Machine-Learning Methods for Earthquake Ground Motion

Analysis and Simulation

Arzhang Alimoradi, M.ASCE1; and James L. Beck, M.ASCE2

Abstract: This paper presents a novel method of data-based probabilistic seismic hazard analysis (PSHA) and ground motion simulation, ver-
ified using previously recorded strong-motion data and machine-learning techniques. The procedure consists of three parts: (1) selection of an
orthonormal set of basis vectors called eigenquakes to represent characteristic earthquake records; (2) estimation of response spectra for the
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

anticipated level of shaking for a scenario earthquake at a site using Gaussian process regression; and (3) optimal combination of the eigen-
quakes to generate time series of ground acceleration consistent with the response spectral ordinates obtained in the second part. The paper
discusses the benefits of applying such machine-learning methods to strong-motion databases for PSHA and ground motion simulation, par-
ticularly in large urban areas where dense instrumentation is available or expected. The effectiveness of the proposed methodology is exhibited
using four scenario examples for downtown Los Angeles. Advantages, disadvantages, and future research needs for this machine-learning ap-
proach to PSHA are discussed. DOI: 10.1061/(ASCE)EM.1943-7889.0000869. © 2014 American Society of Civil Engineers.
Author keywords: Ground motion; Earthquake engineering; Simulation models; Optimization; Machine learning; Principal component
analysis; Gaussian process regression; Genetic algorithms; Intensity measures; Probabilistic seismic hazard analysis.

Introduction megathrust earthquake in the Sendai region, Japan, is estimated to

have caused approximately US$20–30 billion in loss. All of these
events occurred at or near the center of urban developments.
Probabilistic Seismic Hazard Assessment Probabilistic methods of seismic hazard analysis have been de-
Probabilistic seismic hazard analysis (PSHA) has been the subject of veloped to provide quantitative measures of the likelihood of the
many studies since the seminal publication of Cornell (1968) occurrence of seismic events (or exceedance of a certain level of
(Kramer 1996; Bazzurro and Cornell 1999; Bozorgnia and Bertero intensity of shaking such as spectral acceleration) at a site over
2004; McGuire 2004, 2008; Delavaud et al. 2009; Musson 2009; a period of time. Generally, the outcome of these studies is expressed
Scherbaum et al. 2009; Wang and Takada 2009). Recently, several in the form of a set of uniform hazard spectra that show the level of
critiques of PSHA have been made (Bommer and Abrahamson intensity at different periods corresponding to different probabilities
2006, 2007a, b; Klugel 2007, 2009; Wang and Zhou 2007; Strasser of exceedance. The PSHA procedures rely on regression models,
and Bommer 2009; Geller 2011). The scarcity of recorded strong- often called ground motion prediction equations (GMPEs) or at-
motion data from large earthquakes, the harsh consequences as- tenuation relationships, that estimate the measure of intensity as a
sociated with rare seismic events, and the uncertainties about the function of magnitude, distance, and other model-specific param-
source, path, and site effects of future events make the problem of eters depending on the model and its intended use.
seismic hazard assessment important both from an academic point The current standard analytical procedures of PSHA can be
of view and from considerations of societal safety. This problem summarized into four steps:
is further exacerbated by the vast urban developments that are 1. Identification and geometrical characterization of all sources
spreading into seismogenic regions and by the aging infrastructure of seismicity affecting the site of interest and probabilistic
(ASCE 2011), the result of which is evident from the global description of source-to-site distances.
earthquake losses in the recent past. The January 12, 2010, M7.0 2. Development of a recurrence relationship for temporal distri-
Haiti earthquake killed more than 220,000 people and injured bution of event recurrence at each source; the Gutenberg-
300,000 more. The February 27, 2010, M8.8 Maule, Chile, earth- Richter relationship is often developed, with consideration of
quake collapsed four buildings and severely damaged 50 other characteristic earthquakes from recorded data and catalogs of
buildings. The September 3, 2010, M7.1 earthquake and its February historical events to provide a probabilistic description of the
22, 2011, M6.3 aftershock in Canterbury, New Zealand, caused rate of occurrence of different magnitudes of events at the site.
approximately US$12 billion in loss. The March 11, 2011, M9.0 3. Selection and implementation of a proper GMPE to estimate
the measure of shaking for pairs of magnitude and distance
from Steps 1 and 2.
1
Graduate Student, Dept. of Mathematics, Southern Methodist Univ., 4. Use of total probability theorem to consider uncertainty in
Dallas, TX 75205 (corresponding author). E-mail: [email protected] earthquake size, location, and ground motion parameter by
.org integrating over all magnitudes and distances to arrive at the
2
George W. Housner Professor of Engineering and Applied Science,
probability of shaking exceeding a predefined value over
California Institute of Technology, Pasadena, CA 91125.
Note. This manuscript was submitted on September 26, 2013; approved a period of time.
on August 18, 2014; published online on September 17, 2014. Discussion With information obtained in Step 4, one can construct a set of
period open until February 17, 2015; separate discussions must be submitted uniform hazard spectra with different probabilities of exceedance
for individual papers. This paper is part of the Journal of Engineering that would serve as a basis for design of a facility (i.e., the min-
Mechanics, © ASCE, ISSN 0733-9399/04014147(13)/$25.00. imum required level of shaking to be considered in the design for

© ASCE 04014147-1 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 events with certain probabilities of occurrence) (Newmark and earliest studies on strong ground motions that a single measure of
Hall 1982). intensity cannot reliably predict the damage potential of such a
Although development of design spectra is currently the ultimate complex process (Housner 1975; Housner and Jennings 1982).
goal of PSHA, these spectra by themselves do not provide enough Building codes do not require use of a specific ground motion
information for modern methods of seismic design and performance selection and modification method, but they do provide brief
assessment. This is because such performance assessment proce- guidelines on the adequacy of such procedures mainly from the
dures require nonlinear dynamic response history analysis of consensus of the earthquake engineering community. For example,
a structural model subject to sets of ground motion time series, Chapter 21 of ASCE 7-05 (ASCE 2006), “Site-Specific Ground
natural or synthetic, that represent levels of shaking intensity pre- Motion Procedures for Seismic Design,” reads as follows: “At least
dicted by the PSHA with considerations of site, source, and path. five recorded or simulated horizontal ground motion acceleration
Therefore, ground motion selection and modification procedures time histories shall be selected from events having magnitudes and
have been developed to select a set of ground motion records and fault distances that are consistent with those that control the
adjust them to match the shaking intensities obtained in PSHA [maximum considered earthquake]. Each selected time history shall
[Shome et al. 1998; Naeim et al. 2004; Baker and Cornell 2006; be scaled so that its response spectrum is, on average, approximately
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

Watson-Lamprey and Abrahamson 2006; Beyer and Bommer 2007; at the level of the [maximum considered earthquake] rock response
Luco and Bazzurro 2007; Kottke and Rathje 2008; Shahbazian and spectrum over the period range of significance to structural re-
Pezeshk 2010; Grigoriu 2011; Pacific Earthquake Engineering sponse.” Consequently, there are different methods of ground mo-
Research Center (PEER) 2011a; Baker 2011]. Usually, this selection tion selection and scaling in use with large discrepancies between
and scaling for seismic design purposes is done by representing the their estimates of structural response quantities resulting from their
strong ground motions and their severity by a few simple parameters. application (PEER 2011a). Furthermore, as has been known for
Historically, peak ground acceleration (PGA) had been used solely a long time, excessive scaling of ground motion records and spectral
for characterization of ground motion with all its shortcomings, but alteration can be problematic (Fig. 1). This issue is even more
more recently, spectral acceleration at the first mode period, Sa ðT1 Þ, relevant nowadays because of the popularity of time-history re-
has been used widely. Vector-based intensity measures (IMs) have sponse analysis procedures. There is more evidence in recent years
also been proposed recently (Baker and Cornell 2005, 2008). The against spectral matching and excessive scaling, and there are
parameters used to characterize ground motion records and their procedures to minimize the adverse effects of altering recorded
severity are commonly referred to as IMs in the performance-based ground motions (Naeim et al. 2004; Grigoriu 2011).
earthquake engineering literature (Krawinkler 2001). However, one An appealing alternative to directly using recorded ground
cannot expect to fully capture the complexity of strong ground motions based on selection and modification is stochastic ground
motions by only a few simple parameters; in fact, it was noted in the motion simulation (Housner and Jennings 1964; Jennings et al.

Fig. 1. Example of ground motion modification to match a design spectrum (note that the near-source pulse is completely removed in the process of
modification); structural response evaluation subject to this record will likely produce questionable results

© ASCE 04014147-2 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 1969; Amin and Ang 1968; Saragoni and Hart 1973; Ahmadi 1979, application is that the posterior probability is not used to express the
1980; Der Kiureghian and Crempien 1989; Conte et al. 1992; confidence in the models, and instead, ad hoc weighting schemes are
Papadimitriou 1990; Papadimitriou and Beck 1992; Atkinson and employed that do not use information available from system data.
Beresnev 2002; Mobarakeh et al. 2002; Gu and Wen 2007; Rezaeian Such schemes may bias the final output of the tree in one way or
and Der Kiureghian 2008; Giaralis and Spanos 2009; Yamamoto and another.
Baker 2011). The guiding philosophy in this case is to construct There are also other challenges in using simplified models of re-
a probabilistic description of the strong ground motion time history ality in engineering analysis. Predictive models do not typically get
that is informed by relevant past strong-motion records and, pref- automatically updated with the arrival of new data. Therefore, it is
erably, some seismological theory. Thus, the uncertainty in future often not clear, when new observations are made, how these new
ground motions at a site is directly modeled probabilistically to observations might affect the prediction of a model and by how much.
generate samples of simulated records rather than using a probabi-
listic description of a few (often just one) ground motion parameters
Motivation for This Study
to select and scale past records.
Stochastic ground motion models are often constructed by As of February 2010, the combined global strong ground motion
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

passing a random process through a set of mathematical filters that databases exceed 105 records (Anderson 2010). With the growing
shape the random process in the time and frequency domains, so that number of high-quality ground motion records available worldwide,
the resulting motion has a form and intensity appropriate for the site. instigated by the availability of low-cost sensors and major in-
Seismological theory can be included, for example, where the filter strumentation programs such as the Advanced National Seismic
is chosen to generate synthetic records with both temporal and System (ANSS) (USGS 2011), application of machine-learning
spectral nonstationarities for a given set of seismic hazard parame- methods for analysis, classification, and simulation of strong mo-
ters (magnitude, distance, etc.) (Ahmadi 1980; Papadimitriou 1990; tion becomes compelling. The vast amount of data generated and
Papadimitriou and Beck 1990, 1992; Conte et al. 1992; Rezaeian and archived around the globe on a daily basis exceeds the individual
Der Kiureghian 2008) and by modifying the source spectrum of an processing capability of human beings but is suitable for the ap-
earthquake over the path of seismic waves to account for the at- plication of machine learning and data mining. Most applications in
tenuation of waves analytically (Boore 1983; Hanks and McGuire seismic hazard analysis and ground motion simulation only make
1981; Atkinson and Silva 1997). Historically, stochastic simulation use of a small number of previously recorded data. There is a crucial
methods have received many applications in the design of critical need for creation of systematic procedures for analysis of large
infrastructure such as nuclear power plants. databases of ground motion. Another motivation is to develop a data-
Both standard PSHA and stochastic simulation methods rely on based method of stochastic ground motion simulation that does not
regression-type models. With available data, predictive models are depend on scaling and modification of recorded ground motions,
easy to construct, understand, and use; but similar to the imple- using information learned from a large number of previously
mentation of any kind of model, care must be practiced when recorded data.
making predictions in the real world. Regression models of a real The authors make no assumptions with regards to the details,
system’s behavior are usually developed by using observations of number, and nature of dynamic processes that work together to
a system’s output to inform the selection of a model structure and its generate the ground motion observations during a seismic event;
parameter values. Their utility is limited to the type of information however, the authors hypothesize that, whatever the nature of these
contained in the regression data and the appropriateness of the processes might be, their effects must be reflected in the waveforms
chosen model class. If future observational data suggest the exis- generated by them (Fig. 2). Therefore, careful study of the generated
tence of other important aspects of the system behavior that were data without imposing any predefined mathematical model may
originally not included in the model or if the model exhibits large reveal information that is not available in common methods of
prediction errors, then the utility of the model would be questionable. ground motion and seismic hazard assessment where the terms in the
In the Bayesian framework of system modeling and analysis, regression equation are usually selected a priori (Alimoradi et al.
one’s knowledge of a system is improved by Bayesian updating 2005). A notable example is the influence of local topography on
based on the data available from the system (Beck and Katafygiotis ground motion observations, a factor that is almost universally
1998; Beck 2010). Any effort to represent an actual system or ignored in major GMPEs (Hough et al. 2010) but that could sig-
process with a mathematical model should be conditioned based on nificantly alter the characteristics of the motion observed at a site.
the information that is available from the system’s data and prior The same is true for other phenomena and processes (basin edge
knowledge about the process. Furthermore, there is usually more effects, trampoline effects, etc.) that are relatively newly discovered
than one possible model class to represent a system. Which model but have important effects on the intensity of motion observed at
class should be used in a given situation? For example, there are a site (Choi et al. 2005; Yamada et al. 2009).
many GMPEs in the literature and in use that have different In this work, the authors use principal component analysis (PCA)
mathematical forms [for a summary, see Abrahamson et al. (2008)]. (Jolliffe 2002) to extract useful information from a large database of
In the Bayesian framework, candidate model classes for a system can ground motion records for these ground motion simulations. The
be assessed by examining their posterior probability based on the authors then use recorded data to develop estimates of ground
system data (Cheung and Beck 2010). This posterior probability motion intensity at a site, using acceleration response spectra over
reflects a trade-off between a data-fit measure for the model class and a wide range of periods, for different probabilities of exceedance by
the information extracted from the data by the model class (Beck employing a Gaussian process (GP) regression formulation (Bishop
2010). Posterior model averaging can then be used where the 2006; Rasmussen and Williams 2006). Examples presented later
probabilistic predictions of each model class are weighted by their show the effectiveness of the procedure.
posterior probability. Another procedure is to use a logic tree where A general concern in data-driven methods is dealing with sit-
the user’s confidence in different models is expressed by a weighting uations where data are scarce, such as large events recorded close to
scheme applied to different branches of the tree to find the most the source. The authors anticipate that this will not be a problem in
critical case or to arrive at an average of different model outputs. The the future as the archives of ground motion records continue to grow.
problem with the usual implementation of logic trees for this Many urban areas in the world where earthquakes cause significant

© ASCE 04014147-3 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 2. Description of the complex physical processes that produce strong ground motion at a site, which are difficult to model accurately

losses have already been heavily instrumented or are expected to be represent the seismic hazard at a site with a specified IM and
instrumented. As a result, the data-driven methods presented here a probability of exceedance. A standard procedure for disaggrega-
are particularly suitable for seismic hazard assessment and ground tion of the seismic hazard can be used to choose appropriate pairs of
motion simulation in large urban areas. magnitude and distance, given the probability of exceedance of the
IM at the site. Data associated with the location, as well as accel-
eration response spectra of all magnitudes and distances previously
Methodology recorded, can be used next in a GP regression to generate estimates
of acceleration spectra of the motion anticipated at the site. Given an
The authors wish to exploit more fully current earthquake ground orthogonal basis of characteristic earthquake records from PCA that
motion databases for PSHA and ground motion simulation by using the authors refer to as eigenquakes, the anticipated waveforms can
machine-learning techniques to analyze a large amount of data. The be synthesized by finding an optimal set of basis coefficients such
objective is not to explicitly model the underlying physical pro- that the linear combination of the eigenquakes would create a signal
cesses that contribute to ground shaking and seismic hazard (Song that has spectral acceleration ordinates as close as possible to those
and Somerville 2010) nor it is to rely on the usual regression-type obtained from the GP regression. The motivation for the term
GMPEs (Abrahamson and Silva 2008; Boore and Atkinson 2008; eigenquakes is that the prefix eigen- means own in German, as used
Campbell and Bozorgnia 2008; Chiou and Youngs 2008; Idriss in eigenvectors and eigenvalues of matrices and eigenfaces in pattern
2008). Physical modeling, although valuable in developing an recognition. The ground motion time histories that are generated
understanding of earthquake processes, requires detailed information from the eigenquakes for a given event scenario should be verified
about the seismogenic, tectonic, and geotechnical settings at a site with appropriate recorded data whenever possible.
that is not, in general, available for most engineering projects.
Regression-type GMPEs, on the other hand, only make probabilistic
predictions of one or two parameters describing shaking intensity at Eigenquakes from PCA of a Ground Motion Database
a site, thereby giving an incomplete probabilistic prediction of future
ground motions at the site. The authors define eigenquakes as the PCA time histories that
The proposed methodology is purely data driven (Fig. 3). From characterize earthquake records in a strong-motion database. Be-
recorded ground motion data, the authors extract a set of ortho- cause they are principal components, they form an orthonormal
normal basis vectors using PCA that represent the predominant basis in the space of ground motion records with the same sam-
variations in the time series with temporal and spectral nonstationarities pling interval and duration. Therefore, any such discrete-time strong-
automatically included. From the same data, the authors develop motion record aðtÞ can be expressed as a linear combination of
estimates of the shaking intensity at the site for a given scenario event eigenquakes
using GP regression. A scenario event has a certain moment mag-
P
nEigq
nitude and source-to-site distance along with a particular shear-wave aðtÞ ¼ ai ui ðtÞ (1)
velocity for the top sedimentary layers at the site. The authors then i51
utilize the results of these two procedures to generate ground motion
time histories for a given event scenario. where aðiÞ and uðiÞ 5 basis coefficient and the discrete time history
As an example, the proposed procedure for seismic hazard of the ith eigenquake, i 5 1, . . . , nEigq. Technically, the eigen-
analysis and ground motion simulation would proceed as follows. quakes are eigenvectors of a data matrix, and they are ranked by the
Suppose that ground motion acceleration time series are needed to size of their corresponding eigenvalues.

© ASCE 04014147-4 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 3. Schematic description of the proposed procedure

Fig. 4. Top 10 extracted eigenquakes from the PEER ground motion database

A sample of the eigenquakes that are used in the simulations is Note that the eigenquakes resemble the temporal and spectral non-
shown in Figs. 4 (time domain) and 5 (time-frequency domain). They stationarities of actual earthquake records. This property makes them
are the top 10 principal components based on 530 records from the particularly suitable as a set of basis vectors for ground motion
PEER next generation attenuation (NGA) database (Alimoradi 2011; simulation. They also provide the utility of dimensionality reduction in
PEER 2011b). The power spectral density spectrograms obtained a large database by extracting only the useful information from
from short-time Fourier transforms are calculated for the first 40 s of previously recorded data and by discarding redundancies and noise
the eigenquakes with a Kaiser (also known as the Kaiser-Bessel) (Jolliffe 2002). Therefore, the number nEigq of eigenquakes needed in
window of length 256 data points, making approximately eight Eq. (1) is usually much smaller than the total number of records in the
moving windows along the length of the signals (MATLAB 2009). The strong-motion database.
Kaiser window is a function of the zeroth-order modified Bessel
function of the first kind, and it has two parameters that adjust the
GPs for Response Spectra Regression
width of the side lobes and main lobe. A side-lobe attenuation pa-
rameter of 0.5 is used with 78% overlap and fast Fourier transform Empirical (i.e., purely data-based) model development is an integral
(FFT) length of 512 data points. The sampling frequency is 50 Hz. part of all branches of applied sciences where mathematical models

© ASCE 04014147-5 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 5. Power spectral density spectrograms for the top 10 eigenquakes calculated for the first 40 s with a Kaiser window of length 28 5 256 data points
(making approximately eight moving windows along the length of the signals); the sampling frequency is 50 Hz

  
are developed to describe a complex system’s response given some p½ f ðxÞ ∼ GP mðxÞ, k x, x9 (2)
observations of the system output. Classical parametric regression
analysis has been used widely in the past; however, with a large amount
which are functions of the input variable x. Note that Eq. (2) defines
of data, which is usually the case in modern applications, the task of
a probability distribution over functions of x. There are different
model development by the application of classical regression methods
choices available for the covariance function of a GP. Depending on
becomes challenging. The following issues require careful attention:
the application, some of these choices are (Rasmussen and Williams
1. A set of observed data can be described with different re-
2006)
gression models. Which model should be chosen as most • Squared exponential (SE) covariance
appropriate?
2. It is possible to better fit the data by the mean prediction of "  2 #
  1 x 2 x9
a regression model by increasing its order (number of re- k x, x9 ¼ sf exp 2
2
(3)
gression terms) but with the risk of overfitting the data so that 2 l
it does not do a good job when predicting new data. How can
the problem of overfitting be avoided? • Matérn’s class
3. Nonlinear regression models, such as artificial neural net- pffiffiffiffiffi n pffiffiffiffiffi 
works, have very flexible mathematical forms, although a good   212n 2n x 2 x9 2n x 2 x9
choice of the network architecture is challenging. k x, x9 ¼ Kn (4)
GðnÞ l l
4. Regression models make generalizations based on some train-
ing data from the system, and their predictions are therefore • Brownian motion (Wiener process)
dependent on these data being comprehensive; extrapolation
outside the domain of the data space is risky.    
In summary, regression methods are subject to the challenging k x, x9 ¼ min x, x9 (5)
issues of model selection, overfitting, and extrapolation risk.
• Linear covariance
Regression based on GPs provides a rational framework for
dealing with these problems. Their history of development goes  
back to Kriging methods in geostatistics (Bishop 2006). Rasmussen k x, x9 ¼ s2f þ xx9 (6)
defines GPs as “a generalization of multivariate Gaussian dis-
tributions to infinitely many variables” and “a distribution over where GðnÞ 5 gamma function; KðnÞ 5 modified Bessel function of
functions [with] inference taking place directly in the space of the second kind; and l and n 5 parameters of the covariance. The SE
functions” (Rasmussen 2006; Rasmussen and Williams 2006). covariance function in Eq. (3) is smooth and stationary (invariant to
Similar to a Gaussian distribution, a GP is defined by mean and translation) and will be used in this study, where it is denoted by
covariance functions, mðxÞ and kðx, x9Þ, respectively kðx, x9Þ. The authors take the mean function mðxÞ as zero.

© ASCE 04014147-6 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 The authors use GPs for regression in a supervised learning the GP uses to estimate the parameters of its covariance function
scheme based on training data that consists of a set of n pairs of consists of a set of spectral acceleration values for strong-motion
system input and output, ðX, YÞ 5 fðxi , yi Þ, i 5 1, . . . , ng. The GP records corresponding to different magnitudes, distances, and site
gives a Gaussian probability density for predicting scalar output yp shear-wave velocity values. Examples of posterior predictions are
corresponding to a given vector input xp (Rasmussen 2006) presented in Fig. 6. These figures exemplify the capabilities of
! carefully designed GPs to estimate the spectrum of earthquake
p p KðX, x p ÞT ½KðX, XÞ21 Y ground motion at a site for a scenario based on magnitude, distance,
pðy jx , X, YÞ ¼ N and shear-wave velocity as the input parameters. The quality of
kðxp , xp Þ 2 KðX, x p ÞT ½KðX, XÞ21 KðX, xp Þ
estimates is good even at the lower tail of the distribution of spectral
(7) acceleration, as is shown for a small-magnitude event at a far dis-
tance in Fig. 6.
where KðX, XÞ 5 n 3 n matrix with (i, j) element kðxi , xj Þ; and
KðX, xp Þ 5 n 3 1 vector consisting of the covariance between each Simulating Site-Specific Ground Motions
of the training inputs and the input xp . Using Eigenquakes
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

To estimate the ground-motion spectral acceleration at period T

The number of eigenquakes required to capture the important
and 5% damping ratio as a function of a scenario event that has
characteristics of the strong-motion records in a large database is
a moment magnitude M recorded at a hypocentral distance from the
expected to be sizable. Consequently, estimation of the basis coef-
causative fault R, on a site with a shear-wave velocity of its top 30-m
ficients for the eigenquakes in Eq. (1) to simulate ground motions
layers V30 , GP regression is applied to the logarithm with base 10 of
consistent with a target response spectrum Spa ðTÞ requires large-scale
the spectral acceleration at the period of interest, so y 5 log½Sa ðTÞ
robust global optimization over a high-dimensional space. The
in this theory. The input vector is x 5 ðx1 , x2 , x3 Þ 5 ðM, R, V30 Þ, and
authors use genetic algorithms (GAs) as a robust stochastic opti-
the authors take a SE kernel with a noise term for the covariance
mization method to solve for the optimal coefficients
function of the GP
& (" # ) !2
 2 ’ P
np P
nEigq  
  3 P3 2 xi 2 xi9 3   ^
a ¼ arg mina2RnEigq Sa ai × ui ðtÞ ,Tj 2 Sap Tj
k x, x9ju ¼ ∏ sf ,i exp
2
2
þ s2noise ∏ d xi , xi9
i51 i51 2li i51 j51 i51

(8) (12)

where dðx, x9Þ 5 1 if x 5 x9 and 0 otherwise; and li and sf ,i where the mean or any probability percentile such as 95% of
5 parameters corresponding to correlation length and the prediction- the GP target spectrum is matched in a mean square error sense
error variance. The parameter vector over a discrete set of np periods, Tj , j 5 1, . . . , np , in ½0, Tmax . Here,
a 5 ½a1 , . . . , anEigq T and Sa ðx, TÞ denote the basis coefficients and


u ¼ snoise , sf ,i , li , i ¼ 1, 2, 3 (9) the acceleration response spectrum at period T for time series x.

can be identified by taking the maximum likelihood estimate of u^ Optimization Using GAs and Parallel
from maximization of the likelihood function (Rasmussen 2006) Computing Schemes

1 1 Over the past 25 years, GAs have been used extensively in many
log½ pðYjX, uÞ ¼ 2 YT KðX, XÞ21 Y 2 log½detðKÞ (10) structural mechanics and earthquake engineering applications (Chan
2 2
1997; Raich and Ghaboussi 2000; Chou and Ghaboussi 2001; Kim
where K depends on u and a constant term in Eq. (10) has been and Ghaboussi 2001; Alimoradi et al. 2006, 2007; Alimoradi and
dropped. Naeim 2006; Foley et al. 2007). This experience with GAs shows
The training data input X is that long computational times may be needed for convergence to
globally optimal solutions. The GA theory and operation can be


X ¼ Mi , Ri , V30,i , i ¼ 1, . . . , n (11) found elsewhere (Goldberg 1989) but a brief description is given
here for comprehensiveness.
corresponding to each of the n records from a strong-motion da- The GAs belong to a group of numerical optimization methods
tabase, whereas the training data output Y has components yi collectively called bioinspired (or nature-inspired) algorithms,
5 log½Sa ðTÞi , i 5 1, . . . , n, calculated for each record. The GP re- which also includes methods such as artificial immune systems, ant
gression then uses Eq. (7) to give the posterior probability density colony optimization, harmony search, big bang–big crunch opti-
function pflog½Spa ðTÞjxp , X, Y, ug ^ for predicting yp 5 log½Sp ðTÞ mization, and swarm optimization. These algorithms use adaptation
a
for a specified scenario x 5 ðM p , Rp , V30
p p
Þ and the given ground and simulation of different natural processes effective for com-
motion data ðX, YÞ. puterized search and optimization. They are rough computational
models of natural evolution. Survival of the fittest and evolution of
the population of species through generations is an effective search
Development of Target Response Spectra
strategy. The GAs start by initiating a randomly generated pop-
The GP procedure is repeated for all spectral periods T of interest to ulation of chromosomes, which, when decoded, create the op-
generate a probabilistic target hazard spectrum for a given scenario timization variables in a search space. The chromosomes must
(M p , Rp , V30
p
). In the examples here, np 5 41 periods are selected compete for survival in an environment with the measure of their
between T 5 0:01 and 10:00 s. The methodology for GP regression fitness being used to rank their optimality. Each chromosome is,
described in the “GPs for Response Spectra Regression” section is therefore, tested in a fitness function based on the optimization
easy to set up. Posterior prediction should be done only for in- objective function to determine its likelihood for survival to the next
terpolation within the data domain, because extrapolation into un- generation. Those that survive, stochastically selected from the
known territory has uncertain accuracy. The set of training data that previous generation, reproduce the population of offspring by the

© ASCE 04014147-7 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 6. Examples of the estimated mean acceleration spectrum from GP regression for two scenario events compared with actual record number
(a) 1818 and (b) 2135 corresponding to these two events; the two test records are not used in the training set; the symmetric 95% confidence intervals
are also shown by vertical bars

actions of crossover and random occasional mutation. The process of Initialize the population (randomly)
creating new generations is continued until a measure of conver- Evaluate initial population (fitness evaluation)
gence to an optimal solution of the fitness function is met. Repeat
In stochastic search, GAs are considered to be a powerful and Perform competitive selection
efficient class of methods. They are efficient, because the action of Apply genetic operators (crossover and mutation) to generate new
the populations of chromosomes (different candidate solutions) can candidate solutions in the population
be done in parallel, which is particularly suitable for high- Evaluate fitness of the candidate solutions in the current
performance computing. They are also powerful, because they im- generation
plement both exploration and exploitation in the search space to Until some convergence criteria is satisfied
locate the optima. As a result, eventual convergence to a solution is
almost always guaranteed if the algorithm is run for a sufficiently The property that individual chromosomes in a population of
large number of generations. a GA can be evaluated in parallel makes GAs particularly suitable
A pseudocode for the GA operations can be given as follows. for parallel computing. The authors used the Mind-Meld server at

© ASCE 04014147-8 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 Caltech’s Center for Advanced Computing Research (CACR) Illustrative Application of Methodology
(2011) for their simulations of ground motion records. When per-
forming the simulations reported here, the Mind-Meld server had To demonstrate the capabilities of the proposed methodology,
eight dual-core 2.8-GHz processors with 1 MB of cache per core. simulations of ground motion acceleration histories at a site in
Fig. 7 shows a flowchart for the GA the authors use for optimal downtown Los Angeles are performed for four scenario events that
estimation of the coefficients of the eigenquakes based on Eq. (12). A are shown in Table 1 by four pairs of a moment magnitude, MW , and
population size of twice the number of the eigenquakes was used in a hypocentral distance, R, with a fixed site V30 value. Two of the
the simulations. The crossover ratio was set to 0.8 (the fraction of the Scenarios (S1 and S2) are for a rare event with a 2% probability of
population, excluding elite chromosomes, created by the crossover exceedance in 50 years, and the other two (S3 and S4) are for
operation), and a Gaussian mutation was used (to alter the chro- a relatively frequent event with a 50% probability of exceedance in
mosomes by adding a random number from a Gaussian distribution 50 years.
to explore the space). The value of these parameters was chosen For each probability of exceedance, ground motion simulations
empirically by trial and error. The performance of a GA depends on are presented corresponding to two different periods: 0.2 s repre-
these parameters, and general strategies for fine-tuning the perfor- senting a typical short structure (approximately a 2-story building)
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

mance of GAs can be found in Goldberg (1989). and 2.0 s representing a midrise building (approximately a 20-story
building). Event Scenarios can be obtained from deaggregation of
the seismic hazard for the site. The authors used the USGS (2008)
interactive deaggregations calculator, which uses the coordinates of
a site, the spectral period, and the site’s soil condition (V30 ), in
addition to a probability of exceedance, to generate a set of Sce-
narios. The authors choose no for the GMPE Deaggs option, so that
the mean hazard is deaggregated independently from any GMPE.
The process of generating the ground motion simulations and the
results are discussed next.
The site is located at the intersection of 7th Street and Metro in
downtown Los Angeles with a latitude of 34.04869 and a longitude
of 2118:258775 (Fig. 8). The mean values of magnitude and dis-
tance are used for the simulations and are shown in Table 1. There are
two sets of simulations for Scenarios S2 and S4, because they were
more challenging during high-dimensional optimization of the
coefficients of the eigenquakes. Therefore, the number of required
eigenquakes was lowered from 100 in Scenario S2a to 20 in Scenario
S2b and from 40 in Scenario S4a to 20 in Scenario S4b for better
convergence.
A total of 266 earthquake records within a radius of 0.5 (ap-
proximately 50 km) of the site were collected from the PEER NGA
database and used in the GP regression to arrive at mean spectral
acceleration ordinates for different Scenarios. The data that were
utilized included events with moment magnitudes from 4.27 to 7.36
recorded from seismic sources between 8.88 and 226.98 km away
(not to be confused with the 50-km radius for selection of the stations
around the site). The source-to-site distance can be larger than 50 km.
The GP spectra results are shown in Fig. 9 (solid curves). Note the
different spectral shapes that are present at this site for different
Scenarios. Given a GA to determine the optimal set of coefficients,
simulation of ground motion histories can be done simply by a linear
combination of the eigenquakes as in Eq. (12). The resulting
pseudoacceleration spectra (PSA) are shown in Fig. 9 (dashed
curves) for the four Scenarios, and the corresponding acceleration
Fig. 7. Flowchart of the ground motion simulation procedure time histories are shown in Fig. 10. The difficulty in arriving at

Table 1. Scenarios Considered for Ground Motion Simulation

Probability of exceedance Elapsed Minimum Number of Number of
Scenario in 50 years (%) T (s) M R (km) V30 (m=s) time (min) function value generations eigenquakes
S1 2 0.2 6.65 6.6 388 327 0.14 48 40
S2a 2.0 6.91 11.3 388 536 0.44 87 100
S2b 504 0.31 81 20
S3 50 0.2 6.64 23.8 388 123 0.04 17 40
S4a 2.0 7.07 43.8 388 256 0.98 37 40
S4b 416 0.18 67 20
Note: The last four columns are for the GA optimization. All Scenarios are located in downtown Los Angeles.

© ASCE 04014147-9 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 8. Location of the stations used in generating ground motion simulations for Scenarios S1–S4 (©2011 Google, Map data ©2011 Google)

Fig. 9. Target and simulated response spectra for downtown Los Angeles using optimal combinations of the eigenquakes

a global solution for optimization of the coefficients of the eigen- procedure. Instead, in the first part, the concept of eigenquakes
quakes for Scenarios S2 and S4 is evident in Fig. 9. from the PCA of a ground motion database is introduced and is
shown to be effective in characterizing the database as a set of
Concluding Remarks orthonormal basis vectors that can be used for simulation of ground
motion records. In the second part of the procedure, GP regression
A three-part procedure for ground motion simulation based on ap- is used to estimate the response spectrum of the ground motion
plying machine-learning techniques to a strong-motion database expected at a site for a given scenario event based on previously
is presented. The problem of selection and scaling of recorded recorded strong-motion data. In the third part, the ground motion
ground motions for design purposes that has been the subject of is simulated by a superposition of eigenquakes that optimally
many controversies in the past decade is avoided in the proposed matches the estimated response spectrum from the second part.

© ASCE 04014147-10 J. Eng. Mech.

J. Eng. Mech. 2015.141.

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

Fig. 10. Simulated ground acceleration time histories for downtown Los Angeles

The proposed procedure is particularly suitable for large urban NGA program through Dr. Yousef Bozorgnia of the University of
areas where dense strong ground motion instrumentation exists California at Berkeley and by Professor Masumi Yamada of Kyoto
or is expected to grow, thereby producing increasing numbers of University. Professor Sami Masri of the University of Southern
records over time. California provided great advice during this study for which both
Several improvements of the proposed procedure could be the authors are thankful.
subject of future work. A Bayesian framework is desirable for
updating the eigenquakes when new records are added to a strong- References
motion database. Variational PCA (Bishop 1999) is particularly
suitable for this purpose, because it would allow for a Bayesian Abrahamson, N., et al. (2008). “Comparisons of the NGA ground-motion
probabilistic formulation of the classical PCA. Automatic selection relations.” Earthq. Spectra, 24(1), 45–66.
of the appropriate model dimensionality (the number of eigenquakes Abrahamson, N., and Silva, W. (2008). “Summary of the Abrahamson &
to retain) would then be possible. A drawback of the third part of the Silva NGA ground-motion relations.” Earthq. Spectra, 24(1), 67–97.
procedure is its reliance on robust global optimization in high- Ahmadi, G. (1979). “Generation of artificial time histories compatible with
dimensional spaces using a GA. It was found that sometimes the given response spectra—A review.” Solid Mech. Arch., 4(3), 207–239.
GA spent most of the computational time around local optima. Ahmadi, G. (1980). “A note on the Wiener-Hermite representation of the
Alternative optimization schemes, such as regularization to induce earthquake ground acceleration.” Mech. Res. Commun., 7(1), 7–13.
sparsity in the eigenquake superposition, could be explored. Another Alimoradi, A. (2011). “Earthquake ground motion simulation using novel
variant would be to select the eigenquake superposition to take into machine learning tools.” EERL 2011-01, Earthquake Engineering Re-
account the ranks of the eigenquakes based on the size of their PCA search Laboratory, California Institute of Technology, Pasadena, CA.
eigenvalues rather than treating all eigenquakes uniformly as in the Alimoradi, A., Miranda, E., Taghavi, S., and Naeim, F. (2006). “Evolu-
tionary modal identification utilizing coupled shear-flexural response—
current formulation.
Implication for multistory buildings. Part I: Theory.” Struct. Des. Tall
Spec. Build., 15(1), 51–65.
Alimoradi, A., and Naeim, F. (2006). “Evolutionary modal identification
Acknowledgments
utilizing coupled shear-flexural response—Implication for multistory
buildings. Part II: Application.” Struct. Des. Tall Spec. Build., 15(1), 67–
The Earthquake Engineering Research Institute (EERI)/FEMA Na- 103.
tional Earthquake Hazards Reduction Program (NEHRP) Profes- Alimoradi, A., Pezeshk, S., and Foley, C. M. (2007). “Probabilistic
sional Fellowship and a California Institute of Technology (Caltech) performance-based optimal design of steel moment-resisting frames. II:
Visiting Associateship awarded to the first author provided financial, Applications.” J. Struct. Eng., 10.1061/(ASCE)0733-9445(2007)133:
logistical, and technical support for this project. Their support is 6(767), 767–776.
greatly appreciated by both authors. Caltech’s Center for Advanced Alimoradi, A., Pezeshk, S., Naeim, F., and Frigui, H. (2005). “Fuzzy pattern
Computing Research (CACR) was instrumental in providing needed classification of strong ground motion records.” J. Earthquake Eng.,
computational resources for the parallel processing tasks involved 9(3), 307–332.
in this research, and their assistance is greatly appreciated. The work Amin, M., and Ang, A. H.-S. (1968). “Nonstationary stochastic models of
in this paper used recorded data that were provided by the PEER- earthquake motions.” J. Engrg. Mech. Div., 94(2), 559–584.

© ASCE 04014147-11 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 Anderson, J. G. (2010). “Source and site characteristics of earthquakes that Chan, E. (1997). “Optimal design of building structures using genetic
have caused exceptional ground accelerations and velocities.” Bull. algorithms.” EERL-97-06, Earthquake Engineering Research Labora-
Seismol. Soc. Am., 100(1), 1–36. tory, California Institute of Technology, Pasadena, CA.
ASCE. (2006). “Minimum design loads for buildings and other structures.” Cheung, S. H., and Beck, J. L. (2010). “Calculation of posterior probabilities
7-05, Reston, VA. for Bayesian model class assessment and averaging from posterior
ASCE. (2011). “Report card for America’s infrastructure.” Æhttp://www samples based on dynamic system data.” Comput. Aided Civ. Infrastruct.
.infrastructurereportcard.org/æ (Apr. 2011). Eng., 25(5), 304–321.
Atkinson, G. M., and Beresnev, I. (2002). “Ground motions at Memphis and Chiou, B. S. J., and Youngs, R. R. (2008). “An NGA model for the average
St. Louis from M 7.5–8.0 earthquakes in the New Madrid seismic zone.” horizontal component of peak ground motion and response spectra.”
Bull. Seismol. Soc. Am., 92(3), 1015–1024. Earthq. Spectra, 24(1), 173–215.
Atkinson, G. M., and Silva, W. (1997). “An empirical study of earthquake Choi, Y., Stewart, J. P., and Graves, R. W. (2005). “Empirical model for
source spectra for California earthquakes.” Bull. Seismol. Soc. Am., basin effects accounts for basin depth and source location.” Bull.
87(1), 97–113. Seismol. Soc. Am., 95(4), 1412–1427.
Baker, J. W. (2011). “Conditional mean spectrum: Tool for ground-motion Chou, J.-H., and Ghaboussi, J. (2001). “Genetic algorithm in structural
selection.” J. Struct. Eng., 10.1061/(ASCE)ST.1943-541X.0000215, damage detection.” Comput. Struct., 79(14), 1335–1353.
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

322–331. Conte, J. P., Pister, K. S., and Mahin, S. A. (1992). “Nonstationary ARMA
Baker, J. W., and Cornell, C. A. (2005). “A vector-valued ground motion modeling of seismic motion.” Soil Dyn. Earthquake Eng., 11(7), 411–
intensity measure consisting of spectral acceleration and epsilon.” 426.
Earthquake Eng. Struct. Dynam., 34(10), 1193–1217. Cornell, C. A. (1968). “Engineering seismic risk analysis.” Bull. Seismol.
Baker, J. W., and Cornell, C. A. (2006). “Spectral shape, epsilon and record Soc. Am., 58(5), 1583–1606.
selection.” Earthquake Eng. Struct. Dynam., 35(9), 1077–1095. Delavaud, E., Scherbaum, F., Kuehn, N., and Riggelsen, C. (2009).
Baker, J. W., and Cornell, C. A. (2008). “Vector-valued intensity measures “Information-theoretic selection of ground-motion prediction equations
for pulse-like near-fault ground motions.” Eng. Struct., 30(4), 1048– for seismic hazard analysis: An applicability study using Californian
1057. data.” Bull. Seismol. Soc. Am., 99(6), 3248–3263.
Bazzurro, P., and Cornell, C. A. (1999). “Disaggregation of seismic hazard.” Der Kiureghian, A., and Crempien, J. (1989). “An evolutionary model for
Bull. Seismol. Soc. Am., 89(2), 501–520. earthquake ground motion.” Struct. Saf., 6(2–4), 235–246.
Beck, J. L. (2010). “Bayesian system identification based on probability Foley, C. M., Pezeshk, S., and Alimoradi, A. (2007). “Probabilistic
logic.” Struct. Contr. Health Monit., 17(7), 825–847. performance-based optimal design of steel moment-resisting frames. I:
Beck, J. L., and Katafygiotis, L. S. (1998). “Updating models and their Formulation.” J. Struct. Eng., 10.1061/(ASCE)0733-9445(2007)133:
uncertainties. I: Bayesian statistical framework.” J. Eng. Mech., 10.1061/ 6(757), 757–766.
(ASCE)0733-9399(1998)124:4(455), 455–461. Geller, R. J. (2011). “Shake-up time for Japanese seismology.” Nature,
Beyer, K., and Bommer, J. J. (2007). “Selection and scaling of real 472(7344), 407–409.
accelerograms for bi-directional loading: A review of current practice Giaralis, A., and Spanos, P. D. (2009). “Wavelet-based response spectrum
and code provisions.” J. Earthquake Eng., 11(S1), 13–45. compatible synthesis of accelerograms—Eurocode application (EC8).”
Bishop, C. M. (1999). “Variational principal components.” Proc., 9th Int. Soil. Dyn. Earthquake Eng., 29(1), 219–235.
Conf. on Neural Networks (ICANN’99), Vol. 1, Institution of Engi- Goldberg, D. E. (1989). Genetic algorithms in search, optimization, and
neering and Technology (IET), Stevenage, U.K., 509–514. machine learning, Addison-Wesley, Boston.
Bishop, C. M. (2006). Pattern recognition and machine learning, Springer, Grigoriu, M. (2011). “To scale or not to scale seismic ground-acceleration
New York. records.” J. Eng. Mech., 10.1061/(ASCE)EM.1943-7889.0000226,
Bommer, J. J., and Abrahamson, N. A. (2006). “Why do modern proba- 284–293.
bilistic seismic-hazard analyses often lead to increased hazard estimates?” Gu, P., and Wen, Y. K. (2007). “A record-based method for the generation of
Bull. Seismol. Soc. Am., 96(6), 1967–1977. tridirectional uniform hazard-response spectra and ground motions using
Bommer, J. J., and Abrahamson, N. A. (2007a). “Reply to ‘Comment on the Hilbert-Huang transform.” Bull. Seismol. Soc. Am., 97(5), 1539–
“Why do modern probabilistic seismic-hazard analyses often lead to 1556.
increased hazard estimates?” by Julian J. Bommer and Norman A. Hanks, T. C., and McGuire, R. K. (1981). “The character of high-frequency
Abrahamson’ by Jens-Uwe Klugel.” Bull. Seismol. Soc. Am., 97(6), strong ground motion.” Bull. Seismol. Soc. Am., 71(6), 2071–2095.
2208–2211. Hough, S. E., et al. (2010). “Localized damage caused by topographic
Bommer, J. J., and Abrahamson, N. A. (2007b). “Reply to ‘Comment on amplification during the 2010 M7.0 Haiti earthquake.” Nat. Geosci.,
“Why do modern probabilistic seismic-hazard analyses often lead to 3(11), 778–782.
increased hazard estimates?” by Julian J. Bommer and Norman A. Housner, G. W. (1975). “Measures of severity of earthquake ground
Abrahamson’ by Zhenming Wang and Mai Zhou.” Bull. Seismol. Soc. shaking.” Proc., U.S. National Conf. on Earthquake Engineering,
Am., 97(6), 2215–2217. Earthquake Engineering Research Institute, Oakland, CA, 25–33.
Boore, D. M. (1983). “Stochastic simulation of high-frequency ground Housner, G. W., and Jennings, P. C. (1964). “Generation of artificial
motions based on seismological models of the radiated spectra.” Bull. earthquakes.” J. Engrg. Mech. Div., 90(1), 113–152.
Seismol. Soc. Am., 73(6A), 1865–1894. Housner, G. W., and Jennings, P. C. (1982). Earthquake design criteria,
Boore, D. M., and Atkinson, G. M. (2008). “Ground-motion prediction Earthquake Engineering Research Institute, Oakland, CA.
equations for the average horizontal component of PGA, PGV, and 5%- Idriss, I. M. (2008). “An NGA empirical model for estimating the horizontal
damped PSA at spectral periods between 0.01 s and 10.0 s.” Earthq. spectral values generated by shallow crustal earthquakes.” Earthq.
Spectra, 24(1), 99–138. Spectra, 24(1), 217–242.
Bozorgnia, Y., and Bertero, V. V. (2004). Earthquake engineering: From Jennings, P. C., Housner, G. W., and Tsai, N. C. (1969). “Simulated
engineering seismology to performance-based engineering, CRC Press, earthquake motions for design purposes.” Proc., 4th World Conf. on
Boca Raton, FL. Earthquake Engineering, Chilean Association on Seismology and
Campbell, K. W., and Bozorgnia, Y. (2008). “NGA ground motion model Earthquake Engineering, Santiago, Chile.
for the geometric mean horizontal component of PGA, PGV, PGD and Jolliffe, I. T. (2002). Principal component analysis, Springer, New York.
5% damped linear elastic response spectra for periods ranging from 0.01 Kim, Y.-J., and Ghaboussi, J. (2001). “Direct use of design criteria in
to 10 s.” Earthq. Spectra, 24(1), 139–171. genetic algorithm-based controller optimization.” Earthquake Eng.
Center for Advanced Computing Research (CACR). (2011). “Caltech Struct. Dynam., 30(9), 1261–1278.
Center for Advanced Computing Research.” Æhttp://www.cacr.caltech Klugel, J.-U. (2007). “Comment on ‘Why do modern probabilistic seismic-
.edu/main/æ (May 2011). hazard analyses often lead to increased hazard estimates?’ by Julian J.

© ASCE 04014147-12 J. Eng. Mech.

J. Eng. Mech. 2015.141.

 Bommer and Norman A. Abrahamson.” Bull. Seismol. Soc. Am., 97(6), Raich, A. M., and Ghaboussi, J. (2000). “Evolving structural design sol-
2198–2207. utions using an implicit redundant genetic algorithm.” Struct. Multi-
Klugel, J.-U. (2009). “Comment on ‘Sigma: Issues, insights and challenges’ disciplin. Optim., 20(3), 222–231.
by F. O. Strasser, N. A. Abrahamson, and J. J. Bommer.” Seismol. Res. Rasmussen, C. E. (2006). “Advances in Gaussian processes.” Advances in
Lett., 80(3), 494–498. Neural Information Processing Systems 19, MIT Press, Cambridge, MA.
Kottke, A., and Rathje, E. M. (2008). “A semi-automated procedure for Rasmussen, C. E., and Williams, C. K. I. (2006). Gaussian processes for
selecting and scaling recorded earthquake motions for dynamic analy- machine learning, MIT Press, Cambridge, MA.
sis.” Earthq. Spectra, 24(4), 911–932. Rezaeian, S., and Der Kiureghian, A. (2008). “A stochastic ground motion
Kramer, S. L. (1996). Geotechnical earthquake engineering, Prentice Hall, model with separable temporal and spectral nonstationarities.” Earth-
Upper Saddle River, NJ.
quake Eng. Struct. Dynam., 37(13), 1565–1584.
Krawinkler, H. (2001). Progress and challenges in performance-based
Saragoni, G. R., and Hart, G. C. (1973). “Simulation of artificial earth-
earthquake engineering, Earthquake Engineering Research Institute,
quakes.” Earthquake Eng. Struct. Dynam., 2(3), 249–267.
Oakland, CA.
Scherbaum, F., Delavaud, E., and Riggelsen, C. (2009). “Model selection in
Luco, N., and Bazzurro, P. (2007). “Does amplitude scaling of ground
seismic hazard analysis: An information-theoretic perspective.” Bull.
motion records result in biased nonlinear structural drift responses?”
Earthquake Eng. Struct. Dynam., 36(13), 1813–1835. Seismol. Soc. Am., 99(6), 3234–3247.
Downloaded from ascelibrary.org by University of Leeds on 05/17/15. Copyright ASCE. For personal use only; all rights reserved.

MATLAB 2009 [Computer software]. Natick, MA, MathWorks. Shahbazian, A., and Pezeshk, S. (2010). “Improved velocity and dis-
McGuire, R. K. (2004). Seismic hazard and risk analysis, Earthquake placement time histories in frequency domain spectral-matching pro-
Engineering Research Institute, Oakland, CA. cedures.” Bull. Seismol. Soc. Am., 100(6), 3213–3223.
McGuire, R. K. (2008). “Probabilistic seismic hazard analysis: Early his- Shome, N., Cornell, C. A., Bazzurro, P., and Carballo, J. E. (1998). “Earth-
tory.” Earthquake Eng. Struct. Dynam., 37(3), 329–338. quakes, records, and nonlinear responses.” Earthq. Spectra, 14(3), 469–500.
Mobarakeh, A. A., Rofooei, F. R., and Ahmadi, G. (2002). “Simulation of Song, S. G., and Somerville, P. (2010). “Physics-based earthquake source
earthquake records using time-varying Arma (2,1) model.” Probab. Eng. characterization and modeling with geostatistics.” Bull. Seismol. Soc.
Mech., 17(1), 15–34. Am., 100(2), 482–496.
Musson, R. M. W. (2009). “Ground motion and probabilistic hazard.” Bull. Strasser, F. O., and Bommer, J. J. (2009). “Review: Strong ground motions
Earthquake Eng., 7(3), 575–589. —Have we seen the worst?” Bull. Seismol. Soc. Am., 99(5), 2613–2637.
Naeim, F., Alimoradi, A., and Pezeshk, S. (2004). “Selection and scaling of USGS. (2008). “2008 interactive deaggregations.” Æhttp://geohazards.usgs
ground motion time histories for structural design using genetic algo- .gov/deaggint/2008/æ (Sep. 2013).
rithms.” Earthq. Spectra, 20(2), 413–426. USGS. (2011). “ANSS—Advanced National Seismic System.” Æhttp://
Newmark, N. M., and Hall, W. J. (1982). Earthquake spectra and design, earthquake.usgs.gov/monitoring/anss/æ (Jun. 2011).
Earthquake Engineering Research Institute, Oakland, CA. Wang, M., and Takada, T. (2009). “A Bayesian framework for prediction of
Pacific Earthquake Engineering Research Center (PEER). (2011a). “Ground seismic ground motion.” Bull. Seismol. Soc. Am., 99(4), 2348–2364.
motion selection and modification program.” Æhttp://peer.berkeley.edu/ Wang, Z., and Zhou, M. (2007). “Comment on ‘Why do modern proba-
gmsm/index.htmlæ (Apr. 2011). bilistic seismic-hazard analyses often lead to increased hazard estimates?’
Pacific Earthquake Engineering Research Center (PEER). (2011b). “NGA
by Julian J. Bommer and Norman A. Abrahamson.” Bull. Seismol. Soc.
database.” Æhttp://peer.berkeley.edu/nga/æ (Jan. 2011).
Am., 97(6), 2212–2214.
Papadimitriou, K. (1990). “Stochastic characterization of strong ground
Watson-Lamprey, J., and Abrahamson, N. (2006). “Selection of ground
motion and application to structural response.” Ph.D. thesis, California
motion time series and limits on scaling.” Soil. Dyn. Earthquake Eng.,
Institute of Technology, Pasadena, CA.
Papadimitriou, K., and Beck, J. L. (1990). “Nonstationary stochastic 26(5), 477–482.
characterization of strong-motion accelerograms.” Proc., 4th U.S. Na- Yamada, M., Mori, J., and Heaton, T. (2009). “The slapdown phase in high-
tional Conf. on Earthquake Engineering, Earthquake Engineering Re- acceleration records of large earthquakes.” Seismol. Res. Lett., 80(4),
search Institute, El Cerrito, CA. 559–564.
Papadimitriou, K., and Beck, J. L. (1992). “Stochastic characterization of Yamamoto, Y., and Baker, J. W. (2011). “Stochastic model for earthquake
ground motion and applications to structural response.” Proc., 10th ground motions using wavelet packets.” Proc., 11th Int. Conf. on
World Conf. on Earthquake Engineering, Vol. 2, Balkema, Rotterdam, Applications of Statistics and Probability in Civil Engineering, CRC
Netherlands. Press, Boca Raton, FL.

© ASCE 04014147-13 J. Eng. Mech.

J. Eng. Mech. 2015.141.