The Professional Journal of the Earthquake Engineering Research Institute

PREPRINT
This preprint is a PDF of a manuscript that has been accepted for
publication in Earthquake Spectra. It is the final version that was
uploaded and approved by the author(s). While the paper has been
through the usual rigorous peer review process for the Journal, it has
not been copyedited, nor have the figures and tables been modified for
final publication. Please also note that the paper may refer to online
Appendices that are not yet available.

We have posted this preliminary version of the manuscript online in

the interest of making the scientific findings available for distribution
and citation as quickly as possible following acceptance. However,
readers should be aware that the final, published version will look
different from this version and may also have some differences in
content.

The DOI for this manuscript and the correct format for citing the paper are
given at the top of the online (html) abstract.

Once the final, published version of this paper is posted online, it will
replace the preliminary version at the specified DOI.
 1 Non-Ergodic Site Response in Seismic Hazard
2 Analysis
a)
3 Jonathan P. Stewart M.EERI, Kioumars Afshari b) S.M.EERI, and Christine A.
4 Goulet c) M.EERI

5 Abstract: Probabilistic seismic hazard analyses are usually performed with semi-
6 empirical ground motion models (GMMs) following the ergodic assumption
7 whereby average source, path, and site effects from global databases apply for a
8 specific site of interest. Site-specific site response is likely to differ from the
9 global average conditional on site parameters used in GMMs (typically VS30 and
10 basin depth). Non-ergodic site response can be evaluated using on-site ground
11 motion recordings and/or one-dimensional wave propagation analyses, and allows
12 site-to-site variability to be removed from the within-event standard deviation.
13 Relative to ergodic, non-ergodic hazard analyses often reduce ground motions at
14 long return periods. We describe procedures for replacing the site term in GMMs
15 with a non-ergodic nonlinear mean over its appropriate range of periods
16 (returning to the ergodic mean outside that range). We also present procedures for
17 computing non-ergodic standard deviation by removing site-to-site variability
18 while considering effects of soil nonlinearity. We illustrate application of these
19 procedures, and their effect on hazard curves and uniform hazard spectra, as
20 implemented in OpenSHA.

21 Keywords: non-ergodic, site response, probabilistic seismic hazard analysis,

22 earthquake, ground motion.

23

a)
Professor and Chair, UCLA, Civil & Environmental Engineering Dept., Los Angeles, CA, 90095.
[email protected]
b)
Graduate Student, UCLA, Civil & Environmental Engineering Dept., Los Angeles, CA, 90095. [email protected]
c)
Executive Science Director for Special Projects, Southern California Earthquake Center, USC, Los Angeles, CA,
90089. [email protected]

1
24 INTRODUCTION

25 The vast majority of the path length for seismic waves traveling from source-to-site occurs
26 through rock in the earth’s crust. As the waves approach the surface, they travel through geologic
27 strata having progressively slower seismic velocities. This will tend to bend the wave
28 propagation direction upward, per Snell’s Law, and change the amplitude. These and other
29 effects of the local geology and morphology of the site on the ground motions are collectively
30 referred to as site effects. Several phenomena have the potential to contribute to site effects:

31 1. Local ground response describes the effects on ground motion of relatively shallow
32 sediments (typically tens to hundreds of m in depth) having the slowest velocities.
33 Because the dimensions of these soft sediments are limited, the affected frequencies are
34 typically relatively high (>∼1 Hz). Factors contributing to local ground response will
35 include some combination of impedance effects, soil nonlinearity, and potentially
36 resonance effects.

37 2. Basin effects are related to the deep structure of sediments that are present in many areas.
38 Basins often include soft sediments near the surface that transition with depth to
39 progressively stiffer sediments, including sedimentary rock, before basement conditions
40 (crystalline rock) are encountered. Because the dimensions of basins are often quite large
41 (on the order of several km), the effected frequencies are relatively low (< ∼ 1 Hz).

42 3. Topographic effects are related to irregularities in the ground surface morphology that
43 can produce local amplification. The frequencies affected by topographic effects depend
44 on the scale of the topographic features; a local steep hill will affect higher frequencies
45 than large mountains.

46 Semi-empirical ground motion models (GMMs) are derived from recordings made at
47 accelerograph sites having various combinations of these site response mechanisms.
48 Accordingly, the effects of each mechanism are present in an average sense in GMM predictions,
49 conditional on the considered site parameters, which are typically time-averaged 30 m shear
50 wave velocity (VS30) and some measure of basin depth. Site response predictions derived from
51 global models conditional on such parameters are referred to as ergodic (Anderson and Brune,
52 1999).

2
53 Actual (or non-ergodic) site response will differ from this global average. When the decision
54 is made to consider location-specific site effects, as is common for critical projects,
55 one-dimensional (1D) ground response analyses (GRA) are the most frequently utilized
56 approach. Guidelines for performing such analyses are available elsewhere (NCHRP, 2012;
57 Stewart et al., 2014, hereafter Sea14). The point to be made here is that only some of the physical
58 processes known to produce site effects (essentially, those related to local ground response) can
59 be simulated in 1D GRA, hence error is likely for conditions where other site effects are
60 appreciable. Errors of this sort bias the mean prediction of site response as estimated by GRA.

61 The standard deviation of ground motion estimates is also needed for probabilistic seismic
62 hazard analysis (PSHA). Because ergodic analysis using GMMs consider conditionally averaged
63 site effects, their within-event standard deviation terms include a component of site-to-site
64 variability (e.g., Al Atik et al. 2010). Non-ergodic site response allows this component of
65 variability to be removed, which has been referred to as single-station sigma (Atkinson, 2006).
66 As described further in this paper, when this reduction of standard deviation is taken, epistemic
67 uncertainties in the site amplification model must be considered. A practical question facing
68 many projects is whether site-specific GRA can be considered to reliably estimate non-ergodic
69 site response, thereby justifying the use of a standard deviation model in which site-to-site
70 variability is removed. This issue is discussed in this paper, but we note here that such
71 assumptions have been applied in PSHA for critical projects, including the Pegasos Refinement
72 Project (Renault et al. 2010), the Thyspunt Siting Project (Bommer et al. 2015, Rodriguez-Marek
73 et al. 2014), the Hanford site (Coppersmith et al. 2016), and the South Western US project
74 (GeoPentech, 2015).

75 Ground motion analyses including site effects are performed with varying levels of
76 sophistication. Most common is to perform PSHA for reference site conditions (typically rock),
77 then to deterministically modify the rock motion using the mean site amplification. This is
78 referred to as a hybrid analysis, due to its combination of probabilistic and deterministic
79 methods, which produces a result with an unknown hazard level (e.g., Cramer, 2003; Goulet and
80 Stewart, 2009). Convolution approaches (Bazzurro and Cornell, 2004a; Rathje et al., 2015)
81 provide a more sophisticated modification of the rock hazard, but do not consider changes in
82 standard deviation associated with non-ergodic site response nor differences in controlling
83 sources that occur as site conditions are modified. The aforementioned projects in which non-

3
 84 ergodic site response was considered used convolution with simulation-based site amplification
85 models; as such this approach effectively represents the state-of-practice for non-ergodic PSHA.
86 The only approach for PSHA that rigorously incorporates site amplification effects is to modify
87 the median and standard deviation of ground motion within the hazard integral. To date, this has
88 generally only been possible using GMMs with their site terms; the result is ergodic with its
89 appurtenant issues of possible bias and large standard deviation.

90 In this article, we describe a methodology for PSHA that utilizes a site-specific (non-ergodic)
91 GMM and illustrate its application using capabilities implemented in open-source hazard code
92 OpenSHA (Field et al. 2003). We describe procedures for quantification of site-specific,
93 nonlinear mean site response as derived using GRA results or on-site recordings (the use of
94 which can overcome GRA bias). We describe analysis of within-event standard deviation
95 considering nonlinear effects and removal of the site-to-site variability. Factors contributing to
96 epistemic uncertainty are identified. We then illustrate the proposed approach, and its effect
97 relative to more approximate procedures, on hazard curves and uniform hazard spectra.

98 NOTATION AND PARTITIONING OF GROUND MOTION VARIABILITY

99 PROBABILITY DISTRIBUTION FOR SITE AMPLIFICATION

100 We express site amplification factors (Y) as the ratio of a ground motion intensity measure
101 (IM) on the ground surface (Z) to the value of the same IM on the reference site condition
102 (typically rock), X:

Z
103 Y = or ln Y = ln Z − ln X (1)
X

104 The implementation of site amplification factors in PSHA requires a probability density function
105 for Y, which is usually taken as log-normal. The mean in natural log units is defined using the
106 following nonlinear expression that has proven to be effective for representing X-dependent
107 amplification:

 xIMref + f3 
108 μln Y = Flin + Fnl = f1 + f 2 ln   (2)
 f3 

4
109 Where Flin and Fnl indicate linear and nonlinear model components; f1, f2, and f3 are model
110 parameters; and xIMref is the amplitude of shaking for the reference site condition. The intensity
111 measure for xIMref is often taken as the median, RotD50-component peak ground acceleration
112 (PGA) for rock (where RotD50 refers to the median of all possible rotated horizontal
113 components for a given ground motion; Boore, 2010). Ergodic versions of Eq. (2) include site
114 terms in many GMMs (Abrahamson et al., 2014; Boore et al., 2014; Campbell and Bozorgnia,
115 2014; Chiou and Youngs, 2014). We use Eq. (2) to represent site-specific mean amplification.

116 PARTITIONING OF GROUND MOTION VARIABILITY

117 A particular realization of earthquake ground motion from event i at site j, lnzij, can be viewed as
118 the sum of the mean from a GMM in natural log units, (μlnZ)ij, and an error term,

119 ln z ij = ( μ ln Z )ij + ε ijσ ln Z (3)

120 where ε is a standard normal variate, and σlnZ is the total standard deviation for Z. The GMM
121 mean has mean (ergodic) terms for source or event (FE), path (FP), and site (FS),

122 ( μ ln Z )ij = FE ,i + FP ,ij + FS ,ij (4)

123 The mean terms are written with indices i and j to indicate that they depend on characteristics of
124 the event (e.g., magnitude, focal mechanism) and site (location and site parameters). For any
125 given event and ground motion, the actual source, path, and site effects differ from the ergodic
126 estimate by their respective random effects, denoted ηE,i, ηP,ij, and ηS,j. These random effects
127 represent the bias of the ergodic model for the particular event, source-site path, and site that
128 produced motion zij. Hence, the actual (or non-ergodic) site response is FS,ij + ηS,j, and similar
129 relations apply for the source and path terms.

130 Eq. (3) can be re-written as follows to help visualize the random effects (adapted from
131 Al Atik et al., 2010 with some modification):

132 ln zij = ( μln Z )ij + η E ,i + η P ,ij + η S , j + ε ijφln Y (5)

133 where εij has the same meaning as in Eq. (3) (although the values are now different) and φlnY is a
134 standard deviation term reflecting the variability that remains when these random effects are
135 considered. Effects of site amplification are more clearly expressed by re-writing Eq. (5) as

5
136 ln zij = ( μln X )ij + η E ,i + η P ,ij + μln Y ,ij + ε ijφln Y (6)

137 in which ( μln Z )ij + ηS , j from Eq. (5) is replaced with ( μln X )ij + μln Y ,ij , per Eq. (1).

138 Each of the event, path, and site terms has corresponding standard deviations. Following the
139 notation introduced by Al Atik et al. (2010), the standard deviation of between-event terms, and
140 for repeatable path and site terms are denoted τ, φP2P, and φS2S, respectively. These combine to
141 produce the total standard deviation as follows:

142 σ ln Z = τ 2 + φP22 P + φS22 S + φln2 Y (7)

143 The site-to-site standard deviation (denoted ϕS2S) contributes to the within-event standard
144 deviation (φlnZ) provided by GMMs with ergodic site terms, i.e.,

145 φln Z = φP22 P + φS22 S + φln2 Y (8)

146 Site-to-site variability is not needed when the site response model is non-ergodic, as discussed
147 subsequently. The variability that remains when each of the random effects is accounted for is
148 represented by the φlnY term in Eqs. (5-8). This term strictly represents variability in path and site
149 effects when a non-ergodic model is used. However, prior studies summarized in the Site
150 Response Variability, φlnY section below, indicate that φlnY is dominated by site amplification
151 variability.

152 MEAN SITE RESPONSE

153 Eq. (2) is used herein to represent the natural log mean site response. Expressions of this type
154 can be used in both ergodic and non-ergodic applications. In Eq. (2), the Flin=f1 term represents
155 the weak-motion (visco-elastic) site amplification. The second term in the sum represents the
156 effects of nonlinearity (Fnl); the physical meanings of the f2 and f3 parameters are depicted in
157 Figure 1.

158
159 Figure 1. Schematic depiction of nonlinear component of mean site amplification function. Term f2
160 represents the negative slope of Fnl for xIMref >> f3. Term f3 represents the approximate center of the xIMref
161 range where amplification changes from visco-elastic (independent of xIMref) to log-linearly dependent on
162 xIMref.

6
163 Ergodic site terms take f1 as a function of VS30 and basin depth, while f2 is a function of VS30
164 only. As discussed previously, such site terms can be in error due to site-specific geologic
165 structure that is not captured by VS30 and basin depth parameters. Non-ergodic site terms can be
166 developed from analysis of ground motion recordings made at the site of interest or GRA, as
167 discussed below.

168

7
169 EVALUATION FROM RECORDINGS

170 Analysis of non-ergodic site response from recordings begins with the installation of seismic
171 sensors (accelerometer or seismometer) at or near the target site. Ground motions are then
172 recorded for regional earthquake events, which are of interest when they fall in the
173 magnitude-distance range of applicable GMMs. In the case of active crustal regions, the
174 NGA-West2 GMMs (Bozorgnia et al., 2014) are generally applicable for M > 3 and site-source
175 distances < 400 km. The GMM selected should suitably capture average path effects, otherwise
176 bias in the path term could map to erroneous assessments of site response.

177 Denoting the target site with index j, the intensity measure for event i (following suitable
178 signal processing for noise effects; e.g., Boore and Bommer, 2005) is denoted zij. Residuals (Rij)
179 of these motions relative to a GMM applicable for the tectonic regime are then computed as:

180 Rij = ln zij − ( μ ln Z )ij (9)

181 The residuals in Eq. (9) contain random effects from each event i (ηE,i), which should be
182 removed to evaluate the within-event residual (δWij),

183 δ Wij = Rij − η E ,i (10)

184 The evaluation of ηE,i is non-trivial. For events considered in the development of GMMs, ηE,i is a
185 product of mixed-effects regression procedures, which are commonly used. For events not
186 considered in GMM development, a posteriori estimates of ηE,i (denoted ηˆE ,i ) are needed using

187 available recordings of event i (this requires data from multiple sites, which can typically be
188 retrieved using local sensor networks such as Southern California Seismic Network and Berkeley
189 Digital Seismic Network in southern and northern California, respectively). Appendix A
190 (electronic supplement) describes alternate procedures for this calculation, which consider a
191 variety of factors including the relative sizes of τ and φlnZ, recording-to-recording correlation,
192 and the number of records (Ni). We recommend computation of ηˆE ,i using a Bayesian inference

193 approach (Eq. A6, modified from Stafford 2012). As illustrated in Appendix A, ηˆE ,i from Eq.

194 (A6) nearly matches the mean of event residuals when the number of observations Ni > ∼30.

195 Figure 2 shows example results for 13 recordings made at the Obregon Park site, which has
196 station sequence number SSN 337 in the NGA-West2 site database (Seyhan et al., 2014).

8
197 Residuals were computed using the Boore et al. (2014) GMM (BSSA14 hereafter). This site has
198 large positive residuals, indicating that the site response is stronger than anticipated from the
199 ergodic site term.

200

201 Figure 2. Within-event residuals and their mean for Obregon Park site. The large positive bias indicates
202 under-estimation of site response from the ergodic site term used in the calculation (in this case,
203 BSSA14). SSN indicates the station sequence number for the site in the NGA-West2 flatfile (Seyhan et
204 al. 2014).
205
206 We recommend that random effect ηS,j be taken as the mean of δWij (Eq. A11, denoted ηˆS , j ).

207 As explained in Appendix A, this follows a frequentist inference approach, which is

208 recommended because it provides an unbiased (though uncertain) estimate of ηS,j. Alternate
209 Bayesian methods (Eq. A4) provide a biased mean estimate. Our preference for the unbiased
210 mean is to provide the most accurate possible site response. Uncertainty in the mean estimate,
211 σ ηˆ is considered within the epistemic uncertainty characterization (explained further below).
S

212 We recommend using the observations as reflected through site term ηS,j to set the Flin=f1
213 term in the non-ergodic site amplification model (Eq. 2) (nonlinear parameters will typically be
214 set from simulations, next section). A first estimate of f1 can be taken as:

215 ( f1 ) j = Flinerg + η S , j (11)

216 where Flinerg is the ergodic (VS30- and depth-based) linear site amplification. Eq. (11) is effective if

217 the recordings from which ηS,j is derived are sufficiently weak that the nonlinear component of
218 the GMM site term is not exercised in the residuals calculation (Eq. 9) or if the nonlinear site-
219 specific amplification is well represented by the ergodic model. When the ground motions used
220 to evaluate η S , j are strong enough to produce nonlinearity and the site-specific and ergodic

221 nonlinear models are dissimilar, site residuals can be re-computed by taking site response as the
222 sum of Flinerg and site-specific Fnl. The resulting η S , j values are then used with Eq. (11) to set site-

223 specific f1.

9
224 The linear amplification for the Obregon Park site is shown in Figure 3 in arithmetic units
225 (expf1). The amplification is relative to the reference site condition in the ergodic model, in this
226 case 760 m/s.
227
228

229 Figure 3. Ergodic (BSSA14) and non-ergodic (site-specific) linear site amplification relative to 760 m/s
230 for Obregon Park site.
231

232 EVALUATION FROM SIMULATIONS

233 When GRAs are used to estimate site amplification, they should follow recommended
234 practices for selecting an appropriate computational framework, developing dynamic material
235 properties and their uncertainties, and identifying hazard-appropriate input motions. As described
236 in the Introduction, such recommendations are provided in NCHRP (2012) and Sea14; we
237 assume GRA results compatible with these recommendations in the discussion that follows.

238 GRA produces a series of discrete results, consisting of period-dependent amplification Y

239 given a particular set of dynamic soil properties and a particular input motion with amplitude
240 xIMref. In most cases, Eq. (2) is used to predict the RotD50-component of ground motion Z, in
241 which case, per Baker and Cornell (2006) xIMref should be similarly computed for a pair of input
242 motions using the RotD50 component, which is well estimated by the geometric mean (Shahi
243 and Baker 2014). Hence, even when only one scaled component of an input record is used in
244 ground response simulations, a combination of the two original components (as RotD50 or
245 geometric mean) should be used to represent its amplitude as xIMref.

246 When multiple input motions are used, potentially also with multiple realizations of uncertain
247 soil properties, a distribution of xIMref-Y results is obtained as shown in Figure 4. This distribution
248 can typically be fit reasonably well using the mean amplification function in Eq. (2). Routines
249 for performing these regression fits are provided in electronic supplements to Sea14. When the
250 reference site ground motions are derived for a single hazard level, usually that prescribed for
251 design purposes (e.g., 475 year return period), values of amplification Y are computed for a
252 relatively narrow range of xIMref. In such cases, it is not possible to regress each of the
253 parameters f1, f2, and f3; Sea14 (Section 2.6) suggest several options for addressing this situation
254 and the aforementioned routines can accommodate this case.

10
255

256 Figure 4. Intensity measure amplification levels from individual GRAs (symbols) following reference
257 site adjustment, fit curves from Eq. (2), and regression coefficients. The xIMref values used in the figures
258 are the RotD50-component. Amplification levels reflect record-to-record variability only (mean soil
259 properties used in the GRA). Empirical site amplification (derived from recordings) is shown as a mean
260 amplification with 95% confidence intervals beyond the limits of the abscissa, to indicate that this
261 amplification is not associated with any specific value of xIMref.
262
263 The example GRA results in Figure 4 are for amplification of PGA at one stiff and two soft
264 soil sites (Obregon Park, El Centro #7, and APEEL #2; site attributes are presented in
265 Appendix B, and details on the example results are given subsequently). These results are from
266 nonlinear GRA performed using Deepsoil (Hashash et al., 2016) following guidelines for
267 parameter selection in Sea14. The trend of site amplification with the input motion PGA
268 demonstrates weak nonlinearity for Obregon Park and strong nonlinearity for APEEL #2. These
269 are typical patterns that reflect the larger strains that develop in soils at soft sites.

270 The site amplification directly provided by the GRA result reflects the surface IM relative to
271 the IM for the site condition at the base of the profile, which is 540 m/s for Obregon Park,
272 508 m/s for El Centro #7, and 1525 m/s for APEEL #2. These conditions are different than the
273 GMM reference condition for X, which is typically 760 to 1100 m/s, hence the GRA-based
274 amplification requires modification to represent Y per Eq. (1). If we denote the VS30 value
275 corresponding to conditions at the base of the profile as VSB30 and the ground motion IM for that
276 condition (corresponding to the GRA input motions) is XB, then the GRA-based site
277 amplification is:

Z
278 YB = (12)
XB

279 The site amplification relative to the reference condition for X can then be evaluated as

280 ln (Y ) = ln (Y B ) + μ ln Y (VSB30 ) (13)

281 where μ ln Y (VSB30 ) is the mean site amplification from an ergodic model for the base-of-profile

282 site condition. Likewise the ground motion amplitude used in the nonlinear site term
283 computation is taken as:

11
284 ln ( x IMref ) = ln ( x IMref
B
) − μ ln Yref (VSB30 ) (14)

B
285 where xIMref is the corresponding value of that IM for the base-of-profile site condition and

286 μ ln Yref (V SB30 ) is the ergodic amplification of that reference IM for the site condition represented

287 by VSB30 . The site amplification and xIMref values shown in Figure 4 were adjusted in this manner.

288 These procedures carry elevated epistemic uncertainty when VSB30 is near or beyond the limits of

289 empirical models (i.e., > 1500-2000 m/s).

290 An important consideration in the interpretation of GRA results is that the computed mean
291 amplification may be biased relative to the true (generally unknown) mean site amplification.
292 Studies that investigate the effectiveness of GRA are most robust when the input motions are
293 known, which only applies for vertical arrays. While numerous studies of data from vertical
294 arrays at individual sites have found reasonably good fits to GRA results (e.g., Borja et al., 1999;
295 Elgamal et al., 2001; Lee et al., 2006; Tsai and Hashash, 2009; Yee et al., 2013;
296 Kaklamanos et al. 2014), other studies that systematically examine a broad set of such arrays
297 generally find misfits for a substantial subset (Thompson et al. 2012; Afshari and Stewart,
298 2015a). This bias occurs because GRA does not capture some of the physical mechanisms that
299 contribute to site response, especially for geologic conditions that cannot be reasonably
300 represented as 1D. The three example sites (Figure 4) illustrate the issue − amplification levels
301 observed from recordings (using procedures from previous section) are shown in the figure along
302 with the simulation fits. Underprediction bias for PGA amplification occurs for Obregon Park;
303 the bias is statistically significant in that the simulation mean is outside the 95% confidence
304 interval for the empirical amplification. The El Centro #7 and APEEL #2 weak motion
305 amplification levels from the data and simulations are relatively compatible.

306 For situations where amplification is evaluated from recordings and the resulting f1 differs
307 from the GRA-based estimate, we recommend adopting the empirical f1 value while maintaining
308 the nonlinear function from GRA (coefficients f2 and f3). This amounts to a vertical shift of the
309 simulation-based amplification function to the empirical value (e.g., to 3.2 in Figure 4a).

310 When ground motion recordings are not available for a site, simulations are the only viable
311 option for developing site-specific amplification functions. The suitability of simulation results
312 derived from 1D analysis requires judgement on the part of the analyst, and is subject to

12
313 significant epistemic uncertainty, as discussed next. However, a situation for which the
314 inadequacy of GRA is clear is at periods higher than the fundamental-mode period of the soil
315 column considered in the simulations (e.g., Baturay and Stewart, 2003), which is typically tens to
316 perhaps hundreds of meters in thickness. At these long periods, essentially no site response is
317 computed because the quarter wavelength is longer than the profile dimension; in the limit the
318 site displaces in an essentially rigid body manner. In contrast, long-period site response is
319 empirically known (from ergodic models) to be pronounced, which is due to macroscopic
320 features of the geologic column, and perhaps three-dimensional basin structure, that may be
321 several km in dimension. As discussed subsequently (Implementation section), we recommend
322 transitioning from simulation-based non-ergodic coefficients to ergodic coefficients at these long
323 periods.

324 EPISTEMIC UNCERTAINTY OF MEAN SITE RESPONSE

325 PSHA with a site-specific (non-ergodic) GMM requires consideration of epistemic

326 uncertainties in the mean site amplification model (this section) and aleatory uncertainty model
327 (next section), typically following a logic tree framework (Abrahamson and Bommer, 2005;
328 Bommer and Scherbaum, 2008). We suggest to characterize uncertainties in the mean model in
329 two ways: (1) uncertainty in f1 (Eq. 2), which reflects the overall level of amplification; and (2)
330 uncertainty in the nonlinear model (Fnl), potentially affecting parameters f2 and f3 (Eq. 2).

331 Uncertainty in the overall amplification level (parameter f1 in Eq. 2) should be characterized
332 differently when the model is derived partly from recordings versus being entirely GRA-based.
333 When f1 is set from recordings, epistemic uncertainty is represented by the standard error of the
334 mean of ηS,j (App A, Eq. A12). When f1 is set by GRA, epistemic uncertainties should reflect the
335 degree to which the physical processes modeled in 1D analysis capture the true site response
336 given the local geologic structure. For example, sites having a large impedance contrast might be
337 well characterized by GRA for oscillator periods below the site period. For such conditions,
338 epistemic uncertainty could be taken from the range of simulation results given alternate input
339 motions and variable VS profiles (details for considering such uncertainties are given in Chapter 2
340 of Sea14). If geologic complexities not captured by GRA could reasonably be expected to
341 influence site response (e.g., sites within large sedimentary basins with gradually increasing VS
342 with depth), epistemic uncertainty is driven more by the (admittedly subjective) impacts of those

13
343 complexities. The evaluation of such considerations is beyond the scope of the present paper, but
344 is the subject of ongoing research (preliminary results in Afshari and Stewart, 2016).

345 Characterization of epistemic uncertainty in the nonlinear model (Fnl) is directly related to
346 uncertainty in the soil properties that produce this response, which are the modulus reduction and
347 damping versus shear strain relations. Sea14 (Sections 2.2.2 and 2.3) summarize current models
348 for such uncertainties. The influence of these uncertainties on mean site amplification can be
349 readily considered in appropriate suites of GRA.

350 WITHIN-EVENT STANDARD DEVIATION

351 Our objective in this section is to develop an expression for the within-event standard
352 deviation of surface ground motion, φlnZ, appropriate for use with a site-specific natural log mean
353 in non-ergodic PSHA. Epistemic uncertainty in within-event standard deviations models are also
354 discussed. We note that φlnZ is not required when using convolution approaches, which instead
355 use the within-event standard deviation for reference rock, φlnX, as well as the standard deviation
356 of site amplification, φlnY.

357 EFFECT OF SOIL NONLINEARITY

358 Bazzurro and Cornell (2004a) showed that when the mean site amplification is described as,

359 μln Y = c1 + c2 ln ( x ) (15)

360 the standard deviation for Z can be computed as follows:

( c2 + 1) σ ln2 X + φln2 Y
2
361 σ ln Z ≈ (16)

362 where σlnX represents the total standard deviation of input motions (as used in GRA) and φlnY is
363 as defined previously (dispersion of site amplification). An important feature of Eq. (16) is that
364 values of c2 < 0 (caused by nonlinearity) reduce the ground motion variance. This feature of
365 ground motion has a physical explanation. Consider a distribution of reference ground motion X
366 of sufficient amplitude to produce nonlinearity. The mean realization (μlnX) produces a particular
367 level of amplification according to Eq. (2). Considering now the tails of the X distribution, a low
368 realization (negative epsilon) produces less nonlinearity and hence more amplification than at
369 μlnX, which has the effect of ‘squeezing’ the below-mean tail of the Z distribution. A high X

14
370 realization produces the opposite effect (more nonlinearity, less amplification), squeezing the
371 above-mean Z distribution tail. These reductions in the Z distribution width require reduction of
372 φlnZ, which is accommodated by the c2 term in Eq. (16); such effects of decreased within-event
373 variability for soft soil sites are also observed empirically (e.g., Boore et al., 2014).

374 The use of total standard deviation σlnX in Eq. (16) indicates that both the between- and
375 within-event components of variability are affected by site nonlinearity (per Eq. 7). Depending
376 on how between-event terms are computed, it may or may not be appropriate to reduce τ using
377 the slope term as in Eq. (16). Al Atik and Abrahamson (2010) take the position that between-
378 event variability τ represents the standard deviation of random effects terms ηE,i for rock
379 conditions only, which implies that nonlinear site response is not reflected in the data used in
380 their computation. The effects of nonlinearity are therefore added subsequent to the GMM
381 regression, and appear because of the impact of between-event variability on xIMref. Interestingly,
382 the effects of between-event variability on xIMref are not considered in forward application
383 (median values are used – discussed further in the Correlation Issues section below). Among
384 GMMs, this approach is used by Abrahamson et al. (2014) and Campbell and Bozorgnia (2014).
385 We take a different position that τ is based on ηE,i terms for as-recorded conditions including
386 rock and soil sites. In this case, the average site effect represented by the site term in the GMM
387 affects the event terms and hence their variability. Following this approach, between-event
388 dispersion for application is taken as that implied by the data and is not modified for site
389 nonlinearity, nor is it considered in the specification of xIMref. Among recent GMMs, this
390 approach is used by Boore et al. (2014) and Chiou and Youngs (2014). While we adopt the latter
391 approach, we acknowledge the former could also be applied in combination with appropriate
392 GMMs.

393 Because we take τ as a source attribute and site-independent, we modify Eq. (16) for
394 applicability to within-event variability:

( c2 + 1) φln2 X + φln2 Y
2
395 φln Z ≈ (17)

396 The approximation in Eqs. (16-17) is used because its derivation assumed Y and X as
397 uncorrelated, which is not strictly correct. These equations also apply for the case that the IMs

15
398 for X and xIMref match; we provide in Appendix C expressions for the case that the X and xIMref
399 IMs are different.

400 In the present work, we use a nonlinear function for mean amplification (Eq. 2), requiring
401 revision of Eq. (17) to account for nonlinear site effects. We replace c2 in Eq. (17) with the slope
402 of Eq. (2) in log-space,

d ( μ ln Y )
403 (18)
d ( ln xIMref )
404 The numerator in Eq. (18) can be written as:

f 2 dxIMref
405 d  f 2 ln ( xIMref + f 3 )  = (19)
xIMref + f3

406 The denominator is

dxIMref
407 d ( ln xIMref ) = (20)
xIMref

408 Combining, the slope from Eq. (18) becomes,

d ( μ ln Y ) f 2 xIMref
409 = (21)
d ( ln xIMref ) xIMref + f 3

410 Replacing c2 in Eq. (17) with the slope in Eq. (21), we obtain the following expression for
411 within-event standard deviation:

2
 f x 
412 φln Z ≈  2 IMref + 1 φln2 X + φln2 Y (22)
x 
 IMref + f3 

413 A similar version of Eq. (22) was originally derived by Goulet (2008), and subsequently given
414 by Papaspiliou et al. (2012). As before, Eq. (22) applies for matching X and xIMref IMs; App. C
415 presents equivalent expressions for non-matched IMs. Below we describe the evaluation of φlnX
416 and φlnY.

417 REMOVING EFFECTS OF SITE-TO-SITE VARIABILITY

16
418 The term φlnX in Eq. (22) represents the within-event variability of IMs for reference site
419 condition X, prior to modification for nonlinear site effects. Per Eq. (8), the principal contributor
420 to this standard deviation for non-ergodic applications is path-to-path variability (φP2P). Site-to-
421 site variability (φS2S) would also contribute for ergodic site response. Our challenge is that the
422 within-event standard deviation terms published with GMMs generally do not include this
423 partitioning (exceptions are Kotha et al. 2016 & Landwehr et al. 2016), hence only the total
424 within-event standard deviation is typically known (with contributions from φP2P, φS2S, and φlnY).
425 When these dispersions are taken from a GMM for the reference site condition X, they are
426 referred to as φlnXm, where ‘m’ in the subscript indicates its source is a model (GMM or a stand-
427 alone standard deviation model). Some contemporary GMMs are heteroskedastic, in the sense
428 they include site-dependent standard deviation terms (Abrahamson et al., 2014; Boore et al.,
429 2014; Campbell and Bozorgnia, 2014; Chiou and Youngs, 2014), making them well-suited to
430 estimating φlnXm.

431 For ergodic applications, the problem is trivial − the φlnX term in Eq. (22) is taken as φlnXm.
432 For non-ergodic applications, we propose two alternate procedures:

433 Approach 1: Subtract some fraction of variance φS22 S from published values of φln2 Xm , i.e.,

434 φln2 X = φln2 Xm − FφS22 S (23)

435 The term φS2S has been evaluated in prior work (Figure 5) using ground surface stations with
436 multiple recordings, most of which are of sufficiently low amplitude that the nonlinear effects
437 from the prior section are modest-to-negligible. It typically ranges between 0.3-0.5 and is
438 regionally variable (Rodriguez-Marek et al., 2013; Kaklamanos et al. 2013; Lin et al. 2011).
439 Results in Figure 5 by Rodriguez-Marek et al. (2014) are for downhole records and have lower
440 φS2S than the surface data, indicating much of the variability is from shallow portions of these
441 Kik-net sites. The term F in Eq. (23) ranges from zero to one. A value of F = 0 indicates no
442 confidence that the site amplification factors remove site-specific effects beyond the capability of
443 an ergodic model. A value of F = 1 is fully non-ergodic. We recommend use of F = 1 when f1 is
444 inferred from recordings; when the site amplification model is derived solely from simulations, F
445 should be selected in consideration of the degree to which GRA could be expected to capture the
446 most important site response physics (such considerations are discussed in the Epistemic

17
447 Uncertainty of Mean Site Response section). In developing Eq. (23), we considered whether the
448 subtraction of FφS22 S could alternatively be applied to the product of φln2 X and the nonlinear term
449 in brackets in Eq. (22); the present choice reflects the concept that the reference site ground
450 motions are partially non-ergodic, which are then further modified for nonlinear site effects.

451 Approach 2: Take φlnX from so-called ‘single station’ models (φSSm) developed for region-specific
452 applications, as available.

453 φln X = φSS ,m (24)

454 One such set of models is shown in Figure 6 along with an ergodic model (GeoPentech, 2015).

455 While site-specific within-event standard deviations are typically reduced relative to ergodic
456 models (as provided by Approaches 1 or 2), there is a possibility of especially high variability for
457 particular sites (e.g., Bradley, 2015). We are currently unable to predict such conditions in the
458 absence of recordings.

459

460 Figure 5. Site-to-site standard deviation (φS2S) findings from prior studies. R-Mea13, 14 = Rodriguez-
461 Marek et al. (2013, 2014); Kea13 = Kaklamanos et al. (2013); Lea11 = Lin et al. (2011). All results based
462 on ground surface recordings except R-Mea14, which is based on downhole recordings.
463

464 Figure 6. Single-station standard deviation models (φSSm) developed for South-Western US (SWUS)
465 project (GeoPentech, 2015). Three alternate models are shown, one derived from global data with rupture
466 distances < 50 km, and the other two derived from California data using different magnitude cutoffs (CA-
467 1, CA-2). BSSA14 ergodic model shown for comparison (plotted results applicable for M > 5.5, Rjb < 79
468 km, and VS30 > 300 m/s).
469

470 SITE RESPONSE VARIABILITY, φlnY

471 Prior studies have investigated site response variability based on sensitivity analyses using
472 GRA and based on analysis of ground motion data. The findings of this work have been
473 synthesized by Afshari and Stewart (2015b) and provide a basis for estimation of ϕlnY.

474 Suites of GRAs can evaluate the effects of random realizations of input motions, randomness
475 in VS profiles, randomness in modulus reduction and damping (MRD) curves, and model-to-
476 model variability (through the use of alternate codes). Sources of variability that are not captured

18
477 comprise epistemic uncertainties associated with limitations of GRA with respect to geologic
478 structure and 3D wave propagation effects. Figure 7 presents a compilation of period-dependent
479 site-amplification dispersion values (φlnY) from Kwok et al. 2008, Rathje et al. 2010, Bazzurro
480 and Cornell 2004b, and Li and Assimaki 2011. These studies all considered VS profile and MRD
481 uncertainties, but were inconsistent in their consideration of other sources. Nonetheless, the
482 results exhibit broadly similar features, namely: (1) the level of variability at short periods is
483 quite high at about 0.5-0.6; (2) there is an increased variability near the inelastic period of the
484 soil column considered in the analysis (e.g., about 0.15 sec at the Turkey Flat site considered by
485 Kwok et al. 2008, 1.2 sec at the La Cienega site considered by Li and Assimaki 2011); and (3)
486 beyond the soil column period, the dispersion drops markedly.
487
488

489 Figure 7. Synthesis of standard deviations of site amplification from GRA-based studies (Kea08=Kwok
490 et al. 2008; Rea10=Rathje et al. 2010; BC04=Bazzurro and Cornell 2004b; LA11=Li and Assimaki 2011)
491 and empirical studies based on recordings (R-Mea13, Kea13, Lea11).
492
493 For sites having ground motion recordings from multiple earthquakes, φlnY has been
494 estimated in two general ways. One approach is to partition GMM residuals as in Eq. (5), which
495 requires only ground surface records (Lin et al. 2011). In this case, φlnY is affected by randomness
496 in path and site response. The second approach uses vertical array data in which site
497 amplification (and its variability) is evaluated empirically using surface and downhole recordings
498 (Rodriguez-Marek et al. 2011, Kaklamanos et al. 2013); in such studies, φlnY reflects site
499 response variability only. Afshari and Stewart (2015b) explain how we have inferred values of
500 φlnY from this prior work.

501 Figure 7 envelopes results from these studies for φlnY, which are remarkably consistently
502 within a band of width ∼ 0.1 centered over the period range of interest at ∼0.26−0.32, despite
503 significant differences in the source data types and regions. Importantly, at periods below the site
504 period, this variability is less than suggested by GRA, whereas it is greater at long periods. Our
505 interpretation is that VS profile variability may be overestimated in the Toro (1995) model used
506 in the prior GRA-based studies and that GRA cannot capture site response variability beyond the
507 site period. Based on currently available information, we suggest that φlnY can reasonably be

19
508 estimated as falling in the range 0.26-0.32. Moreover, the relative consistency of φlnY as
509 evaluated from surface recordings (affected by site and path variability) and from
510 surface/downhole recordings (site variability only) suggest that φlnY is dominated by site
511 amplification variability, as noted previously.

512 EPISTEMIC UNCERTAINTY

513 Epistemic uncertainty of the within-event standard deviation model should be considered in
514 PSHA along with uncertainties in the mean model (previous section). Considerations affecting
515 this epistemic uncertainty are:

516 • Uncertainty in nonlinear parameters f2 and f3 from the mean model, which impacts φlnZ
517 per Eq. (22).
518 • Uncertainty in the non-ergodic, reference-site within-event standard deviation per Eq.
519 (23) or (24). If Approach 1 is used (Eq. 23), epistemic uncertainty should include
520 alternate values of parameter F (subjective) and φS 2S (regionally variable, Figure 5).

521 Epistemic uncertainty with Approach 2 involves consideration of alternate values of φSS ,m

522 .

523 IMPLEMENTATION

524 The preceding sections present procedures for evaluating site-specific natural log mean site
525 amplification and within-event standard deviation. When combined with a reference site GMM,
526 the result is a site-specific GMM suitable for non-ergodic PSHA. This approach has been
527 implemented in the open-source hazard code OpenSHA (Field et al. 2003). Here we describe
528 several details required for this implementation including correlation issues and an algorithm for
529 interpolation of coefficients between periods. We then describe input fields required for the
530 OpenSHA routines and currently enabled outputs.

531 CORRELATION ISSUES

532 PSHA provides one hazard curve for each IM at the surface (Z). The multi-variate hazard
533 integral includes an integration across possible values of this ground motion conditional on M,
534 site-source distance, and potentially other controlling variables. The IM for Z is characterized by

20
535 its mean, μlnZ, and total standard deviation, σlnZ, and an individual realization (z) within the
536 hazard integral can be written as:

537 ln z = μln Z + ε ln Zσ ln Z (25)

538 where εlnZ is the standard normal variate used in the integration for Z. Expanding upon Eq. (1),
539 the mean μlnZ depends on X as follows:

540 μln Z = μln X + μln Y | xIMref (26)

541 where μlnX is the natural log mean returned by a GMM for the reference site condition and
542 μln Y | xIMref is the mean site response model from Eq. (2). The quantity xIMref has its own
543 distribution, and a particular realization within that distribution can be written as:

544 ln xIMref = μln IMref + ε ln IMref σ ln IMref (27)

545 where μlnIMref is the natural log mean and σlnIMref is the standard deviation of xIMref .

546 Correlation issues arise because the IM controlling nonlinear site response ( xIMref ) differs

547 from the IM being predicted (Z) in two possible respects, even when both are for the RotD50
548 component of ground motion: (1) they apply for different site conditions and (2) they may be
549 different IMs (e.g., PGA for xIMref and pseudo-spectral acceleration, PSA, for Z). These

550 correlations affect the relationship between εlnZ and ε ln IMref ; since εlnZ is the hazard integrand, we

551 seek to estimate ε ln IMref given εlnZ.

552 When the IM for Z and xIMref match, the conditional mean of ε ln IMref (given εlnZ) can be

553 computed using the Z-X correlation coefficient ρln Z ,ln X :

554 ε ln IMref = ε ln Z ρln Z ,ln X (28)

555 A conditional standard deviation would in principle apply as well, but we do not consider this
556 source of dispersion, which would require another loop in the hazard integral. For the present
557 implementation we take ε ln IMref as its conditional mean and subsequently drop the overbar

558 notation. When the Z and xIMref IMs do not match, a more general relationship is applied,

21
559 ε ln IMref = ε ln Z ρln Z ,ln IMref (29)

560 where correlation coefficient ρln Z ,ln IMref reflects the effects of Z-X correlation and correlation

561 between intensity measures ρln IM1 ,ln IM 2 (previously investigated by Baker and Jayaram, 2008 and

562 Bradley, 2011).

563 Neither ρln Z ,ln X nor ρln Z ,ln IMref is presently known. We include this term in the formulation to

564 offer flexibility to users and because its effects are significant (next section). Correlations ρln Z ,ln X

565 and ρln Z ,ln IMref are generally assumed as null in practice, which has the effect of taking xIMref as its

566 mean value ( ε ln IMref =0). In future work, it should be possible to develop models for these

567 correlations from array data.

568

22
569 COEFFICIENT INTERPOLATION BETWEEN PERIODS

570 In practice, it may be cumbersome to perform the fitting operations for parameters f1, f2, and
571 f3 (as in Figure 4) for all of the periods used to construct a response spectrum. In lieu of this, it is
572 possible for the analyst to compute these parameters for a selected number of periods over the
573 range where GRA results are considered valid (usually this would be periods below the soil
574 column period, elongated for effects of nonlinearity). With these established, values at
575 intermediate periods can be interpolated in a way that captures features of the data while
576 mimicking known trends from ergodic models.

577 Equations implementing such an interpolation scheme have been developed using the Boore
578 et al. (2014) ergodic model, and are presented in Section 2.6.3 of Sea14. For brevity, the details
579 are not presented here, but example results of such an interpolation are given in Figure 8. As
580 shown in the figure, the routines provide the option of transitioning coefficients to ergodic values
581 beyond a user-specified period (denoted Tsite in Figure 8; this is general taken as the elongated
582 site period). The transition occurs over a range Tsite:NTsite, where N > 1.

583

584 Figure 8. An example of interpolation for calculating f1 and f2 values between GRA-based values of f1
585 and f2 at 10 periods for example of El Centro #7 site. For f1, the ergodic model is taken as the sum of VS30-
586 based term and basin depth term from Boore et al. (2014) (BSSA14). Basin depth taken as z1.0 = 1.54 km
587 (depth to 1.0 km/s shear wave isosurface). GRA results adjusted for compatibility with reference
588 condition in BSSA14 (760 m/s) (Eq. 13).
589
590 OPENSHA INPUT FIELDS AND OUTPUT OPTIONS

591 The OpenSHA routines for developing a site-specific GMM as described above are given in
592 the Non-ergodic site response GMPE option as an intensity measure relation (along with a series
593 of ergodic models, i.e., GMMs). Users enter the following information:

594 • GMM for the reference site condition (i.e., for analysis of ground motions denoted X).
595 The reference site condition is defined as that for which the site factor in the GMM is null
596 (in natural log units).
597 • Site parameter corresponding to conditions at base of profile ( VSB30 ) – needed when GRA
598 are used to estimate site effects. This would typically be a relatively firm soil or rock

23
599 condition corresponding to velocities at some depth in the profile. This condition is often
600 different from the reference site condition in the GMM.
601 • The VS30 and depth parameter appropriate for surface conditions (used in the interpolation
602 algorithm).
603 • Coefficients for the mean site amplification model (f1, f2, f3) and standard deviation model
604 (φlnY, and either φSS or φS 2S and F).

605 • Option for upper bound limit to site amplification, Ymax (default is expf1).
606 • Definition of the IM used for xIMref (PGA or PSA at oscillator period of interest).
607 • Correlation coefficient ρln Z ,ln IMref (default is zero).

608 • Option to adjust coefficients to ergodic model at long periods.

609 OpenSHA outputs using this intensity measure relation are the same as with other GMMs –
610 deterministic medians and standard deviations for specified conditions, hazard curves, uniform
611 hazard spectra, and disaggregation.

612 EXAMPLE HAZARD RESULTS

613 Figures B1-B3 presents site location maps, geologic logs, and shear wave velocity (VS)
614 profiles for three sites in California – Obregon Park (OP), El Centro #7 (EC7), and APEEL #2
615 (A2). Geotechnical conditions at the three selected sites are given in Table 1. We perform PSHA
616 for these sites to (1) illustrate the differences between ergodic and non-ergodic site response as
617 implemented in PSHA; (2) demonstrate differences between non-ergodic amplification functions
618 and associated PSHA results derived solely from GRA vs. those derived from the semi-empirical
619 approach (weak motion amplification from recordings combined with nonlinearity from GRA);
620 (3) illustrate the significance of X-Z correlation (described in preceding section) on hazard
621 results; and (4) demonstrate differences between site-specific PSHA and more approximate
622 convolution approaches. For brevity, we do not show the different results obtained from site-
623 specific PSHA and hybrid approaches, which has been shown previously (e.g., Cramer, 2003;
624 Bazzurro and Cornell, 2004a; Goulet and Stewart, 2009). GRA are performed using Deepsoil
625 version 6.1 (Hashash et al., 2016) following the protocols for ground motion selection and
626 parameter selection in Sea14 with one exception − small-strain soil damping Dmin is taken from
627 geotechnical models (Darendeli, 2001; Menq, 2003) without modification.

24
628 Table 1. Geotechnical conditions at example sites selected for analysis

VSB30 VS30 z1.0 Tsite

Site Description
(m/s) (m/s) (km) (sec)
stiff soil overlying weathered rock, lacking
Obregon Park
appreciable impedance contrast over the depth of 540 449 0.56 0.57
(OP)
exploration
very deep soil with soft near-surface sediments,
El Centro #7
without a pronounced impedance contrast over 508 211 1.54 1.5
(EC7)
depth of exploration
soft clay (Bay Mud) overlying stiffer sediments and
APEEL #2
rock, and having large impedance contrasts at the 1525 133 0.09 1.6
(A2)
base of the soft clay and at the soil-rock interface
629

630 Figure 9(a) shows the variation of linear PSA amplification with oscillator period for the
631 three sites. Results are shown from the interpretation of recordings (Eq. 11), GRA, and an
632 ergodic model (BSSA14). The empirical results are shown as a median and 95% confidence
633 interval, which reflect the uncertainty of site term η S , j associated with the finite number of

634 recordings. The GRA results are fit using Eq. (2) for each IM as shown in Figure 4; the result
635 plotted here is exp(f1) (small-strain amplification). The ergodic results reflect VS30-scaling and
636 basin effects from BSSA14. For the OP site, the GRA underprediction bias for PGA originally
637 shown in Figure 4 is seen to persist across the full period range. The ergodic model also
638 underpredicts, but is less biased for T > 1 sec than GRA. EC7 has relatively modest
639 underprediction bias for T < ∼0.3 sec, but substantial bias at longer periods including near an
640 apparent site period at 4 sec (possible fundamental mode for the deep sediment column in this
641 part of the Imperial Valley). The match of GRA to observations is quite good for A2, including
642 amplification near the site period of 1 sec. The ergodic model improves upon GRA predictions at
643 long periods for EC7 but not for A2. Figure 9(b) shows results for a nonlinear case in which
644 xIMref = 5f3; as such, the difference from the linear result reflects the size of the f2 term. Nonlinear
645 effects are small for OP, but pronounced for EC7 and A2 for periods T < ∼ 2 sec. Nonlinearity
646 increases the site period for A2.
647
648

25
649 Figure 9. (a) Linear amplification vs. period as inferred from recordings and as estimated from GRA and
650 ergodic model. (b) Amplification for nonlinear condition in which xIMref/f3 = 5 for ergodic and GRA-based
651 models.
652 Figure 10 shows uniform hazard spectra (UHS) for 2475 year return period computed by
653 OpenSHA with alternate site models: non-ergodic site response based on GRA alone, semi-
654 empirical non-ergodic (f1 set empirically, Fnl from GRE results), the ergodic site model, and
655 ergodic reference site. The analyses using GRA for site response are applied across the full
656 period range (no transition to empirical model at long period, as described in previous section),
657 so as to highlight differences from other modeling approaches. The reference site GMM is taken
658 as BSSA14 and the source model is UCERF3 FM3.1 (Field et al. 2014). The hazard results
659 presented in Figure 10 and subsequently are based on the site-specific natural log mean
660 amplification and standard deviation models. As noted previously, epistemic uncertainties
661 associated with simulation-based site response are the subject of ongoing work, and as such a
662 formal assessment of these effects is deferred to a later paper.

663

664 Figure 10. Uniform hazard spectra for example sites using ergodic and non-ergodic site amplification
665 models. Reference site hazard computed using reduced within-event standard deviation (φS2S removed),
666 in conformance with standard-of-practice for non-ergodic hazard analysis with convolution approach.
667 Non-ergodic models use F=1, and φS2S =0.4. No transition to ergodic model at long periods applied. The
668 jagged appearance of non-ergodic spectra result from a limited number of GMM periods implemented in
669 OpenSHA.
670
671 For the OP site (Figure 10a), the effect of the underprediction bias from the ergodic model
672 and GRA is clear, as is the more favorable performance of the ergodic model at long periods. For
673 EC7 (Figure 10b), the GRA under-predicts long-period (> ∼ 0.7 sec) site amplification associated
674 with deep basin structure. Those long period site effects are reasonably well captured by the
675 ergodic model; results of this sort motivate the use of the ergodic model transition (Figure 8).
676 The non-ergodic UHS for A2 (Figure 10c) exhibits a peak in the spectrum near the site period (∼
677 1 sec) that is captured by GRA − the ability to capture features of this sort is an advantage of
678 non-ergodic site response procedures. The large impedance contrast at A2 causes ground
679 response to dominate the site response, which is demonstrated by the similarity of the GRA and
680 semi-empirical results in Figures 9 and 10.

26
681 Figure 11 shows PGA and 1.5 sec PSA hazard curves for the EC7 site derived using the
682 reference site condition, the semi-empirical non-ergodic site model, and the ergodic site model.
683 These two IMs were selected to illustrate results for cases with nonlinearity varying from strong
684 (PGA) to weak but non-zero (1.5 sec PSA). The non-ergodic site model is applied using both the
685 convolution approach and the recommended site-specific GMM implementation. For the latter,
686 we show results computed with with ρln Z ,ln X set to 0 (default), 0.5, and 1.0. Per typical practice

687 on recent projects (as described in Introduction), the convolution approach is applied with
688 reduced within-event standard deviation, per Eq. (23). Two key findings from these hazard
689 curves are:

690 • As ρln Z ,ln X increases, the nonlinear component of the site term becomes more

691 pronounced, which reduces the hazard (PGA, Figure 11a). Where the site nonlinearity is
692 not significant (1.5 sec PSA, Figure 11b), the effect of ρln Z ,ln X on hazard disappears.

693 • For IMs subject to strongly nonlinear site response (Figure 11a), the use of convolution
694 produces lower hazard at long return periods than site-specific PSHA when applied with
695 the default zero correlation. This occurs because the amplitude of shaking driving the
696 nonlinear site term (xIMref) in convolution is taken from the reference (X) hazard, which is
697 ‘positive epsilon’ (exceeds the mean). When nonlinearity effects are modest (Figure 11b),
698 the differences between convolution and site-specific PSHA results are small.

699
700 Figure 11. Hazard curves for EC7 site derived using non-ergodic and ergodic site amplification models.
701 The non-ergodic (semi-empirical) model is applied in hazard computations with varying levels of
702 reference-to-site correlation. The model is also applied through convolution with the non-ergodic
703 reference rock hazard.
704

705 CONCLUSIONS AND RECOMMENDATIONS

706 This paper, and the accompanying 2016 Joyner Lectures (Stewart, 2016), seek to explain the
707 meaning of non-ergodic site response and the manner by which it can be applied to develop site-
708 specific GMMs for use in hazard analyses. The concept of using site-specific GMMs in PSHA
709 has been proposed previously (McGuire et al., 2001), but a specific methodology has not been
710 described in the open literature and such approaches have seldom been applied in practice. We

27
711 describe a methodology and demonstrate its use (in a module within the OpenSHA hazard code),
712 from which several benefits are evident relative to the use of ergodic models:

713 • Non-ergodic mean site response is estimated using truly site-specific information that
714 improves the accuracy and reliability of hazard estimates and related outcomes (e.g,
715 disaggregation results) relative to ergodic models.
716 • The standard deviation of ground motion is reduced, which tends to lower ground
717 motion estimates for hazard levels of interest in design (return periods of 500 years or
718 more).

719 We show how non-ergodic (i.e., site-specific) site response can be evaluated using on-site
720 ground motion recordings in combination with GRA simulations, or by using GRA simulations
721 alone. The former is preferred when practicable, because GRA results may be biased,
722 particularly at long periods (i.e., beyond the fundamental period of the 1D model). These errors
723 arise from physical processes that affect site response (e.g., surface waves, basin effects) but that
724 cannot be readily incorporated into typical GRA simulations performed for engineering
725 purposes.

726 The methodology presented in this paper can be distilled to the following steps:

727 1. Perform detailed site characterization, including geotechnical borehole logging and
728 geophysical logging (details in Sections 2.2-2.3 of Sea14).
729 2. (Optional, but recommended): Install ground motion instrumentation at the site of interest
730 (accelerometers or seismometers) and record data over time. When such data is
731 unavailable, it may be possible to use recordings from instruments in the vicinity of the
732 site of interest. Future research will investigate the viability of this approach, which will
733 depend on the spatial correlations of random effect terms ηS.
734 3. Analyze ground response using suitable equivalent-linear or nonlinear procedures
735 (NCHRP, 2012; Sections 2.4 of Sea14). Compute amplification values using Eq. (12)
736 (YB) for each simulation outcome (combination of dynamic soil properties and input
737 motion).
738 4. Adjust computed amplification values as needed to correct for incompatibility between
739 the reference condition in the GMM and the base of profile condition ( VSB30 ) using Eqs.
740 (13-14). Adjusted amplification values are denoted as Y.

28
741 5. Select the IM for the reference site condition to use for xIMref. Fit Eq. (2) to adjusted
742 amplification values (Y) over the range of xIMref considered in the simulations.
743 6. (Optional): Compute empirical site response from ground motions recorded at the site
744 using Eq. (11); adjust coefficient f1 from Step 5 as needed to match these results.
745 7. Evaluate the extent to which the computed site response is non-ergodic, which is often
746 subjective. If on-site ground motions are used, we posit that the site response is non-
747 ergodic, which implies that the within-event standard deviation should be taken using Eq.
748 (22) with φlnX taken using either Eq. (23) (with F = 1) or Eq. (24) (alternate expressions
749 for unmatched IMs for X and xIMref given in Appendix C). If recordings are not used, φlnX
750 is best taken using Eqs (23) with F set by judgement (when site response is judged to be
751 dominated by ground response, F can be nearly unity; when this is not the case, F should
752 be less than one).
753 8. Consider Z-X and between-IM correlation effects as explained in the Correlation Issues
754 section. These correlations affect the implementation of the mean amplification function
755 per Eqs (27-29).
756 9. The mean site amplification function (Eq. 2) and within-event standard deviation terms
757 (Eq. 22) are combined with regionally appropriate ground motion models to define site-
758 specific GMMs for use in hazard analyses. It is common for results of such calculations
759 in active regions like California to lower ground motion estimates at long return periods
760 (e.g., 475 or 2475 years) by 20% or more, due in large part to standard deviation
761 reductions.
762 10. Repeat hazard analyses to consider epistemic uncertainties of the mean amplification
763 function, alternate representations of standard deviation (e.g., due to variations in F), and
764 alternate correlation models (especially for Z-X correlation). It is not appropriate to take
765 the standard deviation reduction associated with the use of a non-ergodic model without
766 considering these site response uncertainty effects.

767 Clearly the level of effort involved in developing a non-ergodic site amplification model
768 (with corresponding site-specific GMMs) is greater than the use of relatively simple ergodic
769 models. While conceding that point, our view is that the profession should continue to move in
770 this direction, at least for critical projects. The increased effort provides additional knowledge,
771 which lowers aleatory variability and frequently reduces hazard.

29
772 Moreover, the proposed approach using a site-specific GMM in the hazard integral offers
773 advantages relative to convolution methods that comprise the current state of practice for non-
774 ergodic PSHA. These advantages include consideration of nonlinear effects on within-event
775 standard deviation terms, consideration of rock-soil and between-IM correlation issues as
776 described above, and the identification of controlling sources (through disaggregation) in a
777 manner than considers the site response. We show that the use of convolution methods tends to
778 underestimate hazard when the site response is nonlinear.

779 As described in this paper, two technical challenges will affect projects utilizing the proposed
780 approach: (1) difficulty in identifying a priori when GRA results provide an unbiased estimate of
781 site effects and (2) lack of knowledge regarding surface-reference (Z-X) correlations. Future
782 progress on these subjects will affect specific steps in the proposed procedure, but not the
783 framework itself.

784 ACKNOWLEDGEMENTS

785 The first author would like to thank the Earthquake Engineering Research Institute and
786 Seismological Society of America for the opportunity to present the 2016 Joyner Lecture on this
787 topic. The work presented here was supported by the PEER center, U.S. Geological Survey,
788 California Strong Motion Instrumentation Program, and National Science Foundation. The views
789 and conclusions presented in this paper are those of the authors, and no endorsement is implied
790 on the part of the State of California or the U.S. Federal government. We thank Kevin Milner of
791 the Southern California Earthquake Center for his assistance with OpenSHA implementation.
792 Paolo Zimmaro of UCLA provided technical support at several stages of this work. We thank
793 Norman Abrahamson, Jack Baker, Youssef Hashash, and Ivan Wong for helpful discussions on
794 this topic, and Brendon Bradley, Fabrice Cotton, and Adrian Rodriguez-Marek for their critical
795 evaluation of a previous draft of this paper.

796 REFERENCES

797 Abrahamson, N. A. and Bommer, J. J., 2005. Probability and uncertainty in seismic hazard analysis,
798 Earthquake Spectra, 21, 603–607.

799 Abrahamson, N. A., Silva, W. J., and Kamai, R., 2014. Summary of the ASK14 ground motion relation
800 for active crustal regions, Earthquake Spectra, 30, 1025–1055.

30
801 Afshari, K. and Stewart, J. P., 2015a. Effectiveness of 1D ground response analyses at predicting site
802 response at California vertical array sites, Proc. SMIP2015 Seminar on Utilization of Strong Motion
803 Data, California Strong Motion Instrumentation Program, Sacramento, CA.

804 Afshari, K. and Stewart, J. P., 2015b. Uncertainty of site amplification derived from ground response
805 analysis, Proc. 6th Int. Conf. Earthquake Geotech. Eng., Christchurch, New Zealand, Paper No. 227.

806 Al Atik, L. and Abrahamson, N. A., 2010. Nonlinear site response effects on the standard deviations of
807 predicted ground motions, Bull. Seismol. Soc. Am., 100, 1288–1292.

808 Al Atik, L., Abrahamson, N. A., Bommer, J. J., Scherbaum, F., Cotton, F., and Kuehn, N., 2010. The
809 variability of ground motion prediction models and its components, Seism. Res. Lett., 81 794–801.

810 Anderson, J. G. and Brune, J. N., 1999. Probability seismic hazard analysis without the ergodic
811 assumption, Seismol. Res. Let., 70, 19–28.

812 Atkinson, G. M., 2006. Single-station sigma, Bull. Seism. Soc. Am., 96, 446–455.

813 Baker, J. W. and Cornell, C. A., 2006. Which spectral acceleration are you using? Earthquake Spectra,
814 22, 293–312.

815 Baker, J. W. and Jayaram, N., 2008. Correlation of spectral acceleration values from NGA ground motion
816 models, Earthquake Spectra, 24, 299–317.

817 Baturay, M. B. and Stewart, J. P., 2003. Uncertainty and bias in ground motion estimates from ground
818 response analyses, Bull. Seism. Soc. Am., 93, 2025-2042.

819 Bazzurro, P. and Cornell, C. A., 2004a. Nonlinear soil-site effects in probabilistic seismic-hazard
820 analysis, Bull. Seismol. Soc. Am., 94, 2110–2123.

821 Bazzurro, P. and Cornell, C. A., 2004b. Ground motion amplification in nonlinear soil sites with
822 uncertain properties, Bull. Seismol. Soc. Am., 94, 2090–2109.

823 Boore, D. M. and Bommer, J. J., 2005. Processing of strong-motion accelerograms: needs, options and
824 consequences, Soil Dyn. Eqk. Eng., 25, 93–115.

825 Boore, D. M., 2010. Orientation-independent, nongeometric-mean measures of seismic intensity from two
826 horizontal components of motion, Bull. Seismol. Soc. Am., 100, 1830–1835.

827 Boore, D. M., Stewart, J. P., Seyhan, E., and Atkinson, G. M, 2014. NGA-West 2 equations for predicting
828 PGA, PGV, and 5%-damped PSA for shallow crustal earthquakes, Earthquake Spectra, 30, 1057–
829 1085.

31
830 Bommer, J. J. and Scherbaum, F., 2008. The use and misuse of logic trees in probabilistic seismic hazard
831 analysis. Earthquake Spectra, 24, 997–1009.

832 Bommer, J. J., Coppersmith, K. J., Coppersmith, R. T., Hanson, K. L., Mangongolo, A., Neveling, J.,
833 Rathje, E.M, Rodriguez-Marek, A., Scherbaum, F., Shelembe, R., Stafford, P. J., and Strasser, F. O.,
834 2015. A SSHAC Level 3 probabilistic seismic hazard analysis for a new-build nuclear site in South
835 Africa. Earthquake Spectra, 31, 661–698.

836 Borja, R. I., Chao, H.-Y., Montans, F. J., and Lin, C.-H., 1999. Nonlinear ground response at Lotung
837 LSST site, J. Geotech. Geoenviron. Eng., 125, 187–197.

838 Bozorgnia, Y., Abrahamson, N. A., Al Atik, L., Ancheta, T. D., Atkinson, G. M., Baker, J. W., Baltay,
839 A., Boore, D. M., Campbell, K. W., Chiou, B. S.-J., Darragh, R., Day, S., Donahue, J., Graves, R. W.,
840 Gregor, N., Hanks, T., Idriss, I. M., Kamai, R., Kishida, T., Kottke, A., Mahin, S. A., Rezaeian, S.,
841 Rowshandel, B., Seyhan, E., Shahi, S., Shantz, T., Silva, W., Spudich, P., Stewart, J. P., Watson-
842 Lamprey, J., Wooddell, K., and Youngs, R., 2014. NGA-West2 research project, Earthquake Spectra,
843 30, 973–987.

844 Bradley, B. A., 2011. Correlation of significant duration with amplitude and cumulative intensity
845 measures and its use in ground motion selection, J. Earthquake Eng., 15, 809–832.

846 Bradley, B. A., 2015. Systematic ground motion observations in the Canterbury Earthquakes and region-
847 specific non-ergodic empirical ground motion modeling, Earthquake Spectra, 31, 1735-1761.

848 Campbell, K. W. and Bozorgnia, Y., 2014. NGA-West2 ground motion model for the average horizontal
849 components of PGA, PGV, and 5% damped linear acceleration response spectra, Earthquake Spectra,
850 30, 1087–1115.

851 Chiou, B. S.-J. and Youngs, R. R., 2014. Update of the Chiou and Youngs NGA model for the average
852 horizontal component of peak ground motion and response spectra, Earthquake Spectra, 30, 1117–
853 1153.

854 Coppersmith, K., Bommer, J., Hanson, K., Coppersmith, R., Unruh, J., Wolf, L, Youngs, B., Al Atik, L,
855 Rodriguez-Marek, A., Toro, G., 2014. Hanford Sitewide Probabilistic Seismic Hazard Analysis,
856 Prepared for the U.S. Department of Energy Under Contract DE-AC06076RL01830, and Energy
857 Northwest, Pacific Northwest National Lab Report PNNL-23361, November.

858 Cramer, C. H., 2003. Site-specific seismic-hazard analysis that is completely probabilistic, Bull. Seismol.
859 Soc. Am., 93, 1841–1846.

32
860 Darendeli, M. B., 2001. Development of a New Family of Normalized modulus reduction and material
861 damping curves, PhD Thesis, Department of Civil Engineering, University of Texas, Austin, TX.

862 Elgamal, A., Lai, T., Yang, Z., He, L., 2001. Dynamic soil properties, seismic downhole arrays and
863 applications in practice, Proceedings, 4th International Conference on Recent Advances in
864 Geotechnical Earthquake Engineering and Soil Dynamics, S. Prakash, ed., San Diego, CA.

865 Field, E. H., Jordan, T. H., and Cornell, C. A., 2003. OpenSHA: A Developing Community-Modeling
866 Environment for Seismic Hazard Analysis, Seismol. Res. Lett., 74, 406–419.

867 Field, E. H., Arrowsmith, R. J., Biasi, G. P., Bird, P., Dawson, T. E., Felzer, K. R., Jackson, D. D.,
868 Johnson, K. M., Jordan, T. H., Madden, C., Michael, A. J., Milner, K. R., Page, M. T., Parsons, T.,
869 Powers, P. M, Shaw, B. E., Thatcher, W. R., Weldon, R. J., and Zeng, Y., 2014. Uniform California
870 earthquake rupture forecast, version 3 (UCERF3): The time-independent model, Bull. Seismol. Soc.
871 Am. 104, 1122–1180.

872 Hashash, Y. M. A., Musgrove, M. I., Harmon, J. A., Groholski, D. R., Phillips, C. A., and Park, D., 2016.
873 DEEPSOIL 6.1, User Manual.

874 GeoPentech, 2015. Southwestern United States ground motion characterization SSHAC level 3 –
875 technical report rev. 2, March 2015.

876 Goulet, C. A., 2008. Improving the characterization of seismic hazard for performance-based earthquake
877 engineering design, PhD Thesis, Department of Civil Engineering, University of California,
878 Los Angeles.

879 Goulet, C. A. and Stewart, J. P., 2009. Pitfalls of deterministic application of nonlinear site factors in
880 probabilistic assessment of ground motions, Earthquake Spectra, 25, 541–555.

881 Kaklamanos, J., Bradley, B. A., Thompson, E. M., and Baise, L. G., 2013. Critical parameters affecting
882 bias and variability in site-response analyses using KiK-net downhole array data, Bull. Seismol. Soc.
883 Am., 103, 1733–1749.

884 Kaklamanos, J., Baise, L. G., Thompson, E. M., and Dorfmann, L., 2014. Comparison of 1D linear,
885 equivalent-linear, and nonlinear site response models at six KiK-net validation sites, Soil Dyn. Eqk.
886 Eng., 69, 435–460.

887 Kotha, S.R., D. Bindi, and F. Cotton, 2016. Partially non-ergodic region specific GMPE for Europe and
888 Middle-East, Bull. Earthq. Eng., 14, 1245-1263.

889 Kwok, A. O. L., Stewart, J. P., Hashash, Y. M. A, 2008. Nonlinear ground-response analysis of Turkey
890 Flat shallow stiff-soil site to strong ground motion, Bull. Seismol. Soc. Am., 98, 331–343.

33
891 Landwehr, N., Kuehn, N. M., Scheffer, T., and Abrahamson, N.A., 2016. A nonergodic ground‐motion
892 model for California with spatially varying coefficients. Bull. Seismol. Soc. Am., 106, 2574-2583.

893 Lee, C.-P., Tsai, Y.-B., and Wen, K. L., 2006. Analysis of nonlinear site response using the LSST
894 downhole accelerometer array data, Soil Dyn. Eqk. Eng., 26, 435–460.

895 Li, W. and Assimaki, D., 2011. Site and motion-dependent parametric uncertainty of site-response
896 analyses in earthquake simulations, Bull. Seismol. Soc. Am., 100, 954–968.

897 Lin P.-S., Chiou, B. S.-J, Abrahamson, N. A., Walling, M., Lee, C.-T., and Cheng, C.-T., 2011.
898 Repeatable source, site, and path effects on the standard deviation for ground-motion prediction, Bull.
899 Seismol. Soc. Am. 101, 2281–2295.

900 McGuire, R. K., Silva, W. J., and Costantino, C. J., 2001. Technical basis for revision of regulatory
901 guidance on design ground motions: Hazard-and risk-consistent ground motion spectra guidelines.
902 NUREG/CR‐6728, United States NRC.

903 Menq, F. Y., 2003. Dynamic Properties of Sandy and Gravelly Soils, PhD Thesis, Department of Civil
904 Engineering, University of Texas, Austin, TX.

905 Montgomery, D. C. and Runger, G. C., 2005. Applied statistics and probability for engineers, 6th edition,
906 John Wiley & Sons, Inc., Hoboken, New Jersey.

907 NCHRP, 2012. Practices and Procedures for Site-Specific Evaluations of Earthquake Ground Motions,
908 Synthesis 428 (N Matasovic and YMA Hashash), National Cooperative Highway Research Program,
909 Transportation Research Board, Washington D.C.

910 Papaspiliou, M., Kontoe, S., and Bommer, J. J., 2012. An exploration of incorporating site response into
911 PSHA-Part II: Sensitivity of hazard estimates to site response approaches. Soil Dyn & Earthquake
912 Eng, 42, 316–330.

913 Rathje, E. M., Kottke, A. R., and Trent, W. L., 2010. Influence of input motion and site property
914 variabilities on seismic site response analysis, J. Geotech. Geoenv. Eng., 136, 607–619.

915 Rathje, E. M., Pehlivan, M., Gilbert, R., and Rodriguez-Marek, A., 2015. Incorporating site response into
916 seismic hazard assessments for critical facilities: A probabilistic approach. In Perspectives on
917 Earthquake Geotechnical Engineering. Springer International Publishing, 93–111.

918 Renault, P., S. Heuberger, and N. A. Abrahamson, 2010. PEGASOS Refinement Project: An improved
919 PSHA for Swiss nuclear power plants, Proc. 14th European Conference of Earthq. Engin., Ohrid,
920 Republic of Macedonia, 30 August–3 September 2010, Paper ID 991.

34
921 Rodriguez-Marek, A., Montalva, G. A., Cotton, F., and Bonilla, F., 2011, Analysis of single-station
922 standard deviation using the KiK-net data, Bull. Seismol. Soc. Am. 101, 1242–1258.

923 Rodriguez-Marek, A., Cotton, F., Abrahamson, N. A., Akkar, S., Al Atik, L., Edwards, B., Montalva, G.
924 A., and Dawood, H. M., 2013. A model for single-station standard deviation using data from various
925 tectonic regions, Bull. Seismol. Soc. Am. 103, 3149–3163.

926 Rodriguez-Marek, A., Rathje, E. M., Bommer, J. J., Scherbaum, F., and Stafford, P. J., 2014. Application
927 of single-station sigma and site response characterization in a probabilistic seismic hazard analysis for
928 a new nuclear site, Bull. Seismol. Soc. Am., 104, 1601-1619.

929 Seyhan, E., Stewart, J. P., Ancheta, T. D., Darragh, R. B., and Graves, R. W., 2014. NGA-West 2 site
930 database, Earthquake Spectra, 30, 1007–1024.

931 Shahi, S. K. and Baker, J.W., 2014. NGA-West2 models for ground motion directionality, Earthquake
932 Spectra, 30, 1285–1300.

933 Stafford, P. J., 2012. Evaluation of structural performance in the immediate aftermath of an earthquake: A
934 case study of the 2011 Christchurch earthquake, Int. J. Forensic Engrg., 1, 58-77.

935 Stewart, J.P. 2016. Joyner Lecture: Site response uncertainty and its implications for seismic risk
936 characterization, Seism. Res. Ltrs., 87(2B), pp 516 (abstract).

937 Stewart, J.P., Afshari, K., and Hashash, Y. M. A., 2014. Guidelines for performing hazard-consistent one-
938 dimensional ground response analysis for ground motion prediction, PEER Report No. 2014/16,
939 Pacific Earthquake Engineering Research Center, UC Berkeley, CA.

940 Thompson, E. M., Baise, L. G., Tanaka, Y., and Kayen, R. E., 2012. A taxonomy of site response
941 complexity, Soil Dyn. Earthq. Eng., 41, 32–43.

942 Toro, G. R., 1995. Probabilistic models of site velocity profiles for generic and site-specific ground-
943 motion amplification studies, Technical Report No. 779574, Brookhaven National Laboratory, Upton,
944 NY.

945 Tsai, C. C. and Hashash, Y. M. A., 2009. Learning of dynamic soil behavior from downhole arrays, J.
946 Geotech. Geoenv. Eng., 135, 745–757.

947 Yee, E., Stewart, J. P., and Tokimatsu, K., 2013. Elastic and large-strain nonlinear seismic site response
948 from analysis of vertical array recordings, J. Geotech. Geoenv. Eng., 139, 1789–1801.

35