Professional Documents
Culture Documents
Ground Motion Prediction Equation For Taiwan Subduction Zone Earthquakes
Ground Motion Prediction Equation For Taiwan Subduction Zone Earthquakes
Ground Motion Prediction Equation For Taiwan Subduction Zone Earthquakes
Earthquake Spectra
Ground motion prediction 2020, Vol. 36(3) 1331–1358
Ó The Author(s) 2020
equation for Taiwan Article reuse guidelines:
sagepub.com/journals-permissions
Abstract
A ground motion prediction equation (GMPE) is presented for computing the median
and standard deviation of peak ground acceleration (PGA) and 5% damped pseudo-
spectral acceleration (PSA) for periods between 0.01 s and 5.0 s for probabilistic
seismic hazard analysis (PSHA) and engineering applications in Taiwan. An integrated
strong motion dataset consisting of two subduction earthquake regions was selected
from 3314 recordings from Taiwan with M4.5 to M7.1 and 3376 recordings from
Japan with M6.5 to M9.1. This dataset was then used to validate, and refit where nec-
essary, the function form provided by Abrahamson et al. for application to Taiwan
subduction earthquakes. The proposed model accounts for the extrapolation beha-
viors associated with the large-magnitude scaling and the near-source scaling terms,
both of which were developed empirically by using the combined Taiwan–Japan data-
set. The distance attenuation and site term were developed specifically for the
Taiwan region. The site term is based on two parameters; the time-averaged shear
wave velocity of the top 30 m depth (VS30) and the depth-to-the-shear wave velocity
horizon of 1.0 km/s (Z1.0).
Keywords
Ground motion prediction, subduction zone earthquake, Taiwan earthquake, regres-
sion analysis, residual analysis
Date received: 27 March 2019; accepted: 26 December 2019
1
Department of Civil Engineering, National Taiwan University, Taipei
2
Institute of Earth Sciences, Academia Sinica, Taipei
3
National Center for Research on Earthquake Engineering, Taipei
4
University of California—Berkeley, Berkeley, CA, USA
Corresponding author:
Chin Hsiung Loh, Department of Civil Engineering, National Taiwan University, No. 1, Section 4, Roosevelt Road, Taipei
10617.
Email: lohc0220@ccms.ntu.edu.tw
Figure 1. Tectonic structural profiles of the subduction zones sources: (a) tectonic diagram of Taiwan
(Angelier et al., 1986), (b) tectonic profile of seismic sources in the Ryukyu subduction zone, and (c)
tectonic profile of seismic profile of seismic sources in the Manila subduction zone (SSC TI Team, 2018c).
Introduction
Taiwan is located at the collisional boundary between the Philippine Sea Plate and the
Eurasian plate (Lin and Lee, 2008; Shyu et al., 2016). Figure 1a shows a schematic diagram
showing the major geologic setting in the region. In the southern part of Taiwan, the
Eurasian plate actively sub-ducts eastward under the Luzon volcanic arc on the Philippine
Sea Plate, forming the Manila subduction zone. In the North, the Philippine Sea plate sub-
ducts into the Eurasian plate, forming the Ryukyu subduction zone trench. Taiwan is at the
junction of two subduction zones; as a result, the contact zone of the plates and the subduc-
tion zone intraslab are characterized by frequent earthquakes. Figure 1b and c shows the dif-
ferences in position, geometry, and depth in locations with collision and subduction.
The subduction zone sources are divided into two types: interface and intraslab earth-
quakes. Subduction zone interface earthquakes are low-angle thrust events that occur at
the interface between the subducting and the overriding plate. Subduction zone intraslab
earthquakes occur within the subducting oceanic plate and are high-angle normal faulting
events responding to down-dip tension in the subducting plate (Lin and Lee, 2008;
Youngs et al., 1997), and intraslab earthquakes can be attributed to tension but also to
dehydration of the subducting slab (Preston et al., 2003). A large and destructive interface
earthquake occurred off the coast of northeastern Taiwan on March 31, 2002, M7.1
(Hualien M7.1). In addition, two instraslab known as the Pingtung dual earthquake
occurred on December 26, 2006 (Pingtung M6.9 and Pingtung M7.0) in southwestern
Taiwan. These three earthquakes, identified as subduction earthquakes, had a strong
impact on civil infrastructure and caused severe loss of life and property. In addition to
the Hualien M7.1 and the Pingtung dual earthquakes, many other large earthquakes
occurred in the past century but were not included in the database due to lack of available
recordings. Examples are the 1909 Taipei earthquake M7.3 (Lin and Lee, 2008) and the
1920 surface wave magnitude M8.1 (Hsiao et al., 2014). According to seismic source char-
acterization (SSC) models (SSC TI Team, 2018b), large earthquakes with M . 8.0 may
occur in the future. This is one of the main motivations for the development of ground
motion prediction equations (GMPEs).
A GMPE is typically used in probabilistic seismic hazard analysis (PSHA) to assess the
expected ground motion for a given return period, and this model expresses the predicted
ground motion as a function of magnitude (M), rupture distance (Rrup), the time-averaged
shear wave velocity in the top 30 m depth (VS30), the depth to top of rupture (Ztor), and
the depth-to-the-shear wave velocity horizon of 1.0 km/s (Z1.0). The most recent GMPE
for Taiwan subduction earthquakes (Lin and Lee, 2008) was developed using only data
from northeast Taiwan. Their model is conditioned on hypocentral distance and the com-
mon site terms for rock and soil, which is not recommended to be used as a modern
GMPEs which use the rupture distance (Rrup) and VS30 as the primary inputs. In addition,
Lin and Lee (2008) use a common linear magnitude scaling term that could potentially
lead to an over prediction of ground motion for large magnitude events (e.g. above M8.0),
and thus causing either over-predicting the hazard or large events to dominate the hazard
deaggregation. In fact, this equation should not be used in PSHA studies, according to the
exclusion criteria proposed by Bommer et al. (2010).
A major issue with PSHA that has been raised under the framework of the NCREE project
(GMC TI Team, 2018b) is how to correctly quantify epistemic uncertainty associated with
GMPEs. In other words, GMPEs have to capture the center body and range of the expected
future ground motion in the area of interest. A single model is not suggested, especially for the
case that lacks an indigenous GMPE because different GMPEs are commonly combined
within a logic tree to capture the whole range of ground motion uncertainty (Bommer, 2012;
GeoPentech 2015). The seismic hazard community therefore relies on using a set of GMPEs
developed from the same database and used for the logic tree framework, as described in
Bommer et al. (2005), where each model occupies a different branch of the logic tree and where
the branch assigned with the weight express a degree of uncertainty.
Motivated by the development of a set of GMPEs employed in the logic tree, a number
of robust GMPEs were reviewed. They include the development of GMPEs for subduc-
tion zone based on worldwide data (e.g. Abrahamson et al., 2016; Atkinson and Boore,
2003; Youngs et al., 1997) and local data (Zhao et al. 2016a, 2016b, for Japan; Idini et al.,
2017, and Montalva et al., 2017, for Chile). The review aims to find key modeling para-
meters that can be modified or calibrated to match the empirical data in the region of
interest. Each model functional form has its own feature and is able to provide a good fit
to the empirical data. However, the function form by Abrahamson et al. (2016) (AGA16)
developed for the BC Hydro project is thought to be the most physically realistic because
it contains a bilinear magnitude scaling term with a break (Mref); Mref is a flexible compo-
nent in the model to predict different source scaling for large magnitudes. This allows for
physically appropriate extrapolation of the ground motion to both large and small magni-
tudes. Lack of strong motion data from large events makes the development of a robust
GMPE for the Taiwan subduction zone difficult, which may also lead to the over
prediction of ground motion for assessing the seismic hazard of large earthquakes. A solu-
tion for dealing with this issue has been addressed through the development of models that
combine data from multiple regions, while also accounting for specific regional effects
(Abrahamson et al., 2016; Atkinson and Boore, 2003; Youngs et al., 1997). Therefore, it is
our decision to include data from the Japan region, which provides large magnitude
events, and to adopt the functional form of AGA16 for the development of the model,
which helps to constrain the empirical relations using a physical basic. A difficulty in
adopting the AGA16 model is the nonlinear site response behavior, which requires suffi-
cient ground motion data collected from large-magnitude earthquakes and close recorded
sites with low VS30; however, such recordings in Taiwan are sparse and were mostly
recorded at distances larger than 20 km from a few large-magnitude events. Therefore, the
solution for this is to consider the nonlinear site response developed outside Taiwan; that
is the nonlinearity of the site amplification used in the AGA16.
The objective of this study is to develop a set of GMPEs for Taiwan subduction zones.
The linear site amplification is developed using only local data with the range of VS30
greater than 330 m/s, while the non-linearity site amplification is constrained. The reason
for using VS30 = 330 m/s is that these sites are thought to have very weak nonlinearity.
The development of the proposed model is divided into seven sections. Taiwan ground
motion database describes data selection criteria, followed by Functional form of subduc-
tion GMPE, which describes the modifications of the AGA16 functional form. Then in
Analysis of distance scaling of AGA16 GMPE with respect to Taiwan and Japan data the
functional form is validated against with the empirical data. Model development describes
the regression procedure for model development. Evaluation of model presents the evalua-
tion of the proposed model and compares the model with previous models. Finally,
Summary and conclusion summarizes the main points addressing the applications and the
limitations of the proposed GMPE.
Figure 2. Epicenter distribution of subduction zone earthquake occurred in Taiwan from 1983 to 2016.
and computing the distance for each site for each simulation, and then taking the median
of these values for use in completing the metadata in the Taiwan ground motion database
(SSC TI Team, 2018a). Earthquakes are classified into shallow crustal, deep crustal, sub-
duction interface, and subduction intraslab. Two types of subduction earthquakes can be
recognized by average depth range, and most interface earthquake ruptures are shallower
than those of intraslab earthquakes due to the geometry of Taiwan subduction system.
The database contains high-quality instrumental recordings of strong ground motion
from earthquakes. These recordings were mainly collected from the TSMIP, and partially
collected from the Broadband Array in Taiwan for Seismology (BATs). The TSMIP pro-
gram was initiated in 1991 and deployed by the Seismology Center of the Central Weather
Bureau, Taiwan (CWB) (Liu et al., 1999). The TSMIP stations were in densely populated
regions except for mountainous areas. They are positioned approximately 5 km from each
other on average (3 km in urban areas). The BATs stations system was established by the
Institute of Earth Sciences (IES), Academia Sinica, in 1992, and incorporated with the
CWB in 1999 to increase the total number of BATs stations to 64, covering both mountai-
nous and plain areas of the islands. Since 1993, the TSMIP program has operated more
than 700 stations, which are used together with 64 BATs stations. Currently, there are
more than 764 stations, providing orientation-independent measures of ground motion
(RotI50) as defined in Boore (2010). Figure 2 shows epicenter locations of subduction
earthquakes recorded in Taiwan. This corresponds to 111 events with M ø 4:5, resulting
in 14726 strong motion records, equivalent to 36% of the total number of records in the
database. Also, Figure 2 shows all stations locations grouped by VS30 according to the bin
interval of 120–180 m/s, 180–360 m/s, 360–760 m/s, and 760–1500 m/s.
The database provides the maximum usable distance (Rmax), also known as the cutoff
distance. The concept of Rmax was introduced by Chiou and Youngs (2014) and was
assessed event by event by considering the ground motion level (i.e. peak ground accelera-
tion [PGA]), triggered by earthquakes, set to the instrumentations. For example, the 4 gal
PGA was used for the Taiwan Strong Motion Instrumentation Program (TSMIP) system.
The use of Rmax can help to filter out data that are significantly biased at long distances.
Only data in the distance range 0 to Rmax was used in the development of the GMPE.
Table 1 summarizes the maximum usable distance Rmax of Taiwan subduction earth-
quakes used in this study. Table 1 also shows the number of records per event (Nrec), the
number of usable recordings (Nrec_Rmax), and the number of recordings from BATs sys-
tem. There are 36 BATs recordings in the Taiwan data subset, which is a minor contribu-
tion to the total regression dataset.
As there is a lack of ground motion data recorded from large-magnitude subduction
earthquakes in Taiwan, Japan subduction ground motion data for M . 6.5 were extracted
from the NGA-sub ground motion database (NGAsub_Mega_Faltfile_RotD50_R170930,
from Brian Chiou personal communication, 2017). These data were used together with
Taiwan subduction data. The cutoff magnitude-dependent distance (Rmax) truncation for
Japan was guided by Zhao et al. (2016a). The Rmax was initially set to 200 km (i.e. consis-
tent with Rmax developed for a M6.5 earthquake from Taiwan) for an M6.5 earthquake;
this distance increases linearly with magnitude to 300 km for an M9.0 earthquake. The
Rmax for each individual event can be shortened or extended depending on the data quality
visually inspected so as to select the unbiased data with distance.
The final dataset for regression was developed on the basis of the following criteria: (1)
Select the data in the distance interval of 0–Rmax; (2) select data with VS30 greater than
330 m/s; (3) select the event with a focal depth less than 100 km; (4) select the events with
at least five remaining records after all other criteria are applied; and (5) select only the
response spectral values for spectral frequencies greater than 1.25 times the high-pass cor-
ner frequency used in the record processing. Under the consideration of these criteria, a
total number of 3314 recordings obtained from 51 Taiwan subduction earthquakes were
selected; 21 of these were interface events and 30 were intraslab events. The number of
Japanese data was 3746, selected from 15 subduction interface and 10 subduction intraslab
earthquakes. The magnitude–distance–region distribution and magnitude–depth–region
distribution are shown in Figure 3. The magnitudes of the Taiwan interface were M4.8–
M7.1 with distances ranging 1–180 km. The magnitudes of Taiwan intraslab events were
M4.6–M 7.0 with distances ranging 19–280 km. The magnitudes of Japan intraslab earth-
quakes were M6.7–M7.5 with rupture distances ranging 43–345 km. The magnitudes of
Japan interface earthquakes were M6.5–M9.1 with distances ranging 26–330 km. The
depths to the top of rupture (Ztor) of Taiwan earthquakes are displayed in the range 0–
35 km for interface, while intraslab earthquakes mostly occur at Ztor to a depth greater
than 35 km. The depths to Ztor of Japan earthquakes are displayed in the range 0–60 km
for both interface and intraslab events.
The distribution by VS30 is given in Figure 4. It shows selected records for Japan and
Taiwan. There are more soft rock sites (VS30 < 760 m/s) than rock sites (VS30 . 760 m/
s); the number of soft rock site and rock site records are 3354 and 392 for Japan and 2943
and 284 for Taiwan, respectively.
(a) (b)
Figure 3. (a) Magnitude–distance–region distribution and (b) magnitude–Ztor–region distribution of final
regression dataset.
(a) (b)
Figure 4. Bar chart showing VS30 bin for records in the selected data: (a) Japan subduction dataset and
(b) Taiwan subduction dataset.
scaling, which can be used to capture the magnitude-scaling rate for large subduction
earthquakes. There are also some modifications compared with the model of Abrahamson
et al. (2016). Instead of using hypocentral depth and hypocentral distance, the closest dis-
tance to the rupture plane (Rrup ) and the depth to the top of rupture (Ztor ) are used for both
types of earthquakes because finite-fault models are determined for the events in the data-
base. Formulations for Ztor were modified to reflect trends in data. The fore-arc to back-
arc model is dropped because there is no metadata specifying whether a record available in
the database was in a fore-/back-arc region. The function form of the median model is
described as follows:
where PSA is the pseudo-spectral acceleration. The predictor variables are M = moment
magnitude; Rrup = closest distance (km) to the rupture plane; Ztor = depth (km) to the top
of the rupture; Feve is a dummy variable (0 for interface and 1 for intraslab). VS30 = travel
time average shear wave velocity (m/s) of the top 30 m of soil. The random errors h and e
denote the between-event residual and within-event residual, respectively.
The magnitude FMAG ðM, uÞ has the following functional form:
(
a4 M Mref + a13 ð10 M Þ2 for M<Mref
FMAG ðM, uÞ = , ð2Þ
a5 M Mref + a13 ð10 M Þ2 for M.Mref
where Mref is 7:8 + DC1 (DC1 is used to capture the epistemic uncertainty) and represents a
break in magnitude scaling in AGA16.
The depth scaling FDEP ðZtor , uÞ functional form is given by
a10 ðminðZtor , 40Þ 20Þ for interface
FDEP ðZtor , uÞ = ð3Þ
a11 ðminðZtor , 80Þ 40Þ for intraslab:
The caps for depth scaling were chosen because there is not significant increase in
between-event residual as Ztor increases beyond 40 and 80 km.
The distance scaling FDIST ðM, R, uÞ has the following function form:
FDIST ðM, R, uÞ = ða2 + a14 Feve + a3 ðM 7:8ÞÞln Rrup + C4 expða9 ðM 6ÞÞ + a6 Rrup : ð4Þ
where PGA1000 = median PGA value at VS30 = 1000 m/s, and VS = minðVS30 , 1000Þ. In addi-
tion, the basin effect term was also considered in the model because it can help to capture
additional site amplification that is not captured by using VS30 alone. This leads to
reduction of the sigma value. The site FBASIN ðZ1:0 , uÞ functional form is chosen based on
Boore et al. (2014):
Z1:0
FBasin ðZ1:0 , uÞ = a8 min ln ,1 : ð6Þ
Z1:0ref
Here, Z1:0ref is reference sediment depth which can be inferred from the VS30–Z1.0 relation-
ship when only VS30 is available. We validate the relationship developed by Kuo et al.
(2017) using 478 profiles, mainly used in this study. By examination the Kuo et al. (2017)
model against the current use dataset, we found that there is some degree of misfit at
VS30 \ 500 m/s; as a result, the Kuo et al. (2017) model was modified, as described in
Equation 7, in order to re-center it relative to the sediment depth.
2
3:96 VS30 + 352:72
ln Zref = ln : ð7Þ
2 17502 + 352:72
Here, the functional form FSite ðVS30 , uÞ is used to represent VS30 scaling, and u represents a
coefficient or a set of coefficients needed to determine the site amplification components.
Abrahamson et al. (2016) adopted the results of finite-fault simulation conducted by
Gregor et al. (2002) to constrain the magnitude-dependent fictitious depth related to coef-
ficient C4 and a9 . Abrahamson et al. (2016) also indicated that the geometrical spreading
terms are highly correlated with the saturation term and cannot all be determined from
the data. For this reason and for our limitation, we decided to fix the terms controlling
the near-source distance scaling; thus, the coefficients a3 , C4 , and a9 were fixed to 0.1, 10,
and 0.25, respectively, in our regression model. The distance scaling Equation 8 has two
coefficients that need to be regressed to match the empirical data. Coefficient a2 represents
magnitude-independent geometric spreading and is used to fit the ground motion data
with a distance range of 0–100 km, while coefficient a6 measures the curvature of attenua-
tion at large distances.
Because of a high degree of correlation in the estimation of geometric and anelastic
attenuation, the slope and curvature of distance decay are not easy to determine from
regression, especially for data with significant scatter. This behavior can be analyzed by
(a) (b)
Figure 5. Fits of distance scaling function form Equation 8 to the data for the 2011 M9.1 Tohoku
interface earthquake from Japan (a) and for the 2006 M6.9 Pingtung intraslab earthquake from Taiwan
(b). Solid and dashed curves indicate the fitted model plotted for average VS30.
carrying out the fit to each individual event for two cases: the model is analyzed using
Equation 8 with coefficient a6 and the other model is analyzed using Equation 8 without
coefficient a6 (i.e. constrain a6 = 0). Figure 5 illustrates the fits to the PSA at T = 0.01 s
for two well-recorded earthquakes: The 2006 Taiwan Pingtung intraslab earthquake with
M6.9 and the Japan 2011 Tohoku Interface earthquake with M9.1. The median curve in
red is the result of the fitting for the average value of VS30 of the data. As shown in the fig-
ure, both models (with and without a6 ) can be used to satisfactorily capture the data
trends. The fit to the Japanese earthquake shows a substantial curvature, while the fit to
the Taiwanese data shows a little curvature. By constraining a6 = 0, the model fitting to
the Tohoku earthquake exhibits an unexpectedly large prediction at a distance of 10 km;
however, this situation is not significant for the fit to the data from the Pingtung earth-
quake. Therefore, there is a need for the anelastic attenuation term for the model fitting to
the Tohoku earthquake but not the model fitting to the Pingtung earthquake.
By using the same approach, the analyses were performed for 30 events with their data
recorded at wider distance ranges, which is equivalent to 90% of the total number of
Taiwan data used in the regression. All data from the Japan events were selected for the
analysis as they are good representations of recorded motions that attenuates with dis-
tance. The results are shown in Figure 6, which illustrates the geometrical spreading coeffi-
cient a2 , obtained to study regional effects for periods 0.01 and 2.0 s. Based on Figure 6,
the first observation is that there is almost no change in coefficient a2 between readings
with and without coefficient a6 in the Taiwan data. On the contrary, results from the
Japan data show that the model without a6 results in more negative values of a2 compared
with the model with a6 , and this is more obvious for T = 0.01 s than for T = 2 s.
(a) (b)
Figure 6. Comparisons of geometrical spreading, coefficient a2 plotted against magnitude, obtained
from fits to individual earthquakes to study regional effects: the cases were compared without and with
anelastic attenuation a6 .
not supported by the Japan data (Figure 6). This is identical to that of AGA16 for which
the distance scaling of interface is different from that of the intraslab case by using differ-
ent magnitude-independent geometrical spreading coefficients.
Model development
The development of the GMPE is performed using the mixed-effect regression method
proposed by Abrahamson and Youngs (1992) with three stages. Each stage is performed
with a number of steps until each regressed coefficient converges. First, for stage 1, we
temporarily undertake the magnitude scaling term of the AGA16 and perform an iterative
regression process to constrain the path term. Then, stage 2 is performed to determine the
source term while the path term is fixed at the values obtained from stage 1. After con-
straining the Ztor term from the primary model, stages 1 and 2 are repeated to refine the
magnitude scaling with the inclusion of the Ztor effect. In stage 3, the site model is devel-
oped by using region-specific data to constrain the amplification term. In this stage, the
basin depth term can be determined from the analysis of within-event residuals with
respect to the lnðDZ1:0 Þ parameter, which is defined as ln ZZ1:0ref
1:0
. During the regression, the
coefficients a3 , C4 , and a9 are held fixed and used for all periods. The coefficient a9 is fixed
at 0.25 but not 0.4 as it is used in AGA16. This makes our model behaving no magnitude
over saturated even though some studies allow magnitude over saturated of the predicted
motion (e.g. Abrahamson et al., 2016; Boore et al., 2014).
Figure 7. An example of distance scaling for T = 0.01 s, in which different values of coefficient a2 are
compared to gain a sense of the convergent selection of geometrical spreading. Coefficient a2 was
selected after four iterations, each of which is shown by black dashed-dot lines. The dashed lines result
from regression against data grouped into magnitude bins with 0.5 width. This example helps to illustrate
the validity of convergent criteria leading to stable estimation of regressed coefficients.
Figure 8. PSA at T = 0.01 s magnitude scaling with and without the consideration effect of Ztor. The
average event term is evaluated within 0.5 magnitude bins. The slab ground motions were corrected by a
constant in agreement with the interface ground motions.
Figure 9. Variation of coefficient a12 with periods, representing various degrees of site amplification
for each of regional data.
(a) (b)
Figure 10. (a) Plots showing within-event residuals against basin differential parameter DZ1:0 along
with adjusted model (red plus) for 2s. (b) Plot showing the variation of coefficient a8 with period.
residual (dW0 ). Therefore, the total standard deviation of the model is the square root of
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the sum of their squares, that is, sT = t2 + f2S2S + f2SS in which fS2S is standard deviation
of dS2Ss quantifying the site-to-site variability and fSS is the single station standard devia-
tion. The computation was done considering the criteria for stations with at least 10
recordings with regard to a period of 0.01 s, and used the same stations in determining
fSS and fS2S .
series steps described above. Parameters such as a1 , a4 , and a5 were adjusted manually to
achieve a visually smoothed response spectra.
Tables 2 to 4 list the final coefficients of our model development, and the variation of
the relevant coefficients with period are shown in Figure 11. Note that the difference in
geometric coefficients for the interface is slightly different from that for intraslab, which
were shown by the variation of coefficient a2 . The differences in anelastic attenuation and
site effects between Taiwan and Japan are represented by coefficient a6 (Figure 11) and
coefficient a12 (Figure 9), respectively. The values of coefficients a6 and a12 obtained from
combined Taiwan–Japan regression dataset were also plotted for comparison, and these
coefficients were used in the development of the source term.
Evaluation of model
Examination of residuals
In this section, residual analysis was carried out to check for any uncaptured dependence
on the explanatory variables. The residuals are shown for spectral periods of 0.01, 0.2, and
2 s. Figure 12 shows between-event residuals plotted with respect to source parameters M
and Ztor , respectively. There exists a cluster of interface events with a few high values of
between-event residuals appearing at depths of 20 to 30 km; however, these high between-
event residuals were not shown significantly with respect to magnitude. Therefore, we
accepted our model with this local bias against Ztor rather than M. Figure 13 shows within-
event residuals plotted against Rrup and VS30 along with medians in bin. No significant
trend of the within-event residuals is visible with respect to Rrup and VS30 ; this confirms the
validity of the derived model capturing data behaviors in terms of distance attenuation and
Figure 11. Variation of regression coefficients listed in Table 1, Table 2, Table 3, and Supplemental Table
A1 and model variability (t, fS2S , and fSS ) with period.
(a) (b)
Figure 12. Between-event residuals plotted against: (a) magnitude (M) and (b) depth to the top of the
rupture plane (Ztor) spectral period of 0.01, 0.2, and 2 s. Number of event (Ne) and standard deviation
of between-event residuals (t) are shown at the top of each panel.
developed using Rrup and H. The Rhypo and H were used in the development of the LL08.
Even though GMPEs are guided with different uses in distance and depth metrics, the
comparisons are not distinguishable between distance and depth.
Figures 14 and 15 show the comparison of distance attenuations for T = 0.01 s and
2.0 s. The comparisons are shown on each panel by the data grouped into three bins (i.e.
M = 4.5–5.5, M = 5.5–6.5 and M = 6.5–7.5) in which the GMPEs were computed at
three different average M, average Ztor–depth, and average VS30 over each data bin. As
can be seen in these figures, all GMPEs generally provide a good fit to the observation
data; however, our model shows a visually better fit to the observation data in the range
for M = 4.5–5.5 and M = 5.5–6.5 except for M = 6.5–7.5 at which there is some degree
of misfit due to intraslab motion for T = 2.0 s; the misfit is probably due to the regres-
sion dataset with the present of Japanese data, which hindered the fit to M = 6.5–7.5 data
(a) (b)
Figure 13. Within-event residuals plotted against: (a) the rupture distance (Rrup) and (b) the time-
averaged shear wave velocity in the top 30 m depth (Vs30) for spectral period of 0.01, 0.2, and 2 s. Number
of records (Nrec) and standard deviation of within-event residuals (f) are shown at the top of each panel.
bin. For this data bin, LL08 shows a better fit to the data, while other models show over-
prediction. As developed from the Taiwan data, our model and LL08 show the predictions
close to each other for most of the data range.
For the model extrapolation features, Figure 16 shows the prediction for an interface
earthquake with M8.2 at 8 km depth and VS30 of 401 m/s shown for T = 0.01 s and 2 s.
Note that the ground motion data is also plotted along with GMPEs’ prediction, and
GMPEs were computed using a scenario (top panel), which is the average values over the
relevant dataset. The dataset corresponds to a set of events including one event M8 from
Central America & Mexico region, three events M8.2, 8.3 and 8.3 from Japan region, and
three events M8.4, 8.3 and 8.2 from South America. In addition, Figure 17 shows the pre-
diction for an intraslab earthquake with M7.8 at 290 km depth and the site condition VS30
of 435 m/s shown for T = 0.01 s and 2 s. The data bin corresponds to a set of intraslab
events deeper 100 km including one event M8.0 from Alaska region, three events M7.6,
Figure 14. Distance attenuation showing for comparison of the proposed model with previously
published GMPEs (Zh16, LL08, and AGA16) and for T = 0.01 s. The shaded area indicates the prediction
range within one sigma around the median.
7.9, and 8.4 from Japan region, and three events M7.5, 7.7, and 7.8 from South America.
As can be seen in the figures, the set of GMPEs show a wide range of prediction with vari-
ous level of distance attenuation in which the proposed model exhibits moderate distance
Figure 15. Distance attenuation showing the comparison of the proposed model with previously
published GMPEs (Zh16, LL08, and AGA16) and for T = 2.0s. The shaded area indicates the prediction
range within one sigma around the median.
Figure 16. Distance attenuation showing the comparison of models’ extrapolation with large
magnitude M8.2 of interface event for T = 0.01 s and 2.0 s.
attenuation and close to the prediction of the LL08, but notably different from the
AGA16 and the Zh16, especially for T = 2 s. The AGA16 and the Zh16 perform compa-
rable distance attenuation, but their predictions are largely different. The prediction range
of intraslab models have shown to be much broader than those of interface models. This
is because of the LL08 and the Zh16 not considering a cap on the depth scaling term. Our
intraslab model exhibits low predictions for distance less than about 300 km, and it shows
to be close the AGA16 prediction. In terms of the models comparing with the data, the
proposed model shows a generally good fit to the data over wide range of distance,
Figure 17. Distance attenuation showing the comparison of models’ extrapolation with large
magnitude M7.8 of intraslab event deeper than 100 km for T = 0.01 s and 2.0 s.
(a) (b)
Figure 18. Comparison on response spectra for earthquakes M6 (a) and M8 (b) at distance of 50 km
with given VS30 = 760 m/s.
suggesting the applicability of our model to large magnitude and deep intraslab earth-
quakes for PSHA in Taiwan.
Figure 18 shows a comparison of the response spectra for sites with a VS30 = 760 m/s.
The response spectra were computed for two magnitude earthquakes at a rupture distance
of 50 km: M = 6.0 earthquake using Ztor = 20 km and M = 8.0 earthquake using
Ztor = 5 km. Observations from our model reveals that the peak of the response spectra
of M6 is about T = 0.2 s for interface earthquake, while it appears at T = 0.15 s for
intraslab earthquake. The peak of response spectra of M8 is shifted to T = 0.5 s and
T = 0.2 s for interface earthquakes and slab earthquakes, respectively. This is not seen
from global GMPE–AGA16 and Zh16 model in which the peaks are observed to be
around 0.1 to 0.2 s. The LL08 model shows its peak to be near 0.15 s for M = 6.0, while
the peak is seen to be at 0.4 s for M = 8.0. It is understood that the dataset for the devel-
opment of AGA16 was a major part of the Zh16 dataset, while the regional difference
between Taiwan and Japan was already explored in the previous section. This might cause
significant misfit among peaks. However, we need to investigate the case further and need
more data from larger events to prove this point.
We note that one of the limitations in this study is that our GMPE does not consider
the characterization of the attenuation difference between force-arc and back-arc regions
(as indicated in Figure 1), which can be observable in Taiwan especially from the deep sub-
duction earthquakes or at large distances (.100 km). Because of the strong lateral hetero-
geneity, the stronger attenuation in the back-arc was identified to be significant in several
regions (e.g. Japan, Romania, New Zealand) (Beauval et al., 2017; Vacareanu et al., 2015).
We consider use of an average attenuation model acceptable in the current situation since
there is no metadata indicating fore-/back-arc locations in the database we employed. As
the metadata become available, we need to redevelop our GMPE considering distinct
attenuation in the force-arc and back-arc as it is modeled in the development of the global
Abrahamson et al. (2016) GMPE.
Acknowledgments
This research is part of the research program in National Center for Research on Earthquake
Engineering (NCREE). We thank Dr. Brian Chiou, Prof. Kuo-Liang Wen, and Prof. Yin-Nan
Huang for their meaningful discussions. We thank the reviewers for their thoughtful comments,
which greatly helped to improve the manuscript. We also thank Uni-edit (www.uni-edit.net) for edit-
ing and proofreading this manuscript.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this
article.
Supplemental material
Supplemental material for this article is available online.
References
Abrahamson NA and Youngs RR (1992) A stable algorithm for regression analyses using the
random effects model. Bulletin of the Seismological Society of America 82: 505–510.
Abrahamson NA, Gregor N and Addo K (2016) BC hydro ground motion prediction equations for
subduction earthquakes. Earthquake Spectra 32: 23–44.
Angelier J, Barrier E and Chu HT (1986) Plate collision and paleostress trajectories in a fold-thrust
belt: The foothills of Taiwan. Tectonophysics 125: 161–178.
Atkinson GM (1989) Attenuation of the LG phase and site response for the Eastern Canada
telemetered network. Seismological Research Letters 60: 59–69.
Lin P-S and Lee C-T (2008) Ground-motion attenuation relationships for subduction-zone
earthquakes in Northeastern Taiwan. Bulletin of the Seismological Society of America 98:
220–240.
Liu KS, Shin TC and Tsai YB (1999) A free field strong motion network in Taiwan: TSMIP.
Terrestrial Atmospheric and Oceanic Sciences 10: 377–396.
Molas G and Yamazaki F (1995). Attenuation of earthquake ground motion in Japan including deep
focus events. Bulletin of the Seismological Society of America 85: 1343–1358.
Montalva GA, Bastı́as N and Rodriguez-Marek A (2017) Ground-motion prediction equation for
the Chilean subduction zone. Bulletin of the Seismological Society of America 107: 901–911.
Mori J and Helmberger D (1996) Large-amplitude Moho reflections (SmS) from Landers
aftershocks, Southern California. Bulletin of the Seismological Society of America 86: 8.
Preston LA, Creager KC, Crosson RS, et al. (2003) Intraslab earthquakes: Dehydration of the
Cascadia slab. Science 302(5648): 1197–1200.
Rodriguez-Marek A, Cotton FA, Abrahamson N, et al. (2013) A model for single-station standard
deviation using data from various tectonic regions. Bulletin of the Seismological Society of
America 103: 1242–1258.
Rodriguez-Marek A, Montalva GA, Cotton F, et al. (2011) Analysis of single-station standard
deviation using the KiK-net data. Bulletin of the Seismological Society of America 101: 1242–1258.
Shyu JBH, Chen C-F and Wu Y-M (2016) Seismotectonic characteristics of the northernmost
longitudinal valley, eastern Taiwan: Structural development of a vanishing suture. Tectonophysics
692: 295–308.
SSC TI Team (2018a) Fault parameters, Taiwan ground motion characterization. NCREE technical
report, Appendix A, pp. 42–43. Development of the Hazard Input Document for Taiwan Using
SSHAC Level 3 Methodology, prepared for Taiwan Power Company, 4024 pp. Available at:
http://sshac.ncree.org.tw
SSC TI Team (2018b) Fault slip—Rate model, Taiwan ground motion characterization. NCREE
technical report, chapter 6, pp. 5–21. Development of the Hazard Input Document for Taiwan
Using SSHAC Level 3 Methodology, prepared for Taiwan Power Company, 4024 pp. Available
at: http://sshac.ncree.org.tw
SSC TI Team (2018c) Seismic source characterization overview, Taiwan ground motion characterization.
NCREE technical report, chapter 3, pp. 1–46. Development of the Hazard Input Document for
Taiwan Using SSHAC Level 3 Methodology, prepared for Taiwan Power Company, 4024 pp.
Available at: http://sshac.ncree.org.tw
Stewart JP, Douglas J, Javanbarg M, et al. (2015) Selection of ground motion prediction equations
for the global earthquake model. Earthquake Spectra 31: 19–45.
Vacareanu R, Radulian M, Iancovici M, et al. (2015) Fore-arc and back-arc ground motion
prediction model for Vrancea intermediate depth seismic source. Journal of Earthquake
Engineering 19: 535–562.
Wu Y-M, Cheng S-N and Liu TY (2016) A foundation and basis analysis of earthquake catalog and
historical earthquakes database. Taiwan ground motion NCREE technical report, Appendix B.
SSC TI Team. Development of the Hazard Input Document for Taiwan Using SSHAC Level 3
Methodology, prepared for Taiwan Power Company, 4024 pp. Avaiable at: http://sshac.
ncree.org.tw
Youngs R, Chiou BJ, Silva W, et al. (1997) Strong ground motion attenuation relationships for
subduction zone earthquakes. Seismological Research Letters 68: 58–73.
Zhao JX, Jiang F, Shi P, et al. (2016a) Ground-motion prediction equations for subduction slab
earthquakes in Japan using site class and simple geometric attenuation functions. Bulletin of the
Seismological Society of America 106: 1535–1551.
Zhao JX, Liang X, Jiang F, et al. (2016b) Ground-motion prediction equations for subduction
interface earthquakes in Japan using site class and simple geometric attenuation functions.
Bulletin of the Seismological Society of America 106: 1518–1534.
Zhao JX, Zhang J, Asano A, et al. (2006) Attenuation relations of strong ground motion in Japan
using site classification based on predominant period. Bulletin of the Seismological Society of
America 96: 898–913.