and damping D curves兲 associated with a damping curve that is shifted down to smaller
have focused on predicting baseline curves based on soil type, damping ratios. Thus, the generation of G / Gmax and D curves
stress conditions, etc., without any rigorous assessment of vari- should include negative correlation between the curves 共D,NG
ability. The empirical model of Darendeli and Stokoe 共2001兲 rep- ⬍ 0.0兲. To generate correlated G / Gmax and D curves from base-
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
line 共mean兲 curves, the standard deviations from Eqs. 共7兲 and 共8兲, Input motion suites of five, 10, and 20 motions were selected
and D,NG, the following expressions are used for each shear using the Kottke and Rathje 共2008兲 procedure and scaled to fit
strain value 共␥兲 in the curves: the target spectrum and its standard deviation. Motions were
selected from a larger catalog of 105 motions that came from
G/Gmax共␥兲 = 关G/Gmax共␥兲兴mean + 1 · NG 共9兲 earthquake magnitudes between 6.2 and 6.9 and distances be-
tween 5 and 40 km 共Kottke 2006兲. Five suites of each size were
selected and scaled, and the number of motions common to the
D共␥兲 = 关D共␥兲兴mean + D,NG · 1 · D + 冑1 − D,NG
2
· 2 · D suites was limited to about one-half of the suite size. Fig. 4 com-
共10兲 pares two selected suites: one five-motion suite and one 20-
motion suite. Both the five-motion suite and 20-motion suite
where 1 and 2 = uncorrelated random variables with zero mean provide an acceptable fit with the target spectrum, but the 20-
and unit standard deviation; 关G / Gmax共␥兲兴mean and 关D共␥兲兴mean motion suite provides a better fit to the standard deviation across
= baseline values evaluated at strain level ␥; NG and D the periods considered. The standard deviation for the five-motion
= standard deviations computed from Eqs. 共7兲 and 共8兲 at the base- suite varies considerably with period because of the difficultly in
line values of 关G / Gmax共␥兲兴mean and 关D共␥兲兴mean, respectively; and controlling the standard deviation with only five motions. Be-
D,NG = correlation coefficient between G / Gmax and D. Eqs. 共9兲 cause of the additional motions in the 20-motion suite, it can
and 共10兲 are applied at different strain levels, but the same values better model a smooth variation of standard deviation with period.
of 1 and 2 are used at each strain level 共i.e., perfect correlation Fig. 4 also shows the standard deviations for the other five-
between strain levels兲. Because the standard deviation model as- motion and 20-motion suites. For the five-motion suites, the stan-
sumes a normal distribution, cutoffs must be applied to avoid dard deviations not only vary significantly with period but they
unrealistic values 共i.e., G / Gmax ⬎ 1.0; G / Gmax ⬍ 0.0; and D ⬍ 0% also vary considerably between suites because of the difficulty in
are not allowed兲. controlling the standard deviation with only five motions. The
Fig. 3共b兲 displays an example of a modulus reduction and 20-motion suites show less variability in the standard deviation
damping curve pair generated using Eqs. 共7兲–共10兲 and D,NG equal between suites and across periods.
to ⫺0.5. Because of the negative correlation, the modulus reduc- The statistical fit for each suite is summarized in Table 2 in
tion curve that is below the baseline curve is associated with a terms of the RMSE for the target spectrum, the ln RMSE for
damping curve that is above the baseline. However, the relative the standard deviation, and the maximum error of the suite rela-
shifts of the curves are not the same because D,NG is not equal to tive to the target spectrum and target standard deviation. The
⫺1.0. RMSE represents the average percent difference between the tar-
get spectrum and the suite, while ln RMSE represents the aver-
age difference in standard deviation between the target and suite
Influence of Input Motion Selection on Computed 共Kottke and Rathje 2008兲. Generally, as the number of motions in
Response the suite increases, all of the statistical fit parameters improve.
For the fit to the target spectrum, RMSE decreases by a factor of
Site response analyses were performed for the alluvial SCH 2 and the maximum error is reduced below 10% as the suites
site using the baseline shear-wave velocity profile shown in increase from five to 20 motions. However, even the RMSE val-
Fig. 3共a兲 and four sets of baseline nonlinear soil properties de- ues for the five-motion suites 共⬃0.06–0.07兲 represent an average
rived from Darendeli and Stokoe 共2001兲 for a plasticity index of fit within 5–10% of the target. The statistical fit parameters for ln
0, an overconsolidation ratio of 1.0, and mean effective confining are significantly improved by the addition of more motions,
pressures of 50, 250, 550, and 750 kPa 共corresponding depths of although the maximum error is still often above 20%.
4, 19, 45, and 75 m兲. The Vs30 of the site is 315 m/s, which Each of the input motion suites was propagated through the
classifies it as a stiff soil site, and the initial site period based on baseline SCH soil profile to compute the surface response spectra.
the linear elastic transfer function for the site is 0.55 s. The site The median surface response spectrum and the standard deviation
response simulations were performed using the program Strata of the surface response spectra 共ln Sa兲 at each period were com-
共Kottke and Rathje 2009兲. Strata performs EQL site response puted for each suite of motions. Essentially, each suite represents
analysis using time domain input motions or random vibration a reasonable realization of the input motions, such that comparing
theory 共Rathje and Ozbey 2006兲 and can perform Monte Carlo the results from different suites provides an evaluation of the
simulations on the shear-wave velocity profile, nonlinear soil statistical stability of the computed response. Fig. 5 shows the
properties, and depth to bedrock. median surface response spectra for each of the five suites gener-
Suites of input motions were selected to fit a deterministic ated for each size 共i.e., five, 10, and 20 motions兲, along with
target acceleration response spectrum representing a M w = 6.5 and the median input response spectra. There is some suite-to-suite
R = 20 km event from Ambrahamson and Silva 共1997兲. This re- variability in the median response spectra from the five-motion
sponse spectrum has a peak ground acceleration of 0.17 g, and the suites 关Fig. 5共a兲兴, with differences between suites as large as 15–
target standard deviation ranges from 0.5 to 0.75 in natural log 20% at a single period. This variability between suites is reduced
units. This moderate intensity target spectrum allowed for a sig- to 5–10% as the number of motions in a suite is increased to
nificant number of potential motions to be identified such that 20 关Fig. 5共c兲兴. Fig. 5共d兲 displays the standard deviation 共in natural
input suites of different sizes could be selected. A response spec- log units兲 of the median surface spectra for each set of five
trum from a deterministic event was used rather than a uniform suites 共i.e., five suites of five, 10, and 20 motions兲. The standard
hazard spectrum so that the target spectrum was representative of deviation reduces by approximately 50% when the number of
a single earthquake event. As the goal of this investigation was to motions is increased from five to 10, with no significant improve-
address input motion selection for a given target spectrum, the use ment when the number of motions is increased to 20. These
of a target spectrum from a deterministic event rather than from results indicate that a stable median surface spectra, within ⫾15
PSHA should not influence the conclusions. to 20% of any other suite, can be obtained with suites of five
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 4. Suites of five and 20 motions selected to fit the target spectrum
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 5. Median surface and input motion response spectra for suites of 共a兲 five; 共b兲 10; 共c兲 20 motions; and 共d兲 standard deviation of the median
surface response spectra
If one wants to consider the median plus one standard devia- ity 共ln Vs兲 was varied between 0.1 and 0.4. The nonlinear
tion site response, an accurate assessment of the surface ln Sa is modulus reduction and damping curves were generated from
required. These results demonstrate that at least 10 motions, and baseline curves and standard deviation values derived from
preferably 20, are needed to obtain a statistically stable estimate Darendeli and Stokoe 共2001兲, as described previously. Four sets
of ln Sa. Additionally, while site response affects ln Sa, the main of nonlinear curves were generated for each site realization, one
character of ln Sa is controlled by ln of the input suite 共Figs. 4 for each confining pressure 共50, 250, 550, and 750 kPa兲. No in-
and 6兲. Therefore, the input motions should be scaled to fit a terlayer correlation was modeled for the nonlinear curves. 60
target input ln because the surface ln Sa is influenced most by the property randomizations, in which each realization included a ve-
ln of the input motions. locity profile and four sets of nonlinear property curves, were
coupled with the 10 input motions for all analyses to ensure stable
estimates of the surface motion were obtained 共Trent 2008兲. Thus,
Influence of Site Property Variability on Computed all results represent the median and standard deviation computed
Response from a total of 600 analyses.
Fig. 7 displays the site response results for the SCH site with
Monte Carlo simulations of the EQL site response of the SCH site the baseline site characterization. The 10 scaled input motions are
共Fig. 3兲 were performed using a suite of 10 input motions 共Suite shown in Fig. 7共a兲, while the surface response spectra and ampli-
10–1 from Table 2兲 and varying the shear-wave velocity profile fication factors 共AF= Sa,surface / Sa,rock兲 are shown in Figs. 7共b and
and nonlinear soil properties. The original scale factors for the c兲, respectively. The variability about the median values due to
input motions were doubled such that the median peak ground record-to-record variability is quantified by the standard deviation
acceleration of the input motions was 0.35 g. This increase in in natural log units, which is shown for the input rock motions
intensity was used to induce more nonlinearity in the site. The 共ln Sa兲, the surface motions 共ln Sa兲, and the amplification factors
shear-wave velocity profiles were developed from the baseline 共ln AF兲 in Fig. 7共d兲. Figs. 7共b and c兲 demonstrate that the dy-
profile in Fig. 3共a兲 and the Toro 共1995兲 randomization model. The namic response of the site causes amplification of the motion at
Toro 共1995兲 correlation parameters were modified to model a con- almost all periods, with the largest amplification occurring around
stant IL equal to 0.8, and the variability in the shear-wave veloc- the degraded natural period of the site 共0.7–0.9 s兲. The average
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 6. Standard deviation of surface response spectra for suites of 共a兲 five; 共b兲 10; 共c兲 20 motions; and 共d兲 of standard deviations of the surface
response spectra
strain level induced in the surface layer 共i.e., the softest layer兲 changes ln Sa on the order of ⫾20%, the main character of ln Sa
over these 10 motions is 0.20%, while the maximum strain over is controlled by the input motion suite. The standard deviation for
the 10 motions is about 1.5%. the amplification factors 共ln AF兲 is generally smaller than for the
Fig. 7共d兲 demonstrates how site response affects the standard surface response spectra because the amplification factors are nor-
deviation of a suite of motions. The surface motions, on average, malized by the rock spectral acceleration, which reduces the in-
have smaller ln Sa than the rock motions at shorter periods fluence of record-to-record variability 关Fig. 7共d兲兴.
and larger ln Sa than the rock motions at longer periods 共i.e., Fig. 8 displays the median and standard deviation of the sur-
periods greater than about 0.8 s兲. The change in ln Sa is caused by face response spectra computed from Monte Carlo simulations
nonlinear soil amplification, where amplification is a function of with different levels of shear-wave velocity randomization
the intensity of the input motion. At shorter periods, nonlinear soil 共ln Vs兲. Increasing levels of shear-wave velocity randomization
amplification means that less amplification occurs for larger input causes a decrease in the median spectral acceleration, with the
intensities because of increased levels of induced strain and reduction occurring predominantly at periods less than 2.0 s. The
damping. At longer periods, nonlinear soil amplification means reduction ranges from about 5% for ln Vs = 0.1 to about 25% for
that more amplification occurs for larger input intensities because ln Vs = 0.4. This reduction could be as large as 50% for softer sites
of increased levels of induced strain and more period lengthening. 共Trent 2008兲. The decrease in the median surface response spec-
Considering the effect on ln Sa, at shorter periods a larger inten- trum is caused by site realizations that include soft layers and/or
sity motion within a suite will be amplified less than a lower large velocity contrasts, which accumulate more strain and expe-
intensity motion, resulting in less variability between the motions rience more damping and less amplification during seismic load-
at the ground surface. At longer periods 共i.e., longer than the ing. Realizations that include stiffer layers may result in larger
initial site period兲, a larger intensity motion will be amplified amplification; but these larger values are not large enough to
more than a lower intensity motion, resulting in more variability counteract the significant reduction in amplification caused by the
in the surface motions than in the input motions. The transition softer realizations. For ln Vs = 0.2, the mean strain is about the
from smaller surface ln Sa to larger surface ln Sa generally occurs same as for the baseline properties, but the maximum strain over
around the natural period of the site. Note that while site response all realizations is 8%. Obviously, these very large strain levels
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 7. 共a兲 Input rock motions; 共b兲 surface response spectra; 共c兲 amplification factors; and 共d兲 standard deviations for site response calculated with
the baseline soil properties
render the EQL approximation invalid, identifying a drawback of increasing by more than a factor of 2.0 in some cases.
modeling large shear-wave velocity variability in EQL analysis. It is commonly believed that it is conservative to consider
The values of ln Sa for the surface motions indicate that uncertainties in soil properties when performing site response
increasing levels of shear-wave velocity randomization cause analyses. However, Figs. 8 and 9 indicate that including shear-
a slight increase in ln Sa at periods less than about 1.0 s and wave velocity uncertainty in EQL site response analysis through
cause a decrease in ln Sa at periods greater than 1.0 s 关Fig. 8共b兲兴. Monte Carlo simulations consistently produces median surface
These changes are in the opposite direction of the changes in response spectra and amplification factors that are smaller than
ln Sa caused by the site response when using only the baseline those computed from the baseline shear-wave velocity profile.
properties 关i.e., reduction in ln Sa at short periods, and an increase Therefore, it is unconservative to blindly apply shear-wave veloc-
in ln Sa at long periods; Fig. 7共d兲兴. The most pronounced increase ity randomization to site response analyses without considering
in standard deviation occurs at periods around 0.1 s, with in- this change in median response.
creases of 0 to 20% for ln Vs 艋 0.2 and more than 40% for The model of interlayer correlation for the shear-wave velocity
ln Vs 艌 0.3. The decrease in ln Sa at longer periods ranges from randomization can profoundly influence the Monte Carlo simula-
5% to 20%. tion results of site response. Fig. 10 compares the median and
Fig. 9 displays the median amplification factors and the stan- standard deviation of the surface response spectra for shear-wave
dard deviation for the amplification factors 共ln AF兲 for different velocity profiles generated with ln Vs = 0.2 and different levels of
levels of shear-wave velocity randomization. The reductions in interlayer correlation. Constant values of IL between 0.5 and 1.0
median AF are exactly the same as those for surface response were used, as well as the values from the Toro 共1995兲 model
spectra, although the changes in the long period range are more for sites with Vs30 = 180– 360 m / s 共Table 1兲. For the SCH layer
apparent in this representation of the results 共note the increase in thicknesses and depths shown in Fig. 3, IL from the Toro 共1995兲
AF at periods greater than 1.5 s for ln Vs = 0.4兲. Shear-wave ve- model ranged from 0.3 at the surface to a maximum of 0.7 at
locity randomization causes a consistent increase in ln AF at all depth. Results for IL = 0.0 共not shown兲 were similar to those
periods, which is different than the effect on ln Sa. The difference for IL = 0.5. Fig. 10 demonstrates that decreasing IL causes a
is a manifestation of ln Sa at the surface being influenced by ln Sa more significant reduction in the median surface response spec-
of rock motions, while ln AF is less influenced by ln Sa of the trum, although there is little impact on the standard deviation.
rock motions. The change in ln AF can be significant, with ln AF The reduction in the median surface spectrum reaches a maxi-
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 8. Influence of shear-wave velocity randomization on the Fig. 9. Influence of shear-wave velocity randomization on the
共a兲 median; 共b兲 standard deviation of the surface response spectra 共a兲 median; 共b兲 standard deviation of the amplification factors
mum of about 20–25% for IL 艋 0.5. This reduction appears to be intensity, as the degree of induced nonlinearity will affect the
caused by the velocity inversions that are generated within the comparison. When both the shear-wave velocity profile and the
velocity profiles as IL is reduced. Thus, it is important to con- nonlinear property curves are varied, the median amplification
sider the value of IL used in generating the shear-wave velocity factor continues to decrease 共about 20% less than the baseline兲
profiles. and the standard deviation continues to increase. These results
Fig. 11 displays the median and standard deviation of the am- reveal the additive nature of including different sources of site
plification factors computed from Monte Carlo simulations, property variability.
where the modulus reduction and damping curves were varied
共D,NG = −0.5兲. Although significant variability was modeled in the
nonlinear property curves 关Fig. 3共b兲兴, including this variability Conclusions
only caused a reduction of about 5–15% in the amplification fac-
tor at periods less than 1.0 s. Again, modeling the variability Accounting for uncertainties is becoming an important compo-
results in less amplification because softer realizations, in this nent of seismic risk assessment. Quantifying uncertainty and
case softer G / Gmax curves and D curves with higher damping, propagating that uncertainty through numerical simulations pro-
affect amplification more than the stiffer realizations. Also shown vide for more robust estimates of seismic performance. For site
in Fig. 11 are the results for different values of D,NG, but the response analysis, the main sources of uncertainty are the in-
differences are minor. For ln AF, including the variability of the put motion characterization, the shear-wave velocity and nonlin-
G / Gmax and D curves results in an increase of 50–100% at peri- ear property characterization of the soil, and the selection of the
ods less than 1.0 s. The influence of the variability of the G / Gmax method of analysis. This paper focused on the effect of input
and D curves would be less pronounced for smaller intensity motion selection and soil property uncertainty on EQL site re-
input motions 共Kottke 2006兲. sponse simulations.
Fig. 12 compares the effect of modeling the variability of The characteristics of the input rock motions for site response
共1兲 the shear-wave velocity 共ln Vs = 0.2; IL = 0.8兲; 共2兲 the G / Gmax analysis influence site response simulations because of the dif-
and D curves 共D,NG = −0.5兲; and 共3兲 all of these parameters. ferences in the induced nonlinearity. Using the methodology of
Generally, the effect of modeling shear-wave velocity variability Kottke and Rathje 共2008兲, input motion suites of five, 10, and 20
with ln Vs = 0.2 is similar to that observed when modeling the motions were selected and scaled to fit the same target input re-
variability in G / Gmax and D curves, except close to the site pe- sponse spectrum. Including more motions in the input suite re-
riod, where varying the shear-wave velocity has a greater effect. sulted in a slightly better fit to the target spectrum and a
These observations are specific to this site and this input motion significantly better fit to the target standard deviation. The differ-
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
Fig. 10. Influence of interlayer correlation coefficient 共IL兲 on the Fig. 11. Influence of modulus reduction and damping curve random-
共a兲 median; 共b兲 standard deviation of the surface response spectra ization on the 共a兲 median; 共b兲 standard deviation of the amplification
factors
Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org
River plant.” Rep. No. CEERC-91-003, Structural Analysis Div.,
Nuclear Energy Dept., Brookhaven National Laboratory, Upton, N.Y.
Darendeli, M. B., and Stokoe, K. H., II. 共2001兲. “Development of a new
family of normalized modulus reduction and material damping
curves.” Rep. No. GD01-1, Univ. of Texas, Austin, Tex.
Field, E. H., and Jacob, K. H. 共1993兲. “Monte Carlo simulation of the
theoretical site response variability at Turkey Flat, California, given
the uncertainty in the geotechnically derived input parameters.”
Earthquake Spectra, 9共4兲, 669–701.
Hancock, J., Watson-Lamprey, J., Abrahamson, N., Bommer, J.,
Markatis, A., McCoy, E., and Mendis, R., 共2006兲. “An improved
method of matching response spectra of recorded earthquake ground
motion using wavelets.” J. Earthquake Eng., 10共1兲, 67–89.
Idriss, I. M. 共2004兲 “Evolution of the state of practice.” Int. Workshop on
the Uncertainties in Nonlinear Soil Properties and Their Impact on
Modeling Dynamic Soil Response, Pacific Earthquake Engineering
Research Center, Richmond, Calif.
Kottke, A., and Rathje, E. M. 共2008兲. “A semi-automated procedure for
selection and scaling of recorded earthquake motions for dynamic
analysis.” Earthquake Spectra, 24共4兲, 911–932.
Kottke, A., and Rathje, E. M. 共2009兲. “Technical Manual for Strata.” Rep.
No. 2008/10, Pacific Earthquake Engineering Research Center, Berke-
ley, Calif.
Kottke, A. R. 共2006兲. “Impact of input ground motions and site variability
on seismic site response.” MS thesis, Dept. of Civil, Architectural,
and Environmental Engineering, Univ. of Texas, Austin, Tex.
Kramer, S. L. 共1996兲. Geotechnical earthquake engineering, Prentice-
Hall, Upper Saddle River, N.J.
Kwok, A. O., Stewart, J. P., and Hashash, Y. M. A. 共2008兲. “Nonlinear
ground response analysis of Turkey Flat shallow stiff soil site to
strong ground motion.” Bull. Seismol. Soc. Am., 98共1兲, 331–343.
McGuire, R., Toro, G., O ’Hara, T., Jacobson, J., and Silva, W. 共1989兲.
“Probabilistic seismic hazard evaluations at nuclear plant sites in the
Fig. 12. Influence of randomization of different parameters on the 共a兲 central and eastern United States: Resolution of the Charleston earth-
median; 共b兲 standard deviation of the amplification factors quake issue.” Technical Rep. No. NP-6395-D, Electric Power Re-
search Institute, Palo Alto, Calif.
Naeim, F., Alimoradi, A., and Pezeshk, S. 共2004兲. “Selection and scaling
Julian Bommer at Imperial College, London. Financial support
of ground motions time histories for structural design using genetic
was provided by the Lifelines Program of the Pacific Earthquake algorithms.” Earthquake Spectra, 20共2兲, 413–426.
Engineering Research 共PEER兲 Center under Grant No. SA5405-
Powers, M. 共2004兲. “Design ground motion library.” Geotechnical Engi-
15811 and the Nuclear Regulatory Commission under Grant No.
neering for Transportation Projects, Proc., Geo-Trans 2004, Geotech-
NRC-04-07-122. This support is gratefully acknowledged. nical Special Publication, ASCE, Reston, Va., Vol. 126, 778–786.
Rathje, E. M., and Ozbey, M. C. 共2006兲. “Site specific validation of
References random vibration theory-based site response analysis.” J. Geotech.
Geoenviron. Eng., 132共7兲, 911–922.
Roblee, C., Silva, W., Toro, G., and Abrahamson, N. 共1996兲. “Variability
Ambrahamson, N. A., and Silva, W. J. 共1997兲. “Empirical response spec-
in site-specific seismic ground-motion design predictions.” Geotech.
tral attenuation relations for shallow crustal earthquakes.” Seismol.
Spec. Publ., 58共2兲, 1113–1133.
Res. Lett., 68共1兲, 94–127.
Schneider, J. F., Silva, W. J., Chiou, S. J., and Stepp, J. C. 共1991兲. “Es-
Andrade, J. E., and Borja, R. I. 共2006兲. “Quantifying sensitivity of local
timation of ground motion at close distances using the band-limited-
site response models to statistical variations in soil properties.” Acta
white-noise model.” Proc., 4th Int. Conf. on Seismic Zonation, EERI,
Geotech., 1共1兲, 3–14.
Baker, J. W., and Cornell, C. A. 共2006兲. “Spectral shape, epsilon, and Stanford, Calif., Vol. 4, 187–194.
record selection.” Earthquake Eng. Struct. Dyn., 35, 1077–1095. Stewart, J. P. and Kwok, A. O. L. 共2008兲. “Nonlinear seismic ground
Bazzurro, P., and Cornell, A. 共2004a兲. “Ground-motion amplification in response analysis: code usage protocols and verification against ver-
nonlinear soil sites with uncertain properties.” Bull. Seismol. Soc. Am., tical array data.” Geotechnical Engineering & Soil Dynamics IV, D.
94共6兲, 2090–2109. Zeng, M. T. Manzari, and D. R. Hiltunen 共eds兲, ASCE Geotechnical
Bazzurro, P., and Cornell, A. 共2004b兲. “Nonlinear soil-site effects in Special Publication No. 181, 1–24.
probabilistic seismic-hazard analysis.” Bull. Seismol. Soc. Am., 94共6兲, Toro, G., Silva, W., McGuire, R., and Hermann, R. 共1992兲. “Probabilistic
2110–2123. seismic hazard mapping of the Mississippi embayment.” Seismol. Res.
Bommer, J. J., and Acevedo, A. B. 共2004兲. “The use of real earthquake Lett., 63共3兲, 449–475.
accelerograms as input to dynamic analysis.” J. Earthquake Eng., Toro, G. R. 共1995兲 “Probabilistic models of site velocity profiles for
8共1兲, 43–91. generic and site-specific ground-motion amplification studies.” Tech-
Chang, S. W.-Y. 共1996兲. “Seismic response of deep stiff soil deposits.” nical Rep. No. 779574, Brookhaven National Laboratory, Upton, N.Y.
Ph.D. thesis, Univ. of California, Berkeley, Calif. Trent, W. L. 共2008兲 “The influence of soil property variability on
Costantino, C., Heymsfield, E., and Gu, Y. 共1991兲. “Site specific esti- equivalent-linear seismic site response analysis.” MS thesis, Univ. of
mates of surface ground motions for the K-reactor site, Savannah Texas, Austin, Tex.
View publication stats Downloaded 24 Nov 2010 to 146.6.92.162. Redistribution subject to ASCE license or copyright. Visit http://www.ascelibrary.org