Journal of Earthquake Engineering

ISSN: 1363-2469 (Print) 1559-808X (Online) Journal homepage: http://www.tandfonline.com/loi/ueqe20

An Improved Algorithm for Selecting Ground

Motions to Match a Conditional Spectrum

Jack W. Baker & Cynthia Lee

To cite this article: Jack W. Baker & Cynthia Lee (2017): An Improved Algorithm for Selecting
Ground Motions to Match a Conditional Spectrum, Journal of Earthquake Engineering, DOI:
10.1080/13632469.2016.1264334

To link to this article: http://dx.doi.org/10.1080/13632469.2016.1264334

Accepted author version posted online: 10

Jan 2017.

Submit your article to this journal

View related articles

View Crossmark data

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

http://www.tandfonline.com/action/journalInformation?journalCode=ueqe20

Download by: [The UC San Diego Library] Date: 11 January 2017, At: 00:28
 1 An Improved Algorithm for Selecting Ground
2 Motions to Match a Conditional Spectrum
3 Jack W. Baker (Corresponding author) and Cynthia Lee

4 Department of Civil & Environmental Engineering, Stanford University, USA

t
ip
5 This paper describes an algorithm to efficiently select ground motions from a database while

cr
6 matching a target mean, variance and correlations of response spectral values at a range of

7 periods. The approach improves an earlier algorithm by Jayaram et al. (2011). Key steps in the

us
8 process are to screen a ground motion database for suitable motions, statistically simulate

9 response spectra from a target distribution, find motions whose spectra match each statistically
an
10 simulated response spectrum, and then perform an optimization to further improve the

11 consistency of the selected motions with the target distribution. These steps are discussed in
M

12 detail, and the computational expense of the algorithm is evaluated. A brief example selection

13 exercise is performed, to illustrate the type of results that can be obtained. Source code for the
ed

14 algorithm has been provided, along with metadata for several popular databases of recorded and
pt

15 simulated ground motions, which should facilitate a variety of exploratory and research studies.
ce

16 1 Introduction
Ac

17 Selection of ground motions is a topic of great interest as dynamic structural analysis, which

18 requires ground motions as inputs, grows more prevalent (Katsanos et al. 2010; NIST 2011).

19 This selection typically involves searching a ground motion database to find time series

20 produced under appropriate seismological conditions (e.g., earthquake magnitude and source-to-

1
21 site distance), and that have appropriate response spectral values. In some cases, ground motions

22 are selected based on their individual match to a target spectrum; that is, an optimal set of ground

23 motions would have spectra that all perfectly match the target spectrum. In other cases, however,

24 it is important that the ground motions have variability in response spectra that accurately

t
25 represents target distributions from predictive models (e.g., Kramer and Mitchell 2006; Lin et al.

ip
26 2013b). As such, a number of algorithms have been proposed to select ground motions with

cr
27 some form of specified response spectral variability (Bradley 2012; Ha and Han 2016a; b;

us
28 Jayaram et al. 2011; Kottke and Rathje 2008; Wang 2011). Among those algorithms, only

29 Bradley, Ha and Han (2016b) and Jayaram et al. include two features of interest here: accounting
an
30 for correlations among spectral parameters and conditioning on a target spectral acceleration

31 amplitude.
M

32 Traditional practice in active seismic regions has been to search databases of ground motion

33 recordings, but simulated ground motions are receiving increased use. Further, there is a need for
ed

34 comparative research studies where recorded and simulated motions are selected in a comparable
pt

35 manner and their relative impacts on structures are evaluated (e.g., Galasso et al. 2013; Iervolino

36 et al. 2010). In recognition of these trends, data facilitating searches of several popular libraries
ce

37 of recorded and simulated ground motions are provided with this algorithm. A second trend in

38 ground motion libraries is that they are rapidly growing larger (several databases discussed
Ac

39 below have more than 10,000 ground motions), making the computational efficiency of search

40 algorithms more important.

2
41 This manuscript describes an updated version of the algorithm proposed by Jayaram et al.

42 (2011), also utilizing aspects of Bradley (2012). Relative to the Jayaram et al. algorithm, the

43 range of selection options has been broadened and the numerical implementation has been

44 improved to both reduce runtime and improve the statistics of the resulting selected motions.

t
45 Improvements relative to the previous algorithm are noted below, and improvements in

ip
46 numerical efficiency are also reported.

cr
47 2 Target Response Spectra

2.1 Types of Spectral Targets

us
48
an
49 Before discussing the ground motion selection procedure, we first introduce some relevant
M

50 terminology and concepts related to response spectra as targets for ground motion selection.

51 Ground motion models (GMMs) (e.g., Boore et al. 2014) provide the mean and standard
ed

52 deviation of logarithmic spectral acceleration (Sa) at a given period, denoted here as μ ln Sa (Rup,T )
pt

53 and σ ln Sa (Rup,T ), respectively. With this notation, μ denotes a mean, and σ denotes a standard

54 deviation, of the variable noted in subscript. Rup denotes the rupture scenario (defined by the
ce

55 earthquake’s magnitude, distance, rupture mechanism, and other parameters necessary to

Ac

56 evaluate a given GMM) and T denotes the spectral acceleration period. The GMM prediction

57 also generally depends upon one or more parameters defining site conditions such as average

58 shear-wave velocity over the top 30 m of the site (Vs30), but that explicit dependence is omitted

59 from this notation for brevity. Some GMMs (e.g., Abrahamson et al. 2014) also provide

3
60 correlation coefficients for log spectral accelerations at pairs of periods, denoted here as

61 ρ (Ti ,T j ) . If not provided by the GMM, the correlation coefficients can be obtained from a

62 supplemental model (e.g., Baker and Jayaram 2008).

t
63 With the above inputs, we define an “Unconditional Spectrum” as the probability distribution

ip
64 of a response spectrum, given a rupture scenario. The distribution of log spectral acceleration

cr
65 values at multiple periods, given a rupture, is well represented by a multivariate normal

66 distribution (Jayaram and Baker 2008), which is fully specified by the mean and covariance

us
67 matrix for lnSa values an
68 M = [μ ln Sa (Rup,T1 ) μ ln Sa (Rup,T2 ) ... μ ln Sa (Rup,Tp )]T (1)
M

69 (2)
ed
pt

70 where M is a vector of mean values of lnSa at p periods of interest, superscript T denotes a

71 matrix transpose, and Σ is the covariance matrix for lnSa at these same periods. In equation 2
ce

72 we adopt abbreviated notation, σ T ,T = ρ (Ti ,T j )σ ln Sa ( Rup,Ti )σ ln Sa (Rup,T j ) , to denote the covariance

i j
Ac

73 of lnSa at periods Ti and Tj (and σ T2 = σ ln Sa (Rup,Ti )2 is the variance at period Ti).

i

74 The “Unconditional” terminology is used here to emphasize the lack of conditioning on a

75 spectral value, for consistency with the use of the term “Conditional” in the following two

76 definitions. An example Unconditional Spectrum is illustrated in Figure 1a. The mean value
4

4
77 from equation 1, and the standard deviations embedded in equation 2, are plotted in Figure 1a;

78 the period-to-period correlation embedded in equation 2 is apparent in the ground motion spectra

79 plotted in the figure, which are ‘bumpy’ (reflecting a lack of perfect correlation) but do vary with

80 some continuity from period to period.

t
ip
81 The “Conditional Mean Spectrum” (CMS) quantifies mean log spectral acceleration values

cr
82 of a ground motion, conditional on a spectral value at a conditioning period and a rupture

83 scenario

us
84 μ ln Sa(T )|ln Sa(T *) = μ ln Sa (Rup,Ti ) + ρ (Ti ,T *)ε (T *)σ ln Sa (Rup,Ti ) (3)
i
an
85 where T* denotes the conditioning period and ε (T *) is a residual quantifying the difference
M
86 between the conditioning Sa value (Sa(T*)) and its mean value given the considered rupture

ln Sa(T *) − μ ln Sa (Rup,T *)
87 ε (T *) = (4)
ed

σ ln Sa (Rup,T *)

88 A CMS, as calculated using equation 3 and a conditioning period of 1.5s, is illustrated in Figure
pt

89 1b. The term CMS was introduced by Baker and Cornell (2006), and further background is
ce

90 provided in Baker (2011). It is becoming more commonly used to select ground motions for

91 dynamic analysis in several design guidelines (e.g., BSSC 2015; FEMA 2012; TBI Guidelines
Ac

92 Working Group 2010).

93 The “Conditional Spectrum” (CS) is the probability distribution of log spectral acceleration

94 values, conditional on a spectral value at a conditioning period and on a rupture scenario. Unlike

5
 95 the CMS, this spectrum quantifies variability in spectral values at periods other than the

96 conditioning period. If we assume that the distribution is multivariate normal (which is generally

97 reasonable), then the Conditional Spectrum is fully described by conditional means, and a

98 conditional covariance matrix. The conditional means are given by equation 3 and the

t
99 conditional covariance matrix is

ip
cr
Σ cross Σ Tcross
100 Σ cond = Σ − (5)
σ ln Sa (Rup,T *)2

us
101 where Σ is the covariance matrix from equation 2 and Σ cross is a p x 1 matrix of covariances
an
102 between ln Sa(Ti ) and ln Sa(T *) . Visually we can represent this distribution by plotting the mean

103 and +/- one or two standard deviations around the mean, as in Figure 1b. The Conditional
M

104 Spectrum terminology was coined by Abrahamson and Al Atik (2010), but they represented the

105 CS distribution directly by realizations of the spectra rather than an analytical distribution. The
ed

106 Conditional Mean Spectrum was more popular than the Conditional Spectrum prior to

107 approximately 2010, in large part because there was no simple way to select ground motions
pt

108 matching a Conditional Spectrum—a situation rectified by this manuscript and its predecessor
ce

109 algorithms.

110 Bradley (2010) extended the CS to consider ground motion parameters other than response
Ac

111 spectra, and to consider a more general situation where more than one rupture scenario may

112 contribute to occurrence of ground motions with the target amplitude. Bradley refers to the

113 resulting distribution and selection procedure as a Generalized Conditional Intensity Measure

114 (GCIM) approach. This paper focuses on response spectra and single rupture scenarios for
6

6
115 simplicity, but the algorithm could, in principle, be generalized by defining equations 3 and 5 to

116 refer to a general vector of intensity measures and to reflect the impact of multiple rupture

117 scenarios on the target means and covariances. Equations 3 and 5 can also be revised to account

118 for the use of multiple GMMs, consistent with current practice in hazard analysis (Lin et al.

t
119 2013a).

ip
cr
120 To complement the above equations, a few observations may provide intuition about the

121 Conditional Spectrum target illustrated in Figure 1b. First, the response spectra “pinch” to a

us
122 single point at the conditioning period of 1.5s. Since we have specified this amplitude, there is no

123 variability in the spectra at this period. Second, at other periods there is variability in the spectra
an
124 and that variability tends to be larger at periods further from 1.5s. This is a result of the

125 correlation between spectral values: periods close to 1.5s have spectra highly correlated to
M

126 Sa(1.5s), so there is relatively little uncertainty in spectra at these nearby periods, while there is

127 larger spectral variability at the (less-correlated) periods far from 1.5s. This pattern in spectral
ed

128 variability is grossly similar to what is observed if one simply scales a set of ground motions so
pt

129 their spectra are equal at some conditioning period, which somewhat confirms the

130 reasonableness of this target. Third, the mean of the conditional spectrum (i.e., the CMS) reflects
ce

131 the expected response spectral shape, and it accounts for both the spectrum associated with the

132 rupture scenario (via the unconditional mean spectrum of equation 1) and the tendency for high-
Ac

133 amplitude spectral values to be associated with a peak in the spectrum (via the epsilon value and

134 spectral correlations in equation 3).

7
135 It should be intuitive that mean responses obtained from structural analysis are related to the

136 mean amplitude of the input ground motions’ spectra. Further, several studies have shown that

137 considering the full variability in response spectra, rather than only mean values of spectra, can

138 be important for some structural response assessment procedures (e.g., Lin et al. 2013b). This

t
139 motivates the development of tools like that proposed here to facilitate selection of ground

ip
140 motions matching an Unconditional Spectrum or Conditional Spectrum.

cr
141 2.2 Computational Challenges

us
142 With the above methods for quantifying response spectra targets established, here we briefly
an
143 consider ground motion selection strategies (thinking of the case where one wishes to select n

144 ground motions from a database with m candidate motions). It is simple to quickly select ground
M

145 motions to be consistent with a Conditional Mean Spectrum, as only a mean spectrum is

146 relevant: one can simply compute an error metric for each of the m candidate motions (e.g., the
ed

147 sum of squared errors between the ground motion’s spectrum and the target spectrum, potentially

148 after scaling the motion), and then selects the n motions with the smallest error. There may be a
pt

149 benefit to performing a more extensive optimization-based search as discussed below, but the
ce

150 availability of this fast approach makes the optimization-based approach less critical.
Ac

151 It is much more difficult to select ground motions that match a Conditional Spectrum or

152 Unconditional Spectrum target, because the ground motions’ spectra should match both a mean

153 target and a covariance matrix. To find ground motions with an appropriate covariance matrix,

154 one cannot evaluate individual candidates but instead must evaluate a set of n candidates

155 collectively (that is, it is not possible to determine whether an individual ground motion is a
8

8
156 “good” fit to a Conditional Spectrum without knowing the other ground motions it would be

157 paired with). This means that there are m-choose-n combinations of ground motions to

158 consider—too many to search exhaustively for typical situations where n>10 and m>100. It is

159 this problem that the algorithm below addresses, via heuristics to enable a fast search that

t
160 consistently produces ground motion sets closely matching the specified targets.

ip
cr
161 3 Selection Algorithm

us
162 The major steps of the proposed ground motion selection process are illustrated in Figure 2.

163 These steps are described in additional detail in the following subsections, which group the steps
an
164 into distinct conceptual stages.
M

165 3.1 Compute Target Spectrum and Statistically Simulate Realizations

ed

166 The process starts by specifying a target response spectrum (Step 1 in Figure 2). Formally,

167 we specify the mean values of log spectral acceleration, and the covariance matrix for these
pt

168 values. Equations 1 and 2 are utilized if the target is an Unconditional Spectrum, and equations 3

169 and 5 are utilized if the target is a Conditional Spectrum. The provided software includes a
ce

170 function that computes the required mean and covariance matrix if the user specifies the target
Ac

171 earthquake rupture (and the target ε (T *) or Sa(T *) if the target is a Conditional Spectrum). The

172 provided software utilizes the models of Boore et al. (2014) and Baker and Jayaram (2008) to

173 compute the target spectra, but these can be replaced without significant modification to the code

174 if desired.

9
175 While the software was developed to solve the problem of selecting ground motions to match

176 a Conditional Spectrum, it can also be adopted for selecting ground motions to match a code

177 spectrum or some other target, by specifying the target spectrum as the mean spectrum, and

178 setting the covariance matrix to consist of all zeros (i.e., specifying that no variability around the

t
179 target spectrum is desired).

ip
cr
180 Relative to Jayaram et al. (2011), there are a few updates to the target response spectrum

181 calculation. The main program has been generalized so that a single function can handle

us
182 Conditional Spectrum and Unconditional Spectrum targets (previously, separate versions of the

183 software were provided for each target type). Additionally, functionality has been provided so
an
184 that the user can easily match a target Sa(T*) with a given set of rupture values. Because mean

185 rupture values (rather than the full distribution of possible ruptures considered in a seismic
M

186 hazard calculation) are often used in these calculations for convenience, the target Sa(T*) is not

necessarily obtained when these mean values are combined with a mean ε (T *) value from a
ed

187

188 hazard deaggregation. An optional calculation now adjusts the ε (T *) value so that the
pt

189 conditional mean spectrum matches the target Sa(T*), as this was seen by Lin et al. (2013) to be

190 a reasonable approximation strategy in many cases.

ce

191 Step 2 in Figure 2 is to statistically simulate realizations of response spectra from the target
Ac

192 distribution. This is done by sampling from a multivariate normal distribution with the target

193 mean and covariance matrices. Since this simulation step is extremely fast, it is performed

194 multiple times and the set of simulations best matching the target spectrum is utilized for the

195 following steps. We note here that the ‘statistically simulated spectra’ in this step are produced

10

10
196 by sampling from a probability distribution (e.g., Stein 1987); this is distinct from the ‘simulated

197 ground motions’ discussed in the following section, where are produced by numerical evaluation

198 of equations associated with the earthquake rupture and seismic wave propagation process.

3.2 Specify Candidate Ground Motions

t
199

ip
200 Step 3 of the process in Figure 2 specifies candidate ground motions to select from. Relevant

cr
201 metadata from a candidate ground motion database is loaded, including spectral acceleration

us
202 values and rupture parameters for each ground motion. The Jayaram et al. (2011) code included

203 metadata for the NGA-West1 database, consisting of 3551 ground motions from 173 earthquakes
an
204 (Chiou et al. 2008). Here we have added metadata for the NGA-West2 database, which includes

205 21,539 ground motions from 599 earthquakes (Ancheta et al. 2014). Additionally, we have added
M

206 metadata for three databases of numerically simulated ground motions produced by a Southern

207 California Earthquake Center (SCEC) project to validate simulations (Goulet et al. 2015).
ed

208 Simulations were produced on the SCEC Broadband Platform using rupture geometries from

209 seven recent California earthquakes. Ground motions from the “EXSIM” (Atkinson and
pt

210 Assatourians 2015), “GP” (Graves and Pitarka 2015) and “SDSU” (Olsen and Takedatsu 2015)
ce

211 simulation algorithms were compiled for use in this software. Each database includes 13,400

212 ground motions. Because both recorded and simulated ground motion databases are provided in a
Ac

213 compatible format, the authors hope that this tool will facilitate further comparative evaluations

214 of similarities and differences in structural demands caused by recorded versus simulated ground

215 motions with comparable response spectra.

11

11
216 An additional improvement in Steps 2 and 3 of the selection process is that the new target

217 computations and ground motion databases utilize both the RotD50 and RotD100 direction-

218 independent metrics of response spectra for multi-component motions (Boore 2010). These

219 metrics are now used often in ground motion models and engineering analysis procedures

t
220 (Stewart et al. 2011), so their inclusion in the database metadata increases the tool’s relevance.

ip
221 Single-component response spectra are also provided so that users can search for single-

cr
222 component motions if desired.

us
223 Once database metadata has been loaded, it is screened in Step 4 so that only appropriate

224 ground motions are considered for selection. The current code is set up to allow only ground
an
225 motions with appropriate values of earthquake magnitude, source-to-site distance, and Vs30, but

226 the screening can be easily generalized to consider other properties. These so-called causal
M

227 parameters are important to screen in order to assure that the considered time series are

228 reasonably consistent with the conditions of interest in ground motion selection, but they should
ed

229 not be screened so aggressively that an insufficient number of candidate motions remain for the
pt

230 next stage of selection (Tarbali and Bradley 2016). The Jayaram et al. (2011) code did not

231 include this screening step, as its objective was to illustrate other aspects of the selection
ce

232 procedure, but the screening has been added here both to improve the quality of the selected

233 motions and to improve the computational cost of the calculation (since motions excluded at this
Ac

234 stage need not be considered later for selection).

12

12
235 3.3 Ground Motion Selection

236 Step 5 of Figure 2 involves selecting ground motions from the database that best match the

237 statistically simulated spectra. For each statistically simulated spectrum and candidate ground

t
238 motion, the sum of squared errors (SSE) is computed

ip
P
SSE =  ( ln S a (T j ) − ln Sa( s ) (T j ) )
2

cr
239 (6)
j =1

us
240 where ln Sa (Tj ) is the log spectral acceleration of the (optionally scaled) candidate ground
an
241 motion and ln Sa(s) (T j ) is the lnSa of the considered statistically simulated response spectrum.

242 Note that if scaling is not allowed and a target Conditional Spectrum is used, the selected
M
243 motions will not exactly match the target Sa(T*), but equation 6 will encourage selection of

244 motions close to the target and the motions may be suitably similar if choosing from a database
ed

245 having ground motions compatible with the target scenario. For each statistically simulated

246 spectrum, the SSE is computed for all candidate ground motions that have not already been
pt

247 selected, and the motion with the smallest SSE is selected to represent that simulation. The
ce

248 metric of equation 6 is not the only possible selection criterion (e.g., Beyer and Bommer 2007;

249 Buratti et al. 2010), but has been observed to produce satisfactory results; it could easily be
Ac

250 modified by a user if desired (e.g., to put varying weights on the squared errors at varying

251 periods).

252 Simulating spectra from the target distribution (in Step 2), and then searching individual

253 motions to find matches to these simulations, is perhaps the most important step in this algorithm

13

13
254 for overcoming the computational cost that would be required to search suites of ground motions

255 instead of individual motions. Its utility is apparent when noting that most prior algorithms to

256 solve this problem have used this approach (Bradley 2012; Jayaram et al. 2011; Wang 2011),

257 though Ha and Han (2016a) recently proposed a non-simulation-based approach that instead uses

t
258 a limited search of the potential selection combinations.

ip
cr
259 In Step 6 of Figure 2, the selected suite of motions is evaluated to see whether it is

260 sufficiently close to the target distribution. The maximum percentage mismatch of the mean and

us
261 standard deviation of the selected motions’ spectra, relative to their targets, are calculated
an
 m (T ) − μ (T ) 
ln Sa j ln Sa j
262 Errmean = max   × 100 (7)
j  μ (T ) 
 ln Sa j
M

 s (T ) − σ (T ) 
ln Sa j ln Sa j
263 Errstd = max   × 100 (8)
j  σ ln Sa (T j ) 
ed

264 where mln Sa (T j ) is the sample mean of the lnSa values of the selected motions at period Tj and
pt

265 μ ln Sa (T j ) is the target mean from a GMM (in the case of Unconditional Selection) or equation 5
ce

266 (in the case of Conditional Selection). Similarly, sln Sa (T j ) is the sample standard deviation and
Ac

267 σ ln Sa (T j ) is the target standard deviation at period Tj.. Finally, denotes an absolute value. The

268 user can specify a maximum tolerance for the errors defined by equations 7 and 8, and if the

269 errors at this step are less than the tolerance then the selection process is complete.

14

14
270 If the errors are too large, then a finite number of optimization rounds are performed to

271 further improve the selection. Step 7 of Figure 2 involves further optimizing the initial selection

272 if needed. At this stage, the selected set of ground motions are modified by replacing individual

273 ground motions from the set with available motions from the screened database and seeing

t
274 whether the set is improved in its match to the target response spectrum. There are two objective

ip
275 functions available to the user when performing the optimization.

cr
276 In the first case, a weighted sum (over all periods of interest) of squared errors in the

us
277 spectra’s mean values and standard deviations are utilized to evaluate goodness of fit, as follows
an
p

( ) ( )
SSEs =   mln Sa (T j ) − μ ln Sa (T j ) + w sln Sa (T j ) − σ ln Sa (T j ) 
2 2
278 (9)

j=1 

M

279 where SSES denotes the sum of squared errors of the set of ground motions, w is a user-defined

280 weight that assigns relative importance to mismatches in mean versus standard deviation values.
ed

281 In the second case, the d statistic from a Komogorov-Smirnov goodness of fit test (KS test) is
pt

282 used as the metric. The d statistic at a given period is given by the maximum absolute values of

283 the difference between the target distribution’s cumulative distribution function ( Fln Sa(T ) (x) )
ce

284 and the empirical cumulative distribution function ( Fln* Sa(T ) (x) ) given by the selected motions’
j
Ac

285 Sa(Tj) values

286 (
d(T j ) = max Fln* Sa(T ) (x) − Fln Sa(T ) (x)
j j
) (10)

15

15
287 Example d values are illustrated in Figure 3 for a single period. In the software, the d values for

288 all periods of interest are summed and used as the error metric. This metric has been successfully

289 utilized by others in ground motion selection problems (Bradley 2012; Chandramohan et al.

290 2016), and so is incorporated into this algorithm to take advantage of this insight.

t
ip
291 fFigure 3 illustrates the implications of these error metrics using a target distribution for

cr
292 Sa(2s), and Sa(2s) values for two hypothetical sets of candidate ground motions. The target and

293 sample means are labeled on each figure, as well as the d(2s) value; sample standard deviations

us
294 (the third parameter used in error computations) are not shown. In Figure 3a, the candidate

295 motions have nearly the same mean and standard deviation as the target spectrum, and a d(2s)
an
296 value of 0.14. In Figure 3b the candidate ground motions have a sample mean that is 4% larger

297 than the target mean (and, although not shown graphically, the sample standard deviation closely
M

298 matches the target). However, the d(2s)=0.07 value in this case is half of that in Figure 3a. So the

299 SSES error metric would prefer the motions in Figure 3a while the KS test metric would prefer
ed

300 the motions in Figure 3b (keeping in mind that the selection algorithm sums errors across
pt

301 multiple periods rather than considering only a single period as we have here). An additional

302 relevant factor is that the KS test calculation is somewhat slower than the SSES calculation, as
ce

303 will be seen below. Ultimately it is left to the user to select a metric for a given selection case, as

304 both error metrics have merits.

Ac

305 This a greedy optimization algorithm, as it searches for only local improvements (i.e., by

306 replacing only one ground motion at a time), and thus misses opportunities for improvements

307 resulting from replacing two or more ground motions that cause an improvement in aggregate (if

16

16
308 the improvement was not detectable when replacing them one at a time). While such

309 opportunities surely exist, it is not clear that considering them would result in dramatically

310 improved selection results. Further, this choice to use a greedy optimization approach makes the

311 calculation computationally tractable. The optimization algorithm can terminate early if the error

t
312 metrics of equations 7 and 8 are within tolerance.

ip
cr
313 Once a final set of ground motions is determined, in Step 8 of Figure 2 an output file is

314 produced to document the selected ground motions (and scale factors, if scaling was allowed).

us
315 For the NGA-West1 and simulated ground motions, the ground motion time series are available

316 for direct download over the Internet and the selected motions’ URLs are provided in the output
an
317 so that the time series can be obtained automatically. For the NGA-West2 database, the index

318 numbers of the selected ground motions must be copied into the NGA-West2 search tool in order
M

319 to download the ground motion files. Additional detail regarding this download process is

320 provided in the software documentation.

ed

321 4 COMPUTATIONAL EXPENSE

pt
ce

322 The algorithm’s initial record-by-record selection (Step 5) takes time proportional to m x n,

323 where m is the number of candidate ground motions in the database and n is the number of
Ac

324 selected ground motions, because the m motions are compared once each to the n simulations.

325 The optimization (Step 7) also takes time proportional to m x n, because the m-n candidates are

326 evaluated as candidates to replace each of the n previously selected motions. This is much better

327 than the m-choose-n computational expense of the exhaustive search discussed in section 2.2.

17

17
328 Any calculations within the loops over the m x n candidate evaluations were optimized to limit

329 computational cost to the extent possible.

330 We further manage computational expense in a few ways. First, we screen the database (Step

331 4 in Figure 2) to limit the size of m before starting the search for motions to select. While this is

t
ip
332 conceptually simple, it was not implemented in the Jayaram et al. (2011) software. Second,

cr
333 within the optimization stage, we skip all ground motions that need to be scaled by a larger-than-

334 allowable factor, before proceeding to the more expensive calculation of considering the ground

us
335 motion as a potential replacement for a currently selected motion. These first two steps often lead

336 to great reductions in the numbers of considered ground motions when typical restrictions on
an
337 acceptable ground motions and scale factors are used. Third, we stop the optimization early if a

338 selected set of ground motions is sufficiently close to the target spectrum (as evaluated using a
M

339 user-specified error tolerance). Finally, the optimization step of the algorithm—the most

340 expensive step—is optionally parallelized so that each currently selected ground motion is
ed

341 evaluated in parallel to see if a better alternative can be found.

pt

342 The computational cost of the algorithm is illustrated in Figure 4, for varying sizes of
ce

343 selected ground motion sets and varying sizes of the database being searched. A few

344 observations can be made. First, the improvements discussed here have greatly reduced the
Ac

345 algorithm’s run time relative to the previous implementation by Jayaram et al. (2011). A

346 somewhat typical problem of selecting 20 ground motions from a pool of 2000 candidates now

347 requires less than 30 seconds. Second, the cost of the algorithm now scales approximately

18

18
348 linearly with the number of selected ground motions (Figure 4a) and the number of searchable

349 ground motions (Figure 4b), as expected based on the discussion above.

350 The Figure 4 results are produced without allowing early termination of the optimization and

351 without parallelized optimization, in order to provide conservative run times, so users may find

t
ip
352 better performance depending upon their use of these features. The benefits of these features are

cr
353 problem dependent, so no general run time results are provided here. Further speed-up of the

354 algorithm is possible, most obviously by switching to a faster programing language, and by

us
355 restructuring some code for better numerical performance. We consider the current

356 implementation to be well suited, however, for our goals of providing an educational code with
an
357 reasonable run times.
M

358 5 Example Ground Motion Selection

ed

359 To illustrate use of the software, selection of simulated and recorded motions is briefly

360 demonstrated, referring to the numbered steps in Figure 2. In step 1, the target spectrum must be
pt

361 specified; the Conditional Spectrum of Figure 1b is used here as the target RotD50 spectrum.
ce

362 Step 2 statistically simulates realizations of spectra from this distribution, and no user choices

363 need to be made. In step 3, the ground motion database must be specified; here we consider two
Ac

364 alternatives—the NGA-West1 database of recorded ground motions and the Graves and Pitarka

365 database of simulated motions discussed above. In step 4 these databases are screened for

366 suitable ground motions. In this case, because the target spectrum is associated with a magnitude

367 = 6.5, distance = 10 km event, we restricted selection to consider only ground motions within 50

19

19
368 km of an earthquake with magnitude between 6 and 7. Scaling of ground motions was allowed,

369 but scale factors were limited to a maximum of five. These criteria are somewhat consistent with

370 typical ground motion selection procedures (e.g., NIST 2011), but are used here simply to

371 illustrate the selection process and are not recommended values. With these criteria, the NGA-

t
372 West1 database had 582 ground motions satisfying the initial screening, and the GP database had

ip
373 6000.

cr
374 With the initial criteria and database screening performed, ground motions can then be

us
375 selected. In step 5, initial ground motions were selected from each database to match simulated

376 spectra—a step which requires no user choices. In step 6, a maximum 10% error in means and
an
377 standard deviations of the spectral was specified. Neither database satisfied that criterion with the

378 initial selection (the step 6 check), so optimization (step 7) was performed in both cases. The
M

379 SSE objective function of equation 9 was chosen for the optimization. For the NGA-West1

380 database, the max errors in mean and standard deviation were 8% and 34%, respectively, before
ed

381 optimization and 6% and 7% after. For the Graves and Pitarka database, the max errors were
pt

382 41% and 32% before optimization and 29% and 26% after.
ce

383 In step 8, the final selections of ground motions are output. Figure 5 shows the response

384 spectra of the selected motions. We see that, although the Graves and Pitarka selection had
Ac

385 significantly worse error metrics in the previous paragraph, the response spectra plot does not

386 indicate serious deficiencies in general (the periods with poor metrics are the short periods at the

387 left of Figure 5b); these errors would need to be evaluated on a project-specific basis for

388 acceptability. Figure 6 shows example velocity time series for both sets of selected motions. The

20

20
389 two sets of selected motions’ time series are comparable under a cursory visual inspection. These

390 two sets of ground motions would provide useful inputs for a more detailed study of any subtler

391 differences in the recordings and simulations that lead them to produce different structural

392 responses when used as input to a dynamic analysis problem. Such a study is beyond the scope

t
393 of this paper but is the type of exercise that the new software features are intended to facilitate.

ip
cr
394 6 Numerical Implementation

us
395 Matlab source code for the algorithm is available at

396 https://github.com/bakerjw/CS_Selection. The repository includes metadata and documentation

an
397 for five ground motion databases discussed above. In addition to the algorithmic features
M
398 discussed above, this source code also improves upon the previous code from Jayaram et al.

399 (2011) by providing a single general purpose program (the Jayaram et al. code consistent of four
ed

400 separate programs for conditional versus unconditional selection, and single-component versus

401 two-component selection), and better modularization of functions related to the tasks outlined in
pt

402 Figure 2.
ce

403 7 Conclusions
Ac

404 This paper has presented an efficient algorithm for selecting ground motions from a database

405 that match a target response spectrum distribution (i.e., a Conditional Spectrum or Unconditional

406 Spectrum). The motivation for this work is that when the target spectrum has a distribution,

407 rather than a single value, it is not possible to evaluate individual ground motions for selection

21

21
408 without considering them as part of a suite of ground motions that collectively represent the

409 distribution. But evaluating all possible suites of ground motions is impossible when considering

410 large ground motion databases typical in practice today. This algorithm utilizes several heuristics

411 to quickly identify ground motion sets with close match to the target spectrum.

t
ip
412 Note that the code can also be easily adapted for the more common situation where the user

cr
413 wants to select ground motions that each closely match a target spectrum (e.g., for closely

414 matching a design spectrum from a building code, with each motion closely matching the target)

us
415 by specifying the target spectrum as the mean value and setting the target variances to zero. The

416 algorithm’s complexity is more than is needed for this simpler application, but it is well-suited
an
417 for the problem and so a user-defined flag setting the variance to zero is provided in the software.
M
418 The key steps in this algorithm are (1) compute a target response spectrum distribution, (2)

419 statistically simulate response spectra from the target distribution, (3) load and screen a database
ed

420 of candidate ground motions, (4) select ground motions from the database that individually

421 match the statistically simulated spectra, (5) make incremental changes to the initially selected
pt

422 ground motion set to further optimize its fit to the target spectrum distribution. The algorithm
ce

423 follows the general structure of a proposal by Jayaram et al. (2011), but incorporates a number of

424 new features that improve its utility and speed.

Ac

425 Data for a number of new ground motion databases are also provided, to allow users to

426 search recently developed catalogs of recorded or simulated ground motion data. Example

427 selection of comparable recorded and simulated ground motions are illustrated, to demonstrate

428 the feasibility of selecting equivalent motions from differing sources. Matlab source code for the

22

22
429 algorithm has been provided publically for readers interested in using or modifying the

430 algorithm.

431 Acknowledgements

t
432 We thank Nenad Bijelic, Reagan Chandramohan, Beliz Ugurhan and three anonymous reviewers

ip
433 for helpful advice and comments on this work. This work was supported in part by the State of

cr
434 California through the Transportation Systems Research Program of the Pacific Earthquake

us
435 Engineering Research Center (PEER).

436 References
an
437 Abrahamson, N. A., and Al Atik, L. (2010). “Scenario Spectra for Design Ground Motions and
M
438 Risk Calculation.” 9th US National and 10th Canadian Conference on Earthquake Engineering,

439 Toronto, Canada, 12.

ed

440 Abrahamson, N. A., Silva, W. J., and Kamai, R. (2014). “Summary of the ASK14 Ground

441 Motion Relation for Active Crustal Regions.” Earthquake Spectra, 30(3), 1025–1055.
pt

442 Ancheta, T. D., Darragh, R. B., Stewart, J. P., Seyhan, E., Silva, W. J., Chiou, B. S.-J.,
ce

443 Wooddell, K. E., Graves, R. W., Kottke, A. R., Boore, D. M., Kishida, T., and Donahue, J. L.

444 (2014). “NGA-West2 Database.” Earthquake Spectra, 30(3), 989–1005.

Ac

445 Atkinson, G. M., and Assatourians, K. (2015). “Implementation and validation of EXSIM (a

446 stochastic finite-fault ground-motion simulation algorithm) on the SCEC broadband platform.”

447 Seismological Research Letters, 86(1), 48–60.

23

23
448 Baker, J. W. (2011). “Conditional Mean Spectrum: Tool for ground motion selection.” Journal

449 of Structural Engineering, 137(3), 322–331.

450 Baker, J. W., and Cornell, C. A. (2006). “Spectral shape, epsilon and record selection.”

451 Earthquake Engineering & Structural Dynamics, 35(9), 1077–1095.

t
ip
452 Baker, J. W., and Jayaram, N. (2008). “Correlation of spectral acceleration values from NGA

cr
453 ground motion models.” Earthquake Spectra, 24(1), 299–317.

454 Beyer, K., and Bommer, J. J. (2007). “Selection and Scaling of Real Accelerograms for Bi-

us
455 Directional Loading: A Review of Current Practice and Code Provisions.” Journal of
an
456 Earthquake Engineering, 11, 13–45.

457 Boore, D. M. (2010). “Orientation-Independent, Nongeometric-Mean Measures of Seismic

M
458 Intensity from Two Horizontal Components of Motion.” Bulletin of the Seismological Society of

459 America, 100(4), 1830–1835.

ed

460 Boore, D. M., Stewart, J. P., Seyhan, E., and Atkinson, G. M. (2014). “NGA-West2 Equations

461 for Predicting PGA, PGV, and 5% Damped PSA for Shallow Crustal Earthquakes.” Earthquake
pt

462 Spectra, 30(3), 1057–1085.

ce

463 Bradley, B. A. (2010). “A generalized conditional intensity measure approach and holistic

464 ground-motion selection.” Earthquake Engineering & Structural Dynamics, 39(12), 1321–1342.
Ac

465 Bradley, B. A. (2012). “A ground motion selection algorithm based on the generalized

466 conditional intensity measure approach.” Soil Dynamics and Earthquake Engineering, 40, 48–

467 61.

24

24
468 BSSC. (2015). NEHRP Recommended Seismic Provisions for New Buildings and Other

469 Structures. FEMA P-1050, Building Seismic Safety Council, Washington, D.C., 515p.

470 Buratti, N., Stafford, P. J., and Bommer, J. J. (2010). “Earthquake Accelerogram Selection and

471 Scaling Procedures for Estimating the Distribution of Drift Response.” Journal of Structural

t
ip
472 Engineering, 137(3), 345–357.

cr
473 Chandramohan, R., Baker, J. W., and Deierlein, G. G. (2016). “Impact of hazard-consistent

474 ground motion duration in structural collapse risk assessment.” Earthquake Engineering and

us
475 Structural Dynamics, (in press). an
476 Chiou, B., Darragh, R., Gregor, N., and Silva, W. (2008). “NGA Project Strong-Motion

477 Database.” Earthquake Spectra, 24(1), 23–44.

M
478 FEMA. (2012). Seismic Performance Assessment of Buildings. FEMA P-58, Prepared by

479 Applied Technology Council for the Federal Emergency Management Agency.
ed

480 Galasso, C., Zhong, P., Zareian, F., Iervolino, I., and Graves, R. W. (2013). “Validation of

481 ground-motion simulations for historical events using MDoF systems.” Earthquake Engineering
pt

482 & Structural Dynamics, 42(9), 1395–1412.

ce

483 Goulet, C. A., Abrahamson, N. A., Somerville, P. G., and Wooddell, K. E. (2015). “The SCEC

484 Broadband Platform Validation Exercise: Methodology for Code Validation in the Context of
Ac

485 Seismic‐Hazard Analyses.” Seismological Research Letters, 86(1), 17–26.

486 Graves, R., and Pitarka, A. (2015). “Refinements to the Graves and Pitarka (2010) Broadband

487 Ground‐Motion Simulation Method.” Seismological Research Letters, 86(1), 75–80.

25

25
488 Ha, S. J., and Han, S. W. (2016a). “A method for selecting ground motions that considers target

489 response spectrum mean and variance as well as correlation structure.” Journal of Earthquake

490 Engineering, (in press).

491 Ha, S. J., and Han, S. W. (2016b). “An efficient method for selecting and scaling ground motions

t
ip
492 matching target response spectrum mean and variance.” Earthquake Engineering & Structural

493 Dynamics, 45(8), 1381–1387.

cr
494 Iervolino, I., De Luca, F., and Cosenza, E. (2010). “Spectral shape-based assessment of SDOF

us
495 nonlinear response to real, adjusted and artificial accelerograms.” Engineering Structures, 32(9),

496 2776–2792.
an
497 Jayaram, N., and Baker, J. W. (2008). “Statistical Tests of the Joint Distribution of Spectral
M
498 Acceleration Values.” Bulletin of the Seismological Society of America, 98(5), 2231–2243.

499 Jayaram, N., Lin, T., and Baker, J. W. (2011). “A computationally efficient ground-motion
ed

500 selection algorithm for matching a target response spectrum mean and variance.” Earthquake

501 Spectra, 27(3), 797–815.

pt

502 Katsanos, E. I., Sextos, A. G., and Manolis, G. D. (2010). “Selection of earthquake ground
ce

503 motion records: A state-of-the-art review from a structural engineering perspective.” Soil

504 Dynamics and Earthquake Engineering, 30(4), 157–169.

Ac

505 Kottke, A., and Rathje, E. M. (2008). “A Semi-Automated Procedure for Selecting and Scaling

506 Recorded Earthquake Motions for Dynamic Analysis.” Earthquake Spectra, 24(4), 911–932.

507 Kramer, S. L., and Mitchell, R. A. (2006). “Ground Motion Intensity Measures for Liquefaction

508 Hazard Evaluation.” 22(2), 413–438.

26

26
509 Lin, T., Harmsen, S. C., Baker, J. W., and Luco, N. (2013a). “Conditional Spectrum

510 Computation Incorporating Multiple Causal Earthquakes and Ground Motion Prediction

511 Models.” Bulletin of the Seismological Society of America, 103(2A), 1103–1116.

512 Lin, T., Haselton, C. B., and Baker, J. W. (2013b). “Conditional spectrum-based ground motion

t
ip
513 selection. Part I: Hazard consistency for risk-based assessments.” Earthquake Engineering &

514 Structural Dynamics, 42(12), 1847–1865.

cr
515 NIST. (2011). Selecting and Scaling Earthquake Ground Motions for Performing Response-

us
516 History Analyses. NIST GCR 11-917-15, Prepared by the NEHRP Consultants Joint Venture for

517 the National Institute of Standards and Technology, Gaithersburg, Maryland.

an
518 Olsen, K., and Takedatsu, R. (2015). “The SDSU Broadband Ground‐Motion Generation
M
519 Module BBtoolbox Version 1.5.” Seismological Research Letters, 86(1), 81–88.

520 Stein, M. (1987). “Large Sample Properties of Simulations Using Latin Hypercube Sampling.”
ed

521 Technometrics, 29(2), 143–151.

522 Stewart, J. P., Abrahamson, N. A., Atkinson, G. M., Baker, J. W., Boore, D. M., Bozorgnia, Y.,
pt

523 Campbell, K. W., Comartin, C. D., Idriss, I. M., Lew, M., Mehrain, M., Moehle, J. P., Naeim, F.,
ce

524 and Sabol, T. A. (2011). “Representation of Bi-Directional Ground Motions for Design Spectra

525 in Building Codes.” Earthquake Spectra, 27(3), 927–937.

Ac

526 Tarbali, K., and Bradley, B. A. (2016). “The effect of causal parameter bounds in PSHA-based

527 ground motion selection.” Earthquake Engineering & Structural Dynamics, (in press).

528 TBI Guidelines Working Group. (2010). “Guidelines for performance-based seismic design of

529 tall buildings.” Berkeley: University of California (PEER Report No. 2010/05).
27

27
530 Wang, G. (2011). “A ground motion selection and modification method capturing response

531 spectrum characteristics and variability of scenario earthquakes.” Soil Dynamics and Earthquake

532 Engineering, 31(4), 611–625.

533 Figure 1. (a) Unconditional Spectrum associated with magnitude = 6.5, distance = 10 km, Vs30 =

t
534 500 m/s and a California strike-slip rupture. (b) Conditional spectrum associated with the same

ip
535 rupture parameters as (a), and with Sa(1.5s) = 0.3 g. Response spectra consistent with the target
536 distributions are also shown. Calculations use the GMM of Boore et al. (2014) and the

cr
537 correlation model of Baker and Jayaram (2008).

us
an
M
ed

538
539
pt
ce
Ac

28

28
540 Figure 2. Flow chart of major steps in the ground motion selection process, with relevant
541 equation numbers noted in parantheses. Details for each step are discussed in Section 3.

t
ip
cr
us
an
M
ed

542
543
pt
ce
Ac

29

29
544 Figure 3. Illustration of spectral acceleration target distributions and error metrics. (a) Selected
545 ground motions whose mean and standard deviation closely match the target. (b) Selected ground
546 motions whose mean is slightly less than the target mean, but who more closely match the target
547 with regard to the d statistic.

t
ip
cr
us
548
549 an
M
ed
pt
ce
Ac

30

30
550 Figure 4. (a) Time to select a given number of ground motions from a database of 5000 motions,
551 for three different selection approaches. (b) Time to select 20 ground motions from a database of
552 5000, with varying numbers of motions remaining after screening.

t
ip
553
554

cr
us
an
M
ed
pt
ce
Ac

31

31
555 Figure 5. Response spectra of selected ground motions with spectra matching the target
556 Conditional Spectrum of Figure 1b. The mean and mean +/- two standard deviations of the target
557 lnSa distribution are superimposed. (a) NGA-West1 recorded ground motions. (b) Graves and
558 Pitarka simulated ground motions.

t
ip
cr
559

us
an
560
M
ed
pt
ce
Ac

32

32
561 Figure 6. Example velocity time series from the selected motions shown in Figure 5. The time
562 series are scaled to have the same peak ground velocity values for ease of comparison, so no
563 velocity scale is provided. (a) NGA-West1 recorded ground motions. (b) Graves and Pitarka
564 simulated ground motions.

t
ip
cr
us
565
an
M
ed

566
567
pt
ce
Ac

33

33