2016 Zhao Subduction Interface

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Bulletin of the Seismological Society of America, Vol. 106, No. 4, pp. 1518–1534, August 2016, doi: 10.

1785/0120150034

Ground-Motion Prediction Equations for Subduction Interface


Earthquakes in Japan Using Site Class and Simple
Geometric Attenuation Functions
by John X. Zhao, Xuan Liang, Fei Jiang, Hao Xing, Min Zhu, Ruibin Hou,
Yingbin Zhang, Xiaowen Lan, David A. Rhoades, Kojiro Irikura,
Yoshimitsu Fukushima, and Paul G. Somerville

Abstract We presented a set of ground-motion prediction equations (GMPEs) for the


horizontal components of strong-motion records from subduction interface events in
Japan. We assembled and processed in a consistent manner a large set of strong-motion
records from reliably identified subduction interface events up to the end of 2012. The
GMPEs were based on a set of simple geometric attenuation functions, and site class was
based on site period as the site parameter. We adopted a bilinear magnitude-scaling func-
tion hinged at M w 7.1 and found that the magnitude-scaling rate for large events is much
smaller than that for smaller events. To reliably determine the magnitude-scaling rate for
events with M w ≥ 7:1, we used a set of strong-motion records obtained since 1968 to
increase the number of records from large events. A small number of strong-motion
records are from recording stations with inferred site classes using the response spectral
ratio of the horizontal-to-vertical components or a geological description of the surface
soil layers. The effect of site information quality for subduction interface events was
examined using a goodness-of-fit parameter from a dataset with or without the sites
having an inferred site class. The site information quality made a significant difference
at all spectral periods, because the model fit improved significantly when the sites with
inferred classes were excluded. We modeled the effect of volcanic zones using an an-
elastic attenuation coefficient applied to the horizontal portion of the seismic-wave travel
distance within a set of assumed volcanic zones. The within-event residuals were ap-
proximately separated into within-site and between-site components, and the correspond-
ing standard deviations were calculated using a random effects model. The between-site
standard deviations vary significantly among site classes and with spectral periods.

Online Material: Figures showing distributions of between-event and within-event


residuals.

Introduction
Ground-motion prediction equations (GMPEs) are an (2006); Kanno et al. (2006); and Ghofrani and Atkinson
important component in earthquake engineering and prob- (2014). The model by Kanno et al. (2006) was also based
abilistic seismic-hazard analyses. In the last decade or so, on strong-motion records from Japan, but earthquake cat-
many modern GMPEs have been published (e.g., Allen and egories were not used as a model parameter. Abrahamson
Wald, 2009), including the Next Generation Attenuation et al. (2016) developed the most recent model for strong-
(NGA) models mainly based on strong-motion records from motion records from global subduction earthquakes, in-
California and supplemented by shallow crustal records from cluding 2590 records from 63 subduction slab events and
Taiwan, Japan, and Turkey. 953 records from 43 subduction interface (referred to as in-
For subduction zones, a relatively small number terface hereafter) events. Their dataset for interface events
of GMPEs have been developed, including Atkinson and contains far fewer strong-motion records than the dataset
Boore (2003); Zhao, Zhang, et al. (2006); McVerry et al. used in the present study.

1518
GMPEs for Subduction Interface Earthquakes in Japan 1519

Table 1 classified the earthquakes since 1996 in Japan and demon-


Number of Records in Each Site Class for Two strated that the classification they adopted led to the best
Datasets from Japan goodness of fit for the GMPEs using depth-scaled geometric
SC I SC II SC III SC IV Total attenuation functions. The earthquake categories assigned to
Dataset 1 1563 995 360 656 3574
the events used in the present study can be found in the elec-
Dataset 2 1494 786 284 547 3111 tronic supplement to the article by Zhao, Zhou, et al. (2015).
Zhao and Rhoades (2014), Zhao, Zhou, et al. (2016),
and Zhao, Jiang, et al. (2016) showed that the exclusion of
strong-motion records from stations with inferred site class
Zhao, Zhang, et al. (2006) showed that interface events,
leads to improved model fit in terms of maximum log like-
usually with a reverse or thrust focal mechanism and a maxi-
lihood (MLL) for shallow crustal and upper-mantle events
mum depth of 50 km, produce strong ground motions with
and for slab events. In the present study, we will test if the
different frequency content from those of shallow crustal or
exclusion of subduction interface records from sites with
subduction slab events. In their study, all types of events
inferred site class improves the model fit.
were grouped together, and the same site terms were used
We use site class based on site period as the site term,
for all events. A constant term, a linear magnitude term, and
and the site classes are defined in Table 1, together with the
a magnitude squared term were used for the interface events.
approximate National Earthquake Hazards Reduction Pro-
Zhao, Zhang, et al. (2006) used a common depth term and gram site classes (Building Seismic Safety Council [BSSC],
common site terms for each spectral period and for all three 2000) in Zhao, Hu, et al. (2015). We adopted the nonlinear
types of events. A total of 1508 records in their study were site terms based on Zhao, Hu, et al. (2015) and the extended
from interface events, and only a small number of recording nonlinear site terms from Zhao, Zhou, et al. (2016).
stations had a measured shear-wave velocity profile down Zhao and Xu (2012) investigated the magnitude-scaling
to engineering bedrock. Most stations were classified using rates for large interface earthquakes with a moment magnitude
response spectral ratios between the horizontal-to-vertical over 6.5, using strong-motion data from interface earthquakes
(H/V) components as described by Zhao, Irikura, et al. (2006). in Japan. Their study showed that the magnitude-scaling rates
Zhao, Zhang, et al. (2006) used the strong-motion records in were relatively small for large interface earthquakes with a
Japan up to the end of 2003, and many high-quality records moment magnitude over 7.0 compared with those for smaller
have been obtained since then. events. In the present study, we used a bilinear magnitude-
A major disadvantage in the Zhao, Zhang, et al. (2006) scaling function hinged at M w 7.1 as presented by Zhao
study is that the model standard deviations are the same (2014), Zhao, Jiang, et al. (2016), and Zhao, Zhou, et al.
for all three earthquake categories. This feature can have a (2016). The hinge in the bilinear model reduces the effect
significant negative effect on a probabilistic seismic-hazard of the relatively large number of events with an M w ≤ 7:1 on
analysis, if the standard deviations differ among earthquake the magnitude-scaling rate for a relatively small number of
categories. For example, the interface events may have a large earthquakes.
smaller model standard deviation than that for subduction Following the study by Zhao (2010), we used the MLL
slab events, and the seismic hazard from interface events may in the present study, rather than the model standard devia-
be overestimated, whereas the hazards from slab events may tions, as the indicator of goodness of fit for a GMPE in a
be underestimated. similar manner to that used by Zhao, Zhou, et al. (2016). The
In the last two decades, a large number of high-quality advantages of using MLL were described in the Zhao (2010)
strong-motion records from interface events have been ob- and Zhao and Rhoades (2014) studies.
tained, and now there are enough records to establish sepa-
rate GMPEs for each category of events (Zhao and Rhoades,
Strong-Motion Dataset
2014). Zhao, Zhou, et al. (2016) presented a GMPE for shal-
low crustal and upper-mantle events and Zhao, Jiang, et al. In this study, we constructed two datasets. We used the
(2016) presented a GMPE for subduction slab events. The first dataset to derive the magnitude-scaling rates for earth-
models derived for interface events are presented in this quakes with an Mw ≥ 7:1. To have as many large events as
article. possible, we used 76 interface earthquakes in Japan since
Zhao, Zhou, et al. (2015) identified interface earth- 1968, including the large earthquakes from the dataset used
quakes that occur at the interface between the crust and by Zhao, Zhang, et al. (2006). Table 1 shows the number of
the subducting slab or between the mantle wedge and the records in each site class, and 3574 records in total were
subducting slab at a depth of 50 km or less. The other criteria used. The first dataset has 1222 records from 13 interface
are that the focal depth is within 5 km of the subduction earthquakes with an M w ≥ 7:1. All records from earth-
interface boundary defined by Slab1.0 (Hayes et al., 2012), quakes before 1996 and some records from K-NET are from
the focal mechanism must be reverse or thrust, and the dip stations with site classes inferred from geological descrip-
angle for one of the nodal planes is within 15° of the dip tions or H/V response spectral ratios. The inferred site
angle of the subduction interface. Zhao, Zhou, et al. (2015) classes are unlikely to be as reliable as those from the
1520 J. X. Zhao, et al.

We have fault rupture models for a number of interface


earthquakes, and the references for these models were pre-
sented in the electronic supplement to the article of Zhao,
Zhou, et al. (2016).

GMPE Model Forms


Zhao (2010) found that the anelastic attenuation rate for
shallow interface events (0–25 km) is larger than that for re-
cords from deep interface events with a depth over 25 km. In
the present study, we also tested whether the two groups of
events may need different source terms because of the differ-
ent material properties at the shallow and deep interfaces
along the subducting plate. In our dataset for the final model,
798 records are from 11 shallow events. For interface events
with a focal depth of 25 km or less, the following function
was used for the source effects

fmintS mi ; hi   bint hi  γ intS


EQ-TARGET;temp:intralink-;df1;313;533


cintS mi if mi ≤mc
 ; 1
cintS mc  dint mi − mc  if mi >mc

in which γ intS is the constant term for shallow interface


events, bint , cintS , and dint are the regression coefficients for
fault-top depth h and magnitude terms, respectively. Moment
magnitude is denoted by m instead of M w for simplicity, and
mc is a magnitude constant; intS refers to the shallow inter-
Figure 1. (a) The distribution of earthquakes used in the present face events; i refers to the ith event in the dataset.
study with respect to fault depth and moment magnitude and (b) dis- For interface events with a focal depth over 25 km, the
tribution of strong-motion records with respect to source distance following function was used for the source effect
and moment magnitude. The color version of this figure is available
only in the electronic edition.
fmintD mi ; hi   bint hi  γ intS
EQ-TARGET;temp:intralink-;df2;313;366


cintD mi if mi ≤mc
measured shear-wave velocity profiles. However, we expect  ;2
that the imprecision associated with site class for a small cintD mc  dint mi − mc  if mi >mc
number of sites may not significantly affect the estimated
magnitude term for large events. The magnitude term for in which intD means that the term is associated with the deep
events with an M w ≥ 7:1 was, therefore, determined using interface events. Coefficient cintD is the magnitude-scaling rate
the first dataset. We then constructed the second dataset by for events with a magnitude of mc or less.
removing the records from sites with an inferred site class In equations (1) and (2), we used a bilinear magnitude-
in the first dataset. Table 1 shows the number of records in scaling function hinged at M w  mc based on the results of
each site class for the second dataset. A total of 463 records Zhao and Xu (2012), Zhao (2014), and Zhao and Rhoades
were excluded. The second dataset was used to develop (2014), and mc 7.1 was selected for all interface events.
the final GMPE. Figure 1a shows the distribution of earth- For shallow interface earthquakes, the following model
quakes with respect to fault-top depth and magnitude, and was used:
Figure 1b shows the distribution of strong-motion records
loge yi;j   fmintS  gint loge ri;j   gintSL loge xi;j
EQ-TARGET;temp:intralink-;df3a;313;184

with respect to source distance and magnitude for the sec-


ond dataset.  200:0  eintS xi·j  eVint xVi;j  γ int
The 2011 Tohoku earthquake with Mw 9 is the largest
well-recorded earthquake in seismological history. We used  loge A  ξi;j  ηi 3a
413 records from this event and a similar number of records
from the foreshock and aftershocks of the Tohoku earth-  
μi;j
quake. These records provide an invaluable constraint on the EQ-TARGET;temp:intralink-;df3b;313;113 loge  ξi;j  ηi ; 3b
magnitude-scaling rate. yi;j
GMPEs for Subduction Interface Earthquakes in Japan 1521

in which j denotes the jth record in the ith event. Variable y Coefficient gintDL is the large distance geometric attenuation
is for peak ground acceleration (PGA) or 5% damped rate for deep events. Distance constant xinto  10 was also
response spectrum in units of the acceleration due to gravity, used to avoid magnitude–distance oversaturation. For Japan,
and μ denotes the recorded PGA or spectrum. Variable e the closest distance to a deep interface event is approximately
denotes anelastic attenuation rate, x denotes the shortest over 30 km, and oversaturation would occur only when the
distance from a recording station to the fault plane if a fault distance is less than about 10 km if xinto is not used. This term
model is available (otherwise the hypocentral distance), and was retained mainly for ease of model coefficient smoothing
gint denotes geometric attenuation rate. The term gintSL with respect to spectral periods.
used together with gint is the large-distance geometric attenu- Zhao, Zhou, et al. (2015) presents the rock-site factor
ation rate that reduces the geometric attenuation rate over a AmSC I  expS1N . The maximum site amplification ratio
distance of about 50 km. The use of gintSL also leads to a is defined by
negative value for anelastic attenuation rate eintS for all spec-
tral periods. Superscript v indicates association with the vol- EQ-TARGET;temp:intralink-;df7a;313;601 AN max k  AmSC I for SC I sites k  1 7a
canic path. Anelastic attenuation rate eV is applied to the
horizontal distance passing through volcanic zones, denoted
AN max k  AmSC I expSk 
EQ-TARGET;temp:intralink-;df7b;313;569

by xV as presented in figure 3 in Zhao, Zhou, et al. (2016).


Parameter A is the site amplification ratio and contains both for SC II; III; and IV sites k  2; 3; 7; 7b
linear and nonlinear site terms that will be described later.
Random variable ξ is the within-event residual with an aver- in which Sk is the elastic site term, with k  2 for SC II sites,
age value of 0.0 and a standard deviation of σ, referred to as k  3 for SC III sites, and k  4 for SC IV sites for shallow
the within-event standard deviation. Random variable η is the interface earthquakes. For deep events, k  5 for SC II sites,
between-event residual with an average value of 0.0 and a k  6 for SC III sites, and k  7 for SC IV sites. The rock-site
standard deviation τ, referred to as between-event standard term S1N was originally from Zhao, Zhou, et al. (2016) and
deviation. The total residuals are defined in equation (3b). Zhao, Hu, et al. (2015) but was smoothed in the present study.
The distance used for geometric spreading is defined by The nonlinear amplification ratio is given by

EQ-TARGET;temp:intralink-;df4;55;433 ri;j  xinto  xi;j  expc1  c2 Cm  4 EQ-TARGET;temp:intralink-;df8;313;433

loge A  loge AN max 


loge SαMR  β − loge β
 − loge Amax  8
mi if mi ≤ Cmax loge SαReffC  β − loge β
Cm  : 5
Cmax if mi > Cmax
EQ-TARGET;temp:intralink-;df5;55;403

  
loge AN max  loge SαReffC  β − loge SF lnβ
EQ-TARGET;temp:intralink-;df9;313;356

Parameter c1 is a regression coefficient determined by Zhao,


SNC  exp
Zhou, et al. (2015). Parameter c2  1:151 was used for all loge Amax 
spectral periods, and the physical base was described by 1
α
Zhao, Zhou, et al. (2016). We introduced a distance constant −β 9
xinto  10:0 km to avoid magnitude–distance oversaturation,
for which the ground motion at the source decreases with SNC
SMR  SReff f 10
increasing magnitude. This aspect was discussed in the study SReffC SR
EQ-TARGET;temp:intralink-;df10;313;288

by Zhao and Rhoades (2014) and Zhao, Zhou, et al. (2016).


Following their studies and the Zhao, Jiang, et al. (2016)
study, we also used a maximum magnitude Cmax in equa-
EQ-TARGET;temp:intralink-;df11;313;258 SReff  SR I mf 11
tion (5) and we selected Cmax  mc . Discussions on the se-
lection of Cmax can be found in Zhao, Zhou, et al. (2016). EQ-TARGET;temp:intralink-;df12;313;230 SReffC  SRC I mf : 12
As discussed earlier, Zhao (2010) found that the attenu-
ation rate for shallow interface earthquakes was larger than Parameters α  2:0 and β  0:6 were used for all periods,
that for deep interface events because the materials in the shal- and Amax , SRC , and I mf are presented in Zhao, Zhou, et al.
low crust close to the interface and the mantle wedge may (2016), and the procedure for deriving equations (8)–(12) is
have different properties. We found that an anelastic attenua- presented by Zhao, Hu, et al. (2015).
tion term was not necessary for deep interface earthquakes. Adopting different elastic site terms for shallow and
The following equation was used for deep interface records: deep events is purely for model term smoothing, and these
terms, which correspond to the same site class, are sta-
loge yi;j   fmintD  gint loge ri;j 
EQ-TARGET;temp:intralink-;df6;55;127

tistically the same. Because some of the site terms from the
 gintDL loge xi;j  200:0  eVint xVi;j 1D model presented by Zhao, Zhou, et al. (2016) and Zhao,
Hu, et al. (2015) were not smoothed with respect to spectral
 γ int  loge A  ξi;j  ηi : 6 periods, adjustment factors fSR presented in Table 2 were
1522 J. X. Zhao, et al.

used to achieve a broadly smoothed spectrum as described by


Zhao, Jiang, et al. (2016). We adopted deamplification ratios
or a rock-site factor AmSC I  expS1N  for rock sites as
described by Zhao, Zhou, et al. (2016) with minor modifi-
cations for smoothing.
Using Abrahamson and Youngs (1992) algorithm, Zhao
and Rhoades (2014), Zhao, Zhou, et al. (2016), and Zhao,
Jiang, et al. (2016) approximately separated the random var-
iable ξ into a within-site component ξS and a between-site
component ηS . The within-site component ξsk;n has a zero
mean and a within-site standard deviation of σ S . The between-
site component ηsk has a zero mean and a between-site stan-
dard deviation of τS . The within-site, between-site, and total
site standard deviations for interface events are presented in
this article. Figure 2. The effect of site information quality—the increase in
Distance xVi;j is the horizontal distance of the volcanic weighted maximum log likelihood (MLL) without the strong-
path calculated from the straight line between the earthquake motion records from stations with an inferred site class. The color
source and a site passing through volcanic zones described in version of this figure is available only in the electronic edition.
figure 3 of Zhao, Zhou, et al. (2016). Zhao, Zhou, et al.
(2016) also described the limits on the minimum and maxi- that from the first model, we can conclude that the measured
mum volcanic distance for xV . shear-wave velocity down to bedrock indeed improves the
model fit. Figure 2 shows that the improvement is significant
at all spectral periods, with the increase in the weighted MLL
Site Information Effect being over 25 at all spectral periods except for 5.0 s, and the
In a manner similar to that used by Zhao, Jiang, et al. largest increase is over 160. Overall, the exclusion of the re-
(2016) and Zhao, Zhou, et al. (2016), we established two sets cords from sites with an inferred site class improves the model
of GMPEs for the interface events with identical functional fit, and we present the model parameters from the GMPEs
forms. We used the weighted MLL described by Zhao, Zhou, based on the second dataset. The improved model fit sug-
et al. (2016) from the two models to identify the effect of gests the importance of measured site shear-wave velocity
excluding the records from sites with an inferred site class. profiles down to engineering bedrock. PGA is plotted at
If the weighted MLL increases significantly after the records 0.02 s spectral period for the convenience of using a loga-
from sites with inferred site class are excluded, compared with rithm scale in spectral period.

Table 2
Adjustment Factors
Site Class Site Class
Number T (s) I II III IV Number T (s) I II III IV

1 PGA 1.0 1.0 1.165 1.0 20 0.35 0.0 0.701 0.473 1.008
2 0.01 1.0 1.0 0.944 1.0 21 0.4 0.0 0.708 0.487 1.007
3 0.02 1.0 1.0 1.012 1.0 22 0.45 0.0 0.737 0.511 0.981
4 0.03 1.0 1.0 1.100 1.000 23 0.5 0.0 0.748 0.536 0.990
5 0.04 1.0 0.843 0.959 0.557 24 0.6 0.0 0.728 0.540 1.016
6 0.05 1.0 0.663 0.889 0.543 25 0.7 0.0 0.634 0.477 1.022
7 0.06 1.0 0.841 0.946 0.574 26 0.8 0.0 0.0 0.0 1.023
8 0.07 1.0 1.029 1.006 0.648 27 0.9 0.0 0.0 0.0 0.997
9 0.08 1.0 1.235 1.065 0.721 28 1.0 0.0 0.0 0.0 0.948
10 0.09 1.0 1.144 1.093 0.809 29 1.25 0.0 0.0 0.0 0.802
11 0.1 1.0 1.092 1.077 1.015 30 1.5 0.0 0.0 0.0 0.0
12 0.12 0.0 0.945 1.036 0.972 31 2 0.0 0.0 0.0 0.0
13 0.14 0.0 0.624 0.895 0.967 32 2.5 0.0 0.0 0.0 0.0
14 0.15 0.0 0.577 0.860 0.963 33 3 0.0 0.0 0.0 0.0
15 0.16 0.0 0.545 0.822 0.953 34 3.5 0.0 0.0 0.0 0.0
16 0.18 0.0 0.527 0.743 0.967 35 4 0.0 0.0 0.0 0.0
17 0.2 0.0 0.546 0.663 1.005 36 4.5 0.0 0.0 0.0 0.0
18 0.25 0.0 0.596 0.487 1.045 37 5 0.0 0.0 0.0 0.0
19 0.3 0.0 0.623 0.447 1.035

PGA, peak ground acceleration.


GMPEs for Subduction Interface Earthquakes in Japan 1523

Methodology Used in the Regression Analyses


The same strategy of determining model parameters
from Zhao, Zhou, et al. (2016) for crustal and upper-mantle
events was adopted in the present study. The steps are
summarized below.
1. Derive a set of GMPEs using all terms as presented in
equations (1)–(8) as well as others not presented in this
article, including the magnitude-squared term, the geo-
metric attenuation rate with and without magnitude
dependence, and separate depth terms for shallow and
deep events.
2. Test all model coefficients and delete those that are not Figure 3. Magnitude-scaling coefficient dint for large events
statistically significant. and the ratio dint =cint . The color version of this figure is available
only in the electronic edition.
3. Derive the second model from the second dataset, after
the magnitude-scaling coefficient for large events dint
was determined using the first dataset in step 1. fault-top depth to the between-event residuals and a linear
4. Test each term in equations (1)–(8) in turn, starting from function of magnitude and source distance to the within-
the term that has the smallest ratio of the mean value over event residuals for each spectral period. A t-test was carried
its standard deviation. If the term is not statistically out on each coefficient in the linear functions to ensure that
significant, this term can be removed. the coefficient was not statistically significant at a 5% sig-
5. If a term is statistically significant at only a very small nificance level for most terms and a 10% significance level
number of spectral periods, then it is set to zero. for the last couple of terms during the smoothing process.
6. For a term that is statistically significant at a number of
consecutive spectral periods, we interpolate or extrapo- Model Coefficients for the Final GMPE
late the values of this term using a polynomial function
of loge T. The spectral period for PGA is set as 0.005 s Figure 3 shows the values of dint and the ratio of dint =cint .
for the smoothing process. The magnitude-scaling rates for events with an M w > 7:1 are
7. For the linear site terms, we adopted different values for smaller than those for the smaller events. PGA is plotted at
shallow interface events at some spectral periods from 0.005 s spectral period for the convenience of using a loga-
those for the deep interface model purely to allow the rithm scale in spectral period. We selected a constant dint
final model to produce a broadly smoothed spectrum for up to 0.6 s, and then dint increases linearly with increasing
shallow interface events. spectral periods in a natural logarithm scale. Tables 3–5
8. After the model terms were selected, we smoothed γ int present the smoothed coefficients in the model described in
first, and then we used γ int as a constant to carry out a equations (1)–(6), and Table 6 presents the between-event and
regression analysis. The second parameter smoothed was within-event standard deviations. Table 4 also presents the
cintD , and we smoothed S  γ int  mi cintD with mi  7:0. rock-site term S1N from those by Zhao, Zhou, et al. (2016)
Then the magnitude-scaling term could be calculated from with minor adjustment, so as to obtain a visually smoothed
cintD  S − γ int =mi . Assuming that γ intS was to be rock-site spectrum.
smoothed next, we took the smoothed γ int and cintD as con- Figure 4 shows the anelastic attenuation rates for shal-
stants and carried out a regression analysis to determine all low interface events and the anelastic attenuation rate that
other terms. Using the new value for γ intS, we smooth applies to the volcanic path for all interface events. At short
S  γ intS  γ int  mi cintD with mi  7:0. Then γ intS was periods less than 0.3 s, the absolute value for anelastic attenu-
ation rate for volcanic path is larger than that for the shallow
calculated from γ intS  S − cintD mi − γ int . In this manner,
interface event. For spectral periods over 0.3 s, the absolute
all coefficients in equations (1)–(6) were smoothed except
value for the anelastic attenuation rate for the shallow inter-
for the site terms for the shallow interface events, which
face events is much larger than that for the volcanic path.
were adjusted manually to achieve a visually smoothed
Figure 5 presents the coefficients for the linear site terms
spectrum for each site class.
for SC II, SC III, and SC IV sites and deep interface events.
However, the smoothing of the model terms does not guar- The linear site terms for shallow interface events are almost
antee a smoothed spectrum over all spectral periods and the same. All linear site terms have a trough at 0.05–0.1 s,
all parameter ranges, especially at the limits of magnitude, because this spectral period range covers the average site
distance, and depth range of the dataset. During the smooth- period of the SC I site, and the exponential of linear site term
ing process, MLL on average did not decrease by more than expSk  represents the amplification ratios with respect to SC
1.0 for most terms and periods. After all model parameters I sites. The linear site terms for the SC II sites have a maxi-
were smoothed, we fitted a linear function of magnitude and mum value at around 0.3 s, which is the middle period for SC
1524 J. X. Zhao, et al.

Table 3
Model Parameters
T (s) c1 c2 cintD cintS dint γ intS bint gint

PGA −5.301 1.151 1.0997 1.3148 0.553 −3.8953 0.0200 −2.0559


0.01 −5.288 1.151 1.0985 1.3174 0.553 −3.8953 0.0200 −2.0657
0.02 −5.276 1.151 1.0923 1.3192 0.553 −3.8953 0.0200 −2.1023
0.03 −5.268 1.151 1.1069 1.3410 0.553 −3.8953 0.0207 −2.1923
0.04 −5.263 1.151 1.1158 1.3805 0.553 −3.8953 0.0231 −2.2464
0.05 −5.259 1.151 1.1023 1.4325 0.553 −3.8953 0.0271 −2.2934
0.06 −5.255 1.151 1.0861 1.4624 0.553 −3.8953 0.0297 −2.3117
0.07 −5.253 1.151 1.0729 1.4712 0.553 −3.8953 0.0321 −2.3110
0.08 −5.250 1.151 1.0638 1.4643 0.553 −3.8946 0.0320 −2.2878
0.09 −5.248 1.151 1.0586 1.4470 0.553 −3.9018 0.0297 −2.2468
0.10 −5.246 1.151 1.0567 1.4232 0.553 −3.9077 0.0279 −2.2041
0.12 −5.243 1.151 1.0605 1.3683 0.553 −3.9164 0.0247 −2.1201
0.14 −5.240 1.151 1.0714 1.3156 0.553 −3.9227 0.0212 −2.0434
0.15 −5.239 1.151 1.0786 1.2928 0.553 −3.9253 0.0195 −2.0109
0.16 −5.237 1.151 1.0866 1.2732 0.553 −3.9275 0.0179 −1.9830
0.18 −5.235 1.151 1.1047 1.2483 0.553 −3.9313 0.0151 −1.9461
0.20 −5.233 1.151 1.1244 1.2372 0.553 −3.9345 0.0126 −1.9270
0.25 −5.229 1.151 1.1769 1.2239 0.553 −3.9407 0.0077 −1.8988
0.30 −5.226 1.151 1.2297 1.2285 0.553 −3.9455 0.0044 −1.8914
0.35 −5.223 1.151 1.2806 1.2422 0.553 −3.9494 0.0022 −1.8930
0.40 −5.221 1.151 1.3287 1.2608 0.553 −3.9527 0.0000 −1.8953
0.45 −5.218 1.151 1.3739 1.2819 0.553 −3.9556 0.0000 −1.9058
0.50 −5.216 1.151 1.4163 1.3043 0.553 −3.9580 0.0000 −1.9147
0.60 −5.213 1.151 1.4931 1.3502 0.553 −3.9618 0.0000 −1.9274
0.70 −5.210 1.151 1.5607 1.3952 0.560 −3.9648 0.0000 −1.9345
0.80 −5.208 1.151 1.6206 1.4382 0.580 −3.9673 0.0000 −1.9374
0.90 −5.206 1.151 1.6739 1.4788 0.602 −3.9696 0.0000 −1.9373
1.00 −5.204 1.151 1.7217 1.5169 0.622 −3.9720 0.0000 −1.9351
1.25 −5.200 1.151 1.8219 1.6015 0.667 −3.9795 0.0000 −1.9247
1.50 −5.196 1.151 1.9008 1.6728 0.705 −3.9905 0.0000 −1.9119
2.00 −5.191 1.151 2.0148 1.7837 0.768 −4.0265 0.0000 −1.8886
2.50 −5.187 1.151 2.0889 1.8624 0.820 −4.0830 0.0000 −1.8725
3.00 −5.183 1.151 2.1357 1.9171 0.863 −4.1594 0.0000 −1.8635
3.50 −5.181 1.151 2.1625 1.9532 0.902 −4.2542 0.0000 −1.8597
4.00 −5.178 1.151 2.1739 1.9745 0.935 −4.3658 0.0000 −1.8595
4.50 −5.176 1.151 2.1730 1.9836 0.966 −4.4927 0.0000 −1.8615
5.00 −5.174 1.151 2.1620 1.9826 0.994 −4.6331 0.0000 −1.8645

II sites. The linear site terms for the SC III site have a maxi- (digitized from their fig. 6). The between-event standard de-
mum value at about 0.5 s, which is also the middle period for viations of the present study are smaller than those from
SC III sites, whereas the linear site term for SC IV sites is Abrahamson et al. (2016) at most spectral periods. Figure 7a
nearly constant in a 0.8–2.5 s period range before it decreases also shows that the total standard deviations in the present
with increasing spectral periods. study are larger than those from the Abrahamson et al.
Ⓔ Figures S1–S14, available in the electronic supple- (2016) study in a 0.6–2.5 s spectral period range, with the
ment to this article, show the between-event and within-event largest difference being about 13% at 1.0 s spectral periods,
residuals from the final model. Statistical tests were per- whereas the total standard deviations from the present study
formed to make sure that the slope of the trend lines and the are smaller than those from Abrahamson et al. (2016) at the
average residuals are not statistically significant. other periods. Figure 7b shows that the within-event standard
Figure 6 shows the between-event, within-event, and to- deviations from the Abrahamson et al. (2016) study are
tal standard deviations. PGA has the smallest total standard smaller than those of the present study in a 0.15–2.5 s spec-
deviation of 0.669, and the largest total standard is 0.797 at tral period range, with the largest difference being 13% at
0.1 s. On average, the within-event standard deviations are 1.0 s. It is possible that the differences in this period band
about 1.54 times the between-event standard deviations. are caused by using site classes as the site term in the present
Figure 6 also shows the smoothed standard deviations from study, whereas Abrahamson et al. (2016) used V S30 , a shear-
the Abrahamson et al. (2016) study. Figure 7a compares the wave travel-time averaged shear-wave velocity of top 30 m
within-event and total standard deviations from the present soils, as the site term. However, this cannot be confirmed
study with those from the Abrahamson et al. (2016) study without comparing the between-site standard deviations
GMPEs for Subduction Interface Earthquakes in Japan 1525

Table 4
Model Parameters
T (s) gintLD gintLS eVint eintS γ int Rock-Site Factor

PGA 0.5454 1.1336 −0.01123 −0.00628 −4.4986 1.358


0.01 0.5498 1.1336 −0.01125 −0.00625 −4.4589 1.247
0.02 0.5617 1.1336 −0.01127 −0.00616 −4.2581 1.149
0.03 0.5789 1.1336 −0.01158 −0.00572 −3.9180 1.097
0.04 0.4933 0.9881 −0.01203 −0.00532 −3.1142 1.065
0.05 0.4910 0.9044 −0.01256 −0.00503 −2.7604 1.037
0.06 0.5085 0.8877 −0.01312 −0.00528 −2.6409 1.038
0.07 0.5275 0.9049 −0.01359 −0.00569 −2.6562 1.050
0.08 0.5460 0.9421 −0.01382 −0.00619 −2.7527 1.103
0.09 0.5631 0.9865 −0.01393 −0.00673 −2.8992 1.192
0.10 0.5762 1.0355 −0.01395 −0.00718 −3.0770 1.277
0.12 0.5926 1.1353 −0.01381 −0.00793 −3.4828 1.400
0.14 0.6098 1.2342 −0.01351 −0.00853 −3.9161 1.525
0.15 0.6196 1.2813 −0.01333 −0.00879 −4.1348 1.578
0.16 0.6308 1.3266 −0.01312 −0.00902 −4.3524 1.625
0.18 0.6620 1.4113 −0.01269 −0.00927 −4.7803 1.705
0.20 0.6998 1.4885 −0.01223 −0.00942 −5.1944 1.768
0.25 0.7845 1.6521 −0.01108 −0.00959 −6.1580 1.868
0.30 0.8594 1.7813 −0.00998 −0.00952 −7.0200 1.917
0.35 0.9234 1.8844 −0.00898 −0.00933 −7.7915 1.938
0.40 0.9801 1.9676 −0.00808 −0.00911 −8.4955 1.944
0.45 1.0222 2.0355 −0.00727 −0.00888 −9.1135 1.945
0.50 1.0587 2.0914 −0.00656 −0.00866 −9.6852 1.942
0.60 1.1180 2.1764 −0.00534 −0.00824 −10.6895 1.928
0.70 1.1630 2.2360 −0.00437 −0.00787 −11.5460 1.911
0.80 1.1973 2.2783 −0.00359 −0.00755 −12.2872 1.892
0.90 1.2236 2.3085 −0.00296 −0.00726 −12.9363 1.873
1.00 1.2437 2.3299 −0.00244 −0.00700 −13.5100 1.853
1.25 1.2725 2.3585 −0.00153 −0.00644 −14.6903 1.799
1.50 1.2854 2.3665 −0.00097 −0.00597 −15.6030 1.740
2.00 1.2883 2.3554 −0.00043 −0.00518 −16.9001 1.620
2.50 1.2773 2.3311 −0.00023 −0.00451 −17.7366 1.508
3.00 1.2605 2.3041 −0.00016 −0.00393 −18.2714 1.416
3.50 1.2411 2.2779 0.0 −0.00344 −18.5926 1.347
4.00 1.2203 2.2537 0.0 −0.00302 −18.7547 1.304
4.50 1.1986 2.2316 0.0 −0.00267 −18.7935 1.285
5.00 1.1763 2.2115 0.0 −0.00240 −18.7339 1.267

from the two models. The differences can also be caused by periods, and SC IV sites have the largest standard deviations
using different datasets. at long periods. The total site standard deviations appear
Table 5 presents the within-site and between-site standard to have the largest value at the average site period of each
deviations. Figure 8a shows the between-site standard devia- site class.
tions, and Figure 8b shows the within-site standard deviations. The between-event residuals show that the overall mod-
At short periods up to 0.3 s, the between-site standard devia- eling for the 2011 Mw 9.0 Tohoku earthquake and the 2003
tions for SC I and SC II sites are considerably larger than those Mw 8.3 off Tokachi-Oki earthquake is reasonable (see Ⓔ
from the other two site classes. At long periods over 0.7 s, SC Figs. S1–S14); this result validates the selected bilinear mag-
IV sites have the largest between-site standard deviations. In nitude scaling for the interface events. Figure 10 shows the
general, the differences in the between-site standard deviations residual factor of the between-event residuals, exp−ηi , for
vary significantly from one site class to another. The within- each spectral period together with the factor for mean ±1
site standard deviations in Figure 8b are similar for all site between-event standard deviation τ, that is, expτ for
classes and do not vary with spectral periods as much as the these two great subduction interface earthquakes. On average
between-site standard deviations do. Further improvement in across all spectral periods, the response spectra from the
site effect modeling may lead to a reduction in the between- 2003 Mw 8.3 Tokachi-Oki earthquake underestimated by a
site standard deviations. factor of 0.841, and the long-period spectrum at 5.0 s was
Figure 9 compares the total site standard deviations from underestimated significantly, exceeding a mean minus one
four site classes with the within-event standard deviations. between-event standard deviation. For the 2011 M w 9.0 To-
SC III sites have the smallest standard deviations at most hoku earthquake, the spectra on average were overestimated
1526 J. X. Zhao, et al.

Table 5
Site Terms for Deep Subduction Events and Standard Deviations
T (s) S2 S3 S4 S5 S6 S7 σ τ σT

PGA 0.3129 −0.0043 0.2284 0.3129 −0.0043 0.2284 0.553 0.378 0.670
0.01 0.3085 −0.0112 0.2231 0.3085 −0.0112 0.2231 0.553 0.377 0.670
0.02 0.2930 −0.0217 0.2089 0.2930 −0.0217 0.2089 0.553 0.383 0.672
0.03 0.2287 −0.1129 0.1331 0.2287 −0.0829 0.1631 0.554 0.397 0.682
0.04 0.1632 −0.1887 0.0696 0.1632 −0.1887 0.0596 0.564 0.428 0.708
0.05 0.1213 −0.2283 0.0285 0.1213 −0.2283 0.0085 0.569 0.463 0.734
0.06 0.1235 −0.2192 0.0007 0.1235 −0.2192 0.0007 0.583 0.487 0.759
0.07 0.1397 −0.1901 −0.0095 0.1397 −0.1901 0.0055 0.602 0.501 0.783
0.08 0.1639 −0.1550 −0.0049 0.1639 −0.1550 0.0141 0.614 0.501 0.793
0.09 0.2050 −0.1125 0.0305 0.2050 −0.1125 0.0405 0.625 0.495 0.797
0.10 0.2445 −0.0751 0.0608 0.2445 −0.0751 0.0708 0.637 0.478 0.796
0.12 0.3228 0.0150 0.1423 0.3228 0.0150 0.1403 0.646 0.453 0.789
0.14 0.4012 0.0970 0.2270 0.4012 0.0970 0.2020 0.654 0.412 0.773
0.15 0.4362 0.1459 0.2576 0.4362 0.1359 0.2326 0.659 0.404 0.773
0.16 0.4674 0.1879 0.3075 0.4674 0.1729 0.2625 0.664 0.398 0.774
0.18 0.5120 0.2515 0.3597 0.5120 0.2415 0.3197 0.672 0.387 0.776
0.20 0.5393 0.3030 0.4031 0.5393 0.3030 0.3731 0.678 0.382 0.778
0.25 0.5860 0.4269 0.5077 0.5860 0.4269 0.4877 0.659 0.365 0.753
0.30 0.6047 0.5162 0.5778 0.6047 0.5162 0.5778 0.640 0.348 0.729
0.35 0.6064 0.5695 0.6382 0.6064 0.5795 0.6482 0.634 0.360 0.729
0.40 0.6028 0.6237 0.7032 0.6028 0.6337 0.7132 0.627 0.354 0.720
0.45 0.5804 0.6581 0.7508 0.5804 0.6581 0.7508 0.620 0.363 0.719
0.50 0.5569 0.6867 0.7938 0.5569 0.6867 0.7938 0.612 0.364 0.712
0.60 0.5097 0.7122 0.8495 0.5097 0.7122 0.8495 0.612 0.379 0.720
0.70 0.4650 0.7124 0.8798 0.4650 0.7124 0.8798 0.624 0.393 0.738
0.80 0.4244 0.6994 0.8954 0.4244 0.6994 0.8954 0.628 0.396 0.742
0.90 0.3884 0.6800 0.9026 0.3884 0.6800 0.9026 0.628 0.397 0.743
1.00 0.3570 0.6583 0.9053 0.3570 0.6583 0.9053 0.632 0.404 0.750
1.25 0.2967 0.6193 0.9179 0.2967 0.6193 0.9179 0.635 0.404 0.752
1.50 0.2579 0.5829 0.9213 0.2579 0.5829 0.9213 0.643 0.392 0.753
2.00 0.2226 0.5262 0.9171 0.2226 0.5262 0.9171 0.634 0.382 0.741
2.50 0.2184 0.4872 0.9055 0.2184 0.4872 0.9055 0.619 0.393 0.733
3.00 0.2160 0.4570 0.8867 0.2160 0.4570 0.8867 0.599 0.385 0.712
3.50 0.2160 0.4281 0.8588 0.2160 0.4281 0.8588 0.580 0.376 0.692
4.00 0.2160 0.3952 0.8203 0.2160 0.3952 0.8203 0.568 0.377 0.682
4.50 0.2160 0.3549 0.7699 0.2160 0.3549 0.7699 0.552 0.376 0.668
5.00 0.2160 0.3047 0.7071 0.2160 0.3047 0.7071 0.563 0.374 0.676

by about 10% for all spectral periods, and this is a fraction of polynomial of source distance. Clearly, the distribution is
the mean  τ factor. biased at distances within about 70 km. The trend line is very
Figure 11a shows the within-event residuals for all similar to those fitted to the shallow crustal events in the Zhao
records from the Tohoku event, and the trend line suggests (2010) study, and the biased distribution was attributed to the
that the predicted PGAs within a distance of 100 km were Moho reflection effect in that study. The ground motion from a
underestimated. The trend line fitted to the residuals is a shallow interface event (within 25 km) is likely to be affected
second-order polynomial of source distance, and the trend line by Moho reflection, at least in theory. The trend lines in
can be approximated by the following conventional attenua- Figure 11b cannot be described by equation (13) adequately,
tion function and the multisegmented linear geometric attenuation functions
similar to that used by Zhao (2010) can be used to eliminate
EQ-TARGET;temp:intralink-;df13;55;184 loge yj   g loge rj   exj  c; 13
the bias.
in which g is the geometric attenuation coefficient, e is the For NGA models including Abrahamson and Silva
anelastic attenuation coefficient, and c is a constant. The (2008), Boore and Atkinson (2008), Campbell and Bozorgnia
biased residuals distribution within a distance range of 130 km (2008), and Chiou and Youngs (2008), magnitude-dependent
can be corrected by equation (13), with g  −0:280, e  0:0, standard deviations were used, and both between- and within-
and c  1:43. This means that the anelastic attenuation for event standard deviations decrease with increasing earthquake
the shallow interface events is adequate for this earthquake magnitude. The studies by Zhao, Zhou, et al. (2016) for
if equation (13) is used. Figure 11b shows the residuals shallow crustal and upper-mantle events and Zhao, Jiang, et al.
presented in Figure 11a, but the trend line is a third-order (2016) for subduction slab events found that standard
GMPEs for Subduction Interface Earthquakes in Japan 1527

Figure 5. Comparison of elastic site terms for three site classes


Figure 4. Comparison of anelastic attenuation rate with that for in the deep interface model. The color version of this figure is avail-
volcanic path plotted as a percentage. The color version of this fig- able only in the electronic edition.
ure is available only in the electronic edition.

Table 6
Within-Site and Between-Site Standard Deviations for Subduction Interface Earthquakes
SC I SC II SC III SC IV
Period (s) σS τS σ ST σS τS σ ST σS τS σ ST σS τS σ ST

PGA 0.388 0.419 0.571 0.421 0.424 0.597 0.393 0.376 0.544 0.458 0.346 0.574
0.01 0.388 0.420 0.571 0.421 0.424 0.598 0.393 0.377 0.545 0.416 0.348 0.543
0.02 0.386 0.425 0.574 0.420 0.425 0.598 0.394 0.376 0.544 0.415 0.352 0.544
0.03 0.383 0.444 0.586 0.420 0.422 0.595 0.397 0.386 0.553 0.415 0.355 0.546
0.04 0.386 0.474 0.611 0.420 0.434 0.604 0.394 0.393 0.557 0.421 0.362 0.555
0.05 0.386 0.486 0.620 0.419 0.440 0.608 0.399 0.384 0.554 0.419 0.375 0.562
0.06 0.386 0.500 0.632 0.419 0.449 0.614 0.398 0.379 0.550 0.420 0.388 0.571
0.07 0.389 0.531 0.658 0.420 0.466 0.628 0.412 0.377 0.558 0.420 0.410 0.587
0.08 0.397 0.553 0.681 0.424 0.489 0.647 0.406 0.374 0.552 0.419 0.416 0.591
0.09 0.402 0.559 0.688 0.425 0.521 0.672 0.406 0.381 0.557 0.419 0.435 0.604
0.1 0.406 0.562 0.693 0.424 0.540 0.687 0.403 0.409 0.574 0.423 0.469 0.632
0.12 0.411 0.564 0.697 0.424 0.537 0.684 0.415 0.480 0.635 0.426 0.517 0.670
0.14 0.413 0.556 0.692 0.439 0.560 0.712 0.407 0.440 0.599 0.438 0.507 0.670
0.15 0.414 0.552 0.690 0.440 0.574 0.724 0.410 0.427 0.592 0.430 0.504 0.663
0.16 0.416 0.548 0.688 0.446 0.573 0.726 0.404 0.429 0.589 0.437 0.496 0.661
0.18 0.423 0.534 0.681 0.454 0.557 0.719 0.411 0.436 0.600 0.455 0.488 0.667
0.2 0.426 0.528 0.679 0.469 0.532 0.710 0.414 0.437 0.602 0.456 0.481 0.663
0.25 0.424 0.493 0.650 0.468 0.551 0.723 0.410 0.398 0.572 0.452 0.448 0.637
0.3 0.422 0.471 0.632 0.473 0.509 0.695 0.398 0.421 0.579 0.439 0.442 0.623
0.35 0.413 0.474 0.629 0.481 0.485 0.683 0.400 0.457 0.607 0.432 0.442 0.618
0.4 0.413 0.449 0.610 0.470 0.496 0.683 0.393 0.463 0.607 0.423 0.435 0.607
0.45 0.419 0.423 0.595 0.464 0.495 0.679 0.389 0.447 0.593 0.419 0.418 0.592
0.5 0.409 0.415 0.583 0.457 0.473 0.658 0.383 0.435 0.579 0.420 0.432 0.603
0.6 0.395 0.419 0.576 0.450 0.450 0.636 0.385 0.434 0.580 0.403 0.442 0.598
0.7 0.395 0.429 0.583 0.442 0.454 0.634 0.385 0.448 0.591 0.412 0.454 0.613
0.8 0.392 0.427 0.580 0.444 0.448 0.631 0.393 0.440 0.590 0.417 0.468 0.627
0.9 0.395 0.433 0.586 0.434 0.446 0.622 0.392 0.445 0.593 0.407 0.485 0.633
1 0.402 0.453 0.606 0.428 0.445 0.617 0.394 0.422 0.578 0.406 0.498 0.642
1.25 0.408 0.457 0.613 0.425 0.457 0.624 0.402 0.428 0.587 0.419 0.507 0.657
1.5 0.409 0.464 0.619 0.441 0.448 0.628 0.421 0.434 0.604 0.415 0.505 0.653
2 0.415 0.464 0.623 0.437 0.458 0.633 0.412 0.470 0.625 0.409 0.524 0.665
2.5 0.406 0.441 0.599 0.438 0.443 0.623 0.419 0.446 0.612 0.421 0.529 0.676
3 0.397 0.412 0.572 0.440 0.441 0.623 0.399 0.400 0.565 0.414 0.524 0.668
3.5 0.388 0.394 0.553 0.435 0.424 0.608 0.403 0.353 0.536 0.407 0.535 0.672
4 0.390 0.376 0.541 0.442 0.409 0.603 0.398 0.365 0.540 0.396 0.542 0.671
4.5 0.384 0.363 0.528 0.441 0.393 0.591 0.384 0.349 0.519 0.380 0.530 0.652
5 0.394 0.346 0.525 0.461 0.408 0.615 0.375 0.293 0.476 0.371 0.514 0.633
1528 J. X. Zhao, et al.

Figure 6. Variation of within-event, between-event, and total


standard deviations with spectral period. The horizontal lines are
the smoothed standard deviations from the Abrahamson et al.
(2016) study. The color version of this figure is available only in
the electronic edition.

Figure 8. Comparison of (a) between-site and (b) within-site


standard deviations estimated from within-event residuals for four
site classes. The color version of this figure is available only in the
electronic edition.

Figure 9. Comparison of total site standard deviations, esti-


mated from within-event residuals, with the within-event standard
deviations. The color version of this figure is available only in the
electronic edition.

tude over 7.0, and for within-event residuals, each of the last
Figure 7. Comparison of (a) between-event and total standard two magnitude bins has only one event. Figure 12 does not
deviations and (b) within-event standard deviations from the present support an assumption that standard deviations decrease with
study and the Abrahamson et al. (2016) study. The color version of increasing magnitude, and we assume magnitude-independent
this figure is available only in the electronic edition. standard deviations in this study. Also the Abrahamson et al.
(2016) model for global subduction events adopted magni-
deviations do not vary with magnitude in a systematic manner. tude-independent standard deviations.
Figure 12 shows the standard deviations of the residuals in
each magnitude bin of 0.5 magnitude units. For both types of
Predicted Response Spectra
residuals, the standard deviations tend to increase with increas-
ing magnitude up to 7, and then decrease. For between-event Though each model parameter has been smoothed with
residuals, the last magnitude bin has all events with a magni- respect to the logarithm of spectral periods, the model does
GMPEs for Subduction Interface Earthquakes in Japan 1529

Figure 10. Residual factors, the exponential of the negative


between-event residual, for the two largest subduction interface
events, together with the factors for mean  τ. When the residual
factor is larger than 1.0, the spectrum is overestimated, and when
the residual factor is less than 1.0, the spectrum is underestimated,
on average. The color version of this figure is available only in the
electronic edition.

Figure 12. Variation of standard deviation for residuals in each


magnitude bin for (a) between-event residuals and (b) within-event
residuals. The color version of this figure is available only in the
electronic edition.

Figure 13. Predicted rock-site spectra for shallow interface


events with Mw 5–9 and a depth of 15.0 km at a source distance
of 20.0 km. The color version of this figure is available only in
Figure 11. Distribution of peak ground acceleration (PGA) re- the electronic edition.
siduals with respect to source distance (a) with a trend line having a
second-order polynomial of source distance and (b) with a trend line
having a third-order polynomial. The color version of this figure is Figure 13 shows the rock-site spectra from events with a
available only in the electronic edition. magnitude of 5, 6, 7, 8, and 9 at a source distance of 20 km
and a fault-top depth of 15 km. The predicted rock-site PGA
not provide a smoothed spectrum at all magnitude and dis- is 0:053g from Mw 5, 0:16g from M w 6, 0:344g from M w 7,
tance ranges. The adjustment factors for the nonlinear soil 0:59g from M w 8, and 1:03g from Mw 9 events. Among the
site spectra lead to a broadly smoothed nonlinear spectrum. periods that have been modeled, the peak rock-site spectrum
Next, we present the predicted spectra for various magnitude, is at a spectral period of 0.08 s. The differences between the
fault depth, and source distance ranges. spectra from an M w 8 event and the spectra from an M w 7
1530 J. X. Zhao, et al.

Figure 14. Predicted spectra for shallow interface events with Figure 15. Predicted spectra for shallow interface events with
Mw 5–9 and a depth of 15.0 km at a source distance of 20.0 km for Mw 5–9 and a depth of 15.0 km at a source distance of 20.0 km for
(a) an SC I site and (b) an SC II site. The color version of this figure (a) an SC III site and (b) an SC IV site. The color version of this
is available only in the electronic edition. figure is available only in the electronic edition.

event, or between the spectra from an M w 9 event and M w 8 source distances in a 49–83 km range at SC I and SC II sites,
event, are much smaller than those between the spectra from with the largest value being 1:88g. Judging from the residual
Mw 6 and Mw 5 events, because of the marked reduction in distribution in Figures 10 and 11a, the PGAs from the M w 9.0
magnitude scaling for events with an M w ≥ 7:1. event, at a source distance of 20 km that is the closest dis-
Figure 14a shows the predicted spectra for SC I sites, tance for any subduction zone, are not overestimated by the
and the PGA varies from 0:072g for an M w 5.0 event to 1:38g model in the present study.
for an M w 9.0 event, with a fault-top depth of 15 km at a Figure 16 compares the spectra from an Mw 7.0 event,
source distance of 20 km. The PGA is 0:22g for an Mw 6.0 with a fault depth of 15 km and at a source distance of
event, 0:466g for an M w 7.0 event, and 0:8g for an M w 8.0 20.0 km for rock sites and four soil site classes. The PGA
event. The peak SC I spectrum is at about 0.12 s. Figure 14b is 0:34g for rock site, 0:47g for SC I sites, 0:61g for SC
shows that, for SC II sites, the PGA is 0:099g for an M w 5.0 II sites, 0:44g for SC III sites, and 0:52g for SC IV sites.
event, 0:294g for an M w 6.0 event, 0:608g for an M w 7.0 The short-period spectra for SC III and SC IV sites are gen-
event, 0:965g for an M w 8.0 event, and 1:41g for an M w 9.0 erally smaller than the SC II spectra due to nonlinear site
event. For SC III sites, the PGA is 0:072g for an Mw 5.0 event terms. The SC III and SC IV spectra are less than the rock
and increases to 0:996g for an M w 9.0 event, as shown in spectrum in the 0.04–0.09 s period range, a result of nonlin-
Figure 15a. The PGA is 0:214g for an M w 6.0 event, ear site terms. At short periods up to 0.3 s, SC II sites have
0:443g for an Mw 7.0 event, and 0:7g for an M w 8.0 event. the largest spectrum.
For SC IV sites, an M w 5.0 event has a PGA of 0:090g and an Figure 17a shows the spectra from deep interface events
Mw 9.0 event has a PGA of 0:859g, as shown in Figure 15b. with a depth of 26 km and at a distance of 30 km, from an SC
Our model appears to predict a significantly larger spectrum IV site for four magnitude units. The PGA is 0:052g for an
for an M w 9.0 event compared with those by the Abrahamson Mw 5.0 event, 0:13g for an M w 6.0 event, 0:25g for an
et al. (2016) GMPE (their fig. 10). Among the records from Mw 7.0 event, and 0:41g for an M w 8.0 event. Figure 17b
the M w 9.0 event, five records have a PGA over 1:0g, at shows the spectra for rock and SC I spectra, the SC IV
GMPEs for Subduction Interface Earthquakes in Japan 1531

Figure 16. Predicted spectra for an interface event with Mw 7.0


and a depth of 15.0 km at a source distance of 20.0 km for all site
classes. The color version of this figure is available only in the elec-
tronic edition.

nonlinear spectrum, and the SC IV elastic spectrum from an


Mw 9.0 event with a fault depth of 26 km at a source distance
of 30 km. The PGA is 0:451g for a rock site, 0:611g for an
SC I site, 0:77g for elastic soil site spectrum, and 0:63g for
nonlinear SC IV site spectrum. The nonlinear SC IV site
spectrum is markedly smaller than the elastic spectrum for
spectral periods up to about 0.6 s. The spectra for rock site
and SC I sites have a sag centered at about 0.5 s for events
with an M w ≥ 7:0. We attempted eliminating the sag, but the Figure 17. Predicted SC IV spectra for (a) a deep interface
event with a depth of 26.0 km and a source distance of 30.0 km
distribution of residuals is severely biased, and we do not
for four magnitudes, and (b) the predicted spectra for an M w 9 event
have any plausible physical explanation for this problem, on rock, SC I and SC IV sites. Note that significant soil nonlinear
apart from that we do not have any record from an M w 9 response leads to moderate reduction of the elastic soil site spectrum
event in the dataset for deep interface events. in the nonlinear spectrum. The color version of this figure is avail-
Figure 18 shows the attenuation of PGA and spectrum at able only in the electronic edition.
0.5 s with a magnitude of 5, 6, 7, 8, and 9, within a 25–
320 km source distance range at a fault depth of 20 km Japan. These models used simple geometric attenuation func-
for SC IV sites. The fault depth is 14 km for the M w 9 event, tions, and site classes based on site periods were used as
similar to the depth of the 2011 Tohoku earthquake. These the site terms to model both elastic and nonlinear site effects.
figures clearly show the reduced magnitude scaling for The nonlinear soil site terms developed by Zhao, Zhou, et al.
events with M w > 7:0. The reduced attenuation with distance (2015) and Zhao, Zhou, et al. (2016) were used, and a set of
up to 40 km for the M w 9 event is caused mainly by the non- modification factors were adopted so that the predicted non-
linear site term for PGA as shown in Figure 18a and the spec- linear spectrum is broadly smoothed with respect to spectral
trum at 0.5 s as shown in Figure 18b. Figure 19 shows the period.
attenuation of spectra at 1.0 and 3.0 s spectral periods, and Our models adopted a bilinear magnitude-scaling func-
again the magnitude-scaling rates for large events are signifi- tion hinged at Mw 7.1, the same as for Zhao, Zhou, et al.
cantly less than those for events with an Mw up to 7.1. (2016) for shallow crustal and upper-mantle events and for
Figure 20 shows the effect of volcanic path on the pre- Zhao, Jiang, et al. (2016) for subduction slab events. The
dicted response spectra for a shallow interface event with an coefficients for the magnitude term for events with an
Mw 7.0 at a source distance of 60 km for an SC I site. The Mw > 7:1 are much smaller than those for events with
PGA is reduced from 0:13g to 0:11g at a volcanic distance of Mw < 7:1. The magnitude-scaling rate for the large events
20 km, 0:085g at 40 km, and 0:068g at 60 km. was derived from a dataset including all strong-motion re-
cords since 1968.
Discussion and Conclusions For the records before 1996, almost all site classes were
inferred from the response spectral ratio between the H/V
We presented a set of GMPEs based on the strong- components by Zhao, Irikura, et al. (2006) or from geological/
motion records from subduction interface earthquakes in geotechnical descriptions of the surface materials. We tested
1532 J. X. Zhao, et al.

Figure 18. Attenuation for predicted response spectra for five Figure 19. Attenuation for predicted response spectra for five
magnitude units, a fault-top depth of 20 km, and at SC IV sites for magnitudes, a fault-top depth of 20 km, and SC IV sites for periods
(a) PGA and (b) 0.5 s spectral acceleration. The fault-top depth is of (a) 1.0 s and (b) 3.0 s. The color version of this figure is available
14 km for the Mw 9 event. The color version of this figure is avail- only in the electronic edition.
able only in the electronic edition.

the effect of site information quality. Similar to the GMPEs


developed for shallow crustal and upper-mantle events by
Zhao, Zhou, et al. (2016) and for subduction slab events
by Zhao, Jiang, et al. (2016), we found that the goodness-
of-fit parameter of the GMPEs for subduction interface (re-
ferred to as interface) events improved significantly for all
spectral periods after excluding the records from sites with
an inferred site class. The model coefficients presented in this
article were derived based on records from stations with a
measured shear-wave velocity profile down to engineering
bedrock.
The effect of volcanoes on the attenuation of seismic
waves was modeled by applying an anelastic attenuation rate
to the horizontal portion of a straight-line distance (the clos-
est distance) between the fault plane and the recording sta-
tion that passes through the assumed low-Q zones around the Figure 20. Predicted SC I response spectra from an M w 7.0
active volcanoes. Our approach to the modeling of volcanic shallow interface event with a depth of 20 km, and at a source dis-
zones can approximately model the wave propagation path tance of 60 km for four travel distances of 0, 20, 40, and 60 km
within volcanic zones. The color version of this figure is available
along the volcanic front but within the volcanic zone in the only in the electronic edition.
western side of the volcanic front, whereas the binary terms
in the Abrahamson et al. (2016) and Boore et al. (2009)
model the effect of volcanic path across the volcanic front. geometric attenuation rate was used for the deep interface
We found that anelastic attenuation rate is not necessary events, the anelastic attenuation rate for these events is either
for deep interface events with a fault-top depth over 25 km, very small or positive. A plausible explanation is that the
similar to the findings by Zhao (2010). Even if a separate materials in the mantle wedge have a relatively high Q value
GMPEs for Subduction Interface Earthquakes in Japan 1533

compared with the materials in the crustal region close to the supplying the fault model parameters for a number of earthquakes, and
subduction trench. Changjiang Wu from Japan for assistance with the Japan Meteorological
Agency (JMA) catalog. Finally, we would like to thank Eric Thompson
We separated the within-event residuals into between- and two anonymous reviewers for their constructive review comments.
site and within-site residuals using an approximate random
effects model. We found that the between-site standard de-
viations varied significantly with spectral period and among References
the site classes, whereas the within-site standard deviations
are more or less constant. Between-site standard deviation Abrahamson, N. A., and W. J. Silva (2008). Summary of the Abrahamson &
Silva NGA Ground-Motion Relations, Earthq. Spectra 24, no. 1, 67–97.
tends to be large at the average site period of each site class. Abrahamson, N. A., and R. R. Youngs (1992). A stable algorithm for
Similar to the Zhao, Zhou, et al. (2016) model for shal- regression analysis using the random effect model, Bull. Seismol.
low crustal and upper-mantle earthquakes, the discontinuity Soc. Am. 82, no. 1, 505–510.
at the boundaries between shallow and deep interface Abrahamson, N. A., N. Gregor, and K. Addo (2016). BC hydro ground
motion prediction equations for subduction earthquakes, Earthq.
events leads to different predicted spectra for events at the
Spectra 32, no. 1, 23–44.
depth boundary. For example, the predicted spectrum from Allen, T. I., and D. J. Wald (2009). Evaluation of ground-motion modeling
an interface earthquake with a depth slightly smaller than techniques for use in global shakemap—A critique of instrumental
25 km (shallow interface events) would differ from that cal- ground-motion prediction equations, peak ground motion to macro-
culated for a deep interface earthquake with a depth slightly seismic intensity conversions, and macroseismic intensity predictions
larger than 25 km at the same source distance for the same in different tectonic settings, U.S. Geol. Surv. Open-File Rept. 2009–
1047.
site class. We attempted to use continuous variables to Atkinson, G. M., and D. M. Boore (2003). Empirical ground-motion relations
model all the observed differences, but it is almost impos- for subduction-zone earthquakes and their application to Cascadia and
sible to derive a smoothed spectrum with respect to spectral other regions, Bull. Seismol. Soc. Am. 93, no. 4, 1703–1729.
periods for most magnitude and source distance ranges. Boore, D. M., and G. M. Atkinson (2008). Ground-motion prediction
Even if it can be done, it is arguable how accurately the equations for the average horizontal component of PGA, PGV, and
5%-damped PSA at spectral periods between 0.01 s and 10.0 s, Earthq.
boundaries, such as the interface between the mantle wedge Spectra 24, no. 1, 99–138.
and the subducting plate, can be known. Another problem is Boore, D. M., A. A. Skarlatoudis, B. N. Margaris, C. B. Papazachos, and
the uncertainty associated with earthquake classification, C. Ventouzi (2009). Along-arc and back-arc attenuation, site
especially the assumption of a depth tolerance of 5 km response, and source spectrum for the intermediate-depth 8 January
2006 M 6.7 Kythera, Greece, earthquake, Bull. Seismol. Soc. Am. 99,
above or below the subduction interface for defining sub-
no. 4, 2410–2434.
duction interface events (Zhao, Zhou, et al., 2015). We rec- Building Seismic Safety Council (BSSC) (2000). The 2000 NEHRP
ommend that, for an event located at a boundary, the user Recommended Provisions for New Buildings and Other Structures,
takes the averaged ground motion of the models on either Part I (Provisions) and Part II (Commentary), FEMA 368/369.
side of the boundary. This is thought to be a reasonable ap- Washington, D.C.
proach, considering the depth error range associated with Campbell, K. W., and Y. Bozorgnia (2008). NGA ground motion model for
the geometric mean horizontal component of PGA, PGV, PGD and 5%
the event classification. damped linear elastic response spectra for periods ranging from 0.01 to
10 s, Earthq. Spectra 24, no. 1, 139–171.
Data and Resources Chiou, B. S.-J., and R. R. Youngs (2008). An NGA model for the average of
horizontal component of peak ground motion and response spectra,
The strong-motion records are from K-NET and KiK- Earthq. Spectra 24, no. 1, 173–216.
net, administered by the National Research Institute for Earth Ghofrani, H., and G. M. Atkinson (2014). Ground-motion prediction
equations for interface earthquakes of M7 to M9 based on empirical
Science and Disaster Prevention of Japan (http://www. data from Japan, Bull. Earthq. Eng. 12, no. 2, 549–571.
kyoshin.bosai.go.jp/, last accessed July 2015). A small num- Hayes, G. P., D. J. Wald, and R. L. Johnson (2012). Slab1.0: A three-
ber of records are from the Port and Airport (Port and Harbour) dimensional model of global subduction zone geometries, J. Geophys.
Research Institute. A number of rock-site strong-motion re- Res. 117, no. B01302.
cords are from the Pacific Earthquake Engineering Research Kanno, T., A. Narita, N. Morikawa, H. Fujiwara, and Y. Fukushima (2006).
A new attenuation relation for strong ground motion in Japan based on
Center (PEER) strong-motion database (http://peer.berkeley. recorded data, Bull. Seismol. Soc. Am. 96, no. 3, 879–897.
edu/smcat/, last accessed July 2015). McVerry, G. H., J. X. Zhao, N. A. Abrahamson, and P. G. Somerville (2006).
Crustal and subduction zone attenuation relations for New Zealand
Acknowledgments earthquakes, Bull. New Zeal. Soc. Earthq. Eng. 39, no. 1, 1–58.
Zhao, J. X. (2010). Geometric spreading functions and modelling of
The work reported here is partially supported by research grants from volcanic zones for strong-motion attenuation models derived from
the National Science Foundation of China (51278432) and the Southwest records in Japan, Bull. Seismol. Soc. Am. 100, no. 2, 712–732.
Jiaotong University (SWJTU12ZT04), and by a New Zealand Earthquake Zhao, J. X. (2014). Magnitude scaling rates for large earthquakes in ground
Commission 2010 Biennial Research Grant. At an early stage (2011), sup- motion prediction equations for Japan, 2014 Annual Meeting of
port was received from the New Zealand Foundation for Research Science Seismological Society of America, Anchorage, Alaska, 30 April–2
and Technology, New Zealand Hazards Platform Contract C05X0907. The May 2014.
authors would like to thank Jim Cousins and Chris Van Houtte of GNS Zhao, J. X., and D. A. Rhoades (2014). Ground-motion prediction equations
Science for their review of this article. We would like to thank Kimiyuki for subduction zones based on strong-motion records from Japan, GNS
Asano from Disaster Prevention Institute (DPRI), Kyoto University for Science Consultancy Report. 2014/236.
1534 J. X. Zhao, et al.

Zhao, J. X., and H. Xu (2012). Magnitude-scaling rate in ground-motion pre- Institute of Crustal Dynamics
diction equations for response spectra from large subduction interface China Earthquake Administration
earthquakes in Japan, Bull. Seismol. Soc. Am. 102, no. 1, 222–235. 1 Anningzhuang Road, Haidian District
Zhao, J. X., J. S. Hu, F. Jiang, J. Zhou, and D. A. Rhoades (2015). Nonlinear Beijing 100085, China
(X.L.)
site models derived from 1-D analyses for ground-motion prediction
equations using site class as the site parameter, Bull. Seismol. Soc. Am.
105, no. 4, 2010–2022.
Zhao, J. X., K. Irikura, J. Zhang, Y. Fukushima, P. G. Somerville, A. Asano, Y. GNS Sciences
Ohno, T. Oouchi, T. Takahashi, and H. Ogawa (2006). An empirical 1 Fairway Drive, Avalon
site-classification method for strong-motion stations in Japan using Lower Hutt 5010, New Zealand
H/V response spectral ratio, Bull. Seismol. Soc. Am. 96, no. 3, 914–925. (D.A.R.)
Zhao, J. X., F. Jiang, P. Shi, H. Xing, Y. Zhang, P. C. Yu, M. Lu, and D. A.
Rhoades (2016). Ground-motion prediction equations for subduction
slab earthquakes in Japan using site class and simple geometric at- Aichi Institute of Technology
tenuation functions, Bull. Seismol. Soc. Am. 106, no. 4, doi: 10.1785/ Aichi Prefecture
0120150056. Toyota 470-0392, Japan
Zhao, J. X., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi, H. Ogawa, (K.I.)
K. Irikura, H. K. Thio, P. G. Somerville, et al. (2006). Attenuation re-
lations of strong ground motion in Japan using site classification based
on predominant period, Bull. Seismol. Soc. Am. 96, no. 3, 898–913. International Seismic Safety Centre
Zhao, J. X., S. L. Zhou, P. J. Gao, T. Long, Y. B. Zhang, H. K. Thio, M. Lu, Division of Nuclear Installation Safety
and D. A. Rhoades (2015). An earthquake classification scheme for Department of Nuclear Safety and Security
ground-motion prediction equations in Japan, Bull. Seismol. Soc. International Atomic Energy Agency
Am. 105, no. 5, 2750–2763, doi: 10.1785/0120150013. Vienna International Centre
Zhao, J. X., S. L. Zhou, P. J. Gao, Y. B. Zhang, J. Zhou, M. Lu, and D. A. PO Box 100
Rhoades (2016). Ground-motion prediction equations for shallow 1400 Vienna, Austria
(Y.F.)
crustal and upper mantle earthquakes in Japan using site class and
simple geometric attenuation functions, Bull. Seismol. Soc. Am.
106, no. 4, doi: 10.1785/0120150063.
AECOM
915 Wilshire Boulevard, 7th Floor
School of Civil Engineering Los Angeles, California 90017
Southwest Jiaotong University (P.G.S.)
111 1st Northern Section of Erhuan Road
Chengdu 610031
Sichuan, China Manuscript received 30 January 2015;
(J.X.Z., X.L., F.J., H.X., M.Z., R.H., Y.Z.) Published Online 12 July 2016

You might also like