Proceedings of the 9th U.S.

National and 10th Canadian Conference on Earthquake Engineering

Compte Rendu de la 9ième Conférence Nationale Américaine et
10ième Conférence Canadienne de Génie Parasismique
July 25-29, 2010, Toronto, Ontario, Canada • Paper No 1896

SCENARIO SPECTRA FOR DESIGN GROUND MOTIONS AND RISK

CALCULATION

N. A. Abrahamson1 and L. Al Atik2

ABSTRACT

A common method for developing design spectra based on the

probabilistic approach is uniform hazard spectra (UHS). This paper reviews the
basics of UHS and conditional mean spectra and presents a new approach for
developing a suite of realistic scenario earthquake spectra that account for the
correlations and variability of the peaks and troughs in the spectra at different
spectral periods. These scenario spectra can be assigned rates that will allow
reconstituting the hazard curves at different periods. The conditional mean spectra
and the conditional spectra provide methods for breaking the UHS down into
realistic earthquakes. The choice of using a UHS rather than multiple spectra for
the different scenarios should be the decision of the engineering analyst, not the
hazard analyst. If it is worth the additional analysis costs to avoid exciting a broad
period range in a single evaluation, then the engineer should request multiple
scenario spectra from the hazard analyst. For risk calculation for structures that
have a response that depends on widely separated spectral frequencies, then the
conditional spectral approach can be used.

Introduction

A common method for developing design spectra based on the probabilistic approach is
uniform hazard spectra. A uniform hazard spectrum (UHS) is developed by first computing the
hazard at a suite of spectral periods using response spectral attenuation relations. That is, the
hazard is computed independently for each spectral period. For a selected return period, the
ground motion for each spectral period is measured from the hazard curves. These ground
motions are then plotted at their respective spectral periods to form the uniform hazard spectrum.
This process is shown graphically in Figure 1.

The term “uniform hazard spectrum” is used because there is an equal probability of exceeding
the ground motion at any period. Since the hazard is computed independently for each spectral
period, in general, a uniform hazard spectrum does not represent the spectrum of any single

1
Pacific Gas & Electricity Company, 245 Market Street, San Francisco, CA 94105
2
Postdoctoral Scholar, Dept. of Civil Engineering, University of California, Berkeley, CA 94720
1
earthquake. It is common to find that the high frequency (f>5 Hz) ground motions are controlled
by nearby moderate magnitude earthquakes, whereas, the long period (T>1 sec) ground motions
are controlled by distant large magnitude earthquakes.

The “mixing” of earthquakes in the UHS is often cited as a disadvantage of PSHA. There is
nothing in the PSHA method that requires using a UHS. Based on the disaggregation, multiple
spectra (for each important source) can be developed. The reason for using a UHS rather than
using multiple spectra for the individual scenarios is to reduce the number of engineering
analyses required. A deterministic analysis has the same issue. If one deterministic scenario
leads to the largest spectral values for long spectral periods and a different deterministic scenario
leads to the largest spectral values for short spectral periods, it is common practice to develop a
single design spectrum that envelopes the two deterministic spectra. In this case, the design
spectrum also does not represent a single earthquake.

The choice of using a UHS rather than multiple spectra for the different scenarios is the decision
of the engineering analyst, not the hazard analyst. If it is worth the additional analysis costs to
avoid exciting a broad period range in a single evaluation, then the engineer should request
multiple scenario spectra from the hazard analyst.

In this paper, we will review the basics of uniform hazard spectra and conditional mean
spectra and present a new approach for developing a suite of realistic scenario earthquake spectra
that account for the correlations and variability of the peaks and troughs in the spectra at different
spectral periods. These scenario spectra can be assigned rates that will allow reconstituting the
hazard curves at different periods.

Figure 1. Procedure for developing a uniform hazard spectrum with a return period of 500 years

2
 Uniform Hazard Spectra and Deaggregation of the Hazard
The hazard curve gives the combined effect of all magnitudes and distances on the
probability of exceeding a given ground motion level. To provide insight into what events control
the hazard at a given ground motion level and spectral period, the hazard curve is broken down into
its contributions from different earthquake scenarios. This process is called disaggregation
(Bazzurro and Cornell 1999) whereby the fractional contribution of different scenario groups to the
total hazard is computed. The most common form of deaggregation is two-dimensional in
magnitude and distance bins. Formally, it is the conditional probability of the ground motion that
exceeds a target value being generated by an earthquake in the specific magnitude and distance
ranges. For a specified return period and spectral period, the deaggregation in magnitude and
distance allows the dominant earthquake scenario (magnitude and distance) to be identified. The
controlling earthquake scenario will be different for different return period and spectral periods.
An example of a disaggregation of the T=0.2 and T=1.5 sec spectral accelerations for a site
in southern Illinois is shown in Figure 2. At the short spectral periods, the ground motion is
controlled by local earthquakes and at the long spectral periods the ground motion is controlled by
large magnitude earthquakes on the New Madrid zone. The UHS for a return period of 2500 years,
represents an envelope of the ground motion from these two different earthquakes. Therefore,
using the UHS leads to an unrealistic spectrum that cannot occur in a single earthquake. As an
alternative, multiple spectra for different controlling scenarios can be developed and used.

Figure 2. Example disaggregation of the hazard for short period (T=0.2 sec) and long period
(T=1.5 sec) spectral acceleration for a site in southen Illinios.

Conditional Mean Spectra

As an alternative to using UHS, Baker and Cornell (2006) developed the Conditional Mean
Spectra (CMS). The term “conditional” refers to the spectrum being anchored at a selected spectral
acceleration at a reference period of interest (such as the fundamental period of the structure). The
term “mean” refers to the mean value of the number of standard deviations (epsilon) of the ground
motion model at different periods given epsilon at the reference period, described later. The CMS
represents an average spectrum that would occur if the selected spectral acceleration at the
reference spectra period occurs.
3
 To construct a CMS, the first step involves disaggregating the hazard at a reference spectral
period, T0, and a specified return period, RP. The hazard leads to a spectral acceleration Sa(T0,RP).
The disaggreation will determine the magnitude and distance of the controlling earthquake
scenario. Using the parameters of the controlling earthquake scenario, ground motion attenuation
relationships are used to determine the median and standard deviation of the ground motion at
reference period T0 and at all other spectral periods T. The number of standard deviations, ε(T0,RP),
that the UHS at spectral period T0 is above the corresponding median spectral acceleration for the
controlling earthquake scenario is related to the median and standard deviation of the ground
motion for the controlling source and the spectral acceleration from the UHS:
SA(T0,RP) = Sˆa(M,R,T0 ,RP)exp(ε(T0 ,RP)σ (M,T0 )) (1)

where Sˆa(M,R,T0 ,RP) is the median ground motion, M and R are the magnitude and distance from
of the controlling source determined from the disaggregation, and σ is the standard deviation of the
ground motion at period T0. The mean value of epsilon at other spectral periods given the epsilon at
T0 is then found taking into account the correlation of the variability of the ground motion at
different spectral period:
ε (T,T0 ) = ε(T0 ) ρ(T,T0 ) (2)
where ρ is the correlation coefficient of the residuals from ground motion models. The CMS is
given by
CMS(T,T0,RP) = Sˆa(M,R,T,RP)exp(ε (T,T0 )σ (M,T)) (3)

An as example for the southern Illinois site (disaggregation shown in Figure 2), the CMS computed
at T=0.2 sec, T=1.5 sec, and T=3.0 sec are compared to the 2500 year return period UHS in Figure
3. This figure demonstrates how the UHS is an envelope of many different earthquakes and will
not occur in a single earthquake.

Figure 3. Example comparing the CMS at different spectral periods with the UHS for a return
period of 2500 years for a rock site in southen Illinios.
4
An issue with the application of the CMS has been that it is tied to a single reference spectral
period, T0. In the work by Baker and Cornell (2006), the fundamental period of the structure has
been used for the reference period, but what is to be done for structures that do not have a well
defined fundamental period? The concept of the CMS is not restricted to developing a single
spectrum based on the fundamental period of the structure. We can break the UHS down into a
suite of CMS at different T0 values that, when taken together, reproduce the UHS.. The problem is
that you need the CMS at many different T0 values if the full UHS is to be covered. For example,
the three CMS in Figure 3 do not meet the UHS at periods between 0.3 and 1.5 seconds. To keep
the CMS limited to a reasonable number of cases, the CMS can be broadened into scenario spectra,
such that the envelope of the CMS is equal to the UHS.

The choice of using a UHS rather than multiple spectra for the different scenarios is the
decision of the engineering analyst, not the hazard analyst. If it is worth the additional analysis
costs to avoid exciting a broad period range in a single evaluation, then the engineer should
request multiple scenario spectra from the hazard analyst.

In practice, the hazard analyst often only provides the UHS to the engineer in the hazard
report. A hazard report should also include a comparison of the UHS with the spectra from the
individual representative events indentified by the disaggregation. This gives the engineer the
information needed to make a decision whether to evaluate multiple scenarios one at a time or to
envelope the spectra from the multiple scenarios to reduce the number of analyses required.

Conditional Spectra

The use of CMS as a target design spectrum has the advantages of being representative of
realistic earthquake scenarios, but the CMS has two shortcomings: (1) the CMS represents mean
values and does not capture the peak to trough variability and (2) a suite of CMS cannot be
developed that reproduce the hazard curve (Abrahamson and Yunatci, 2010). For risk analyses, we
need to provide suites of ground motions with rates of occurrence that are both realistic and
reproduce the hazard. To address this need, we propose the development of conditional spectra
(CS). These are similar to the CMS, but the use of the mean epsilon has been replaced with
multiple realizations of the epsilons. The procedure for developing CS is given below.

The standard deviation of the peaks and troughs variability around a specific CMS can be
computed from the correlation of the epsilon values:

σ PT (T,T0 , M) = σ (M,T) 1− ρ 2 (T,T0 ) (4)

An example of the peak to trough range around a CMS is shown in Figure 4. This quantifies the
amount of peak to trough variability that should be included in the CS. At the reference period, T0,
there is no peak to trough variability because the spectrum is conditioned on this spectral value.
For spectral periods close to T0, there is only a small peak to trough variability because the epsilon
5
values are strongly correlated for nearby spectral periods. That is, there is typical width to the peak
and troughs. The conditional spectral can be developed by generating suites of correlated epsilon
values that are consistent with the peak to trough range shown in Figure 4. The procedure for
developing the correlated epsilons is given in the next section.

Figure 3. Example of the range of peak to trough variability about the CMS

Once the CS are developed, the second issue for risk analyses is the estimating of the rate of
each CS. The rate of occurrences of the CS are derived from the hazard curve. First, for the
reference period T0, the rate of occurrence of Sa(T0) between z1 and z2 can be compute from the
hazard:

Rate(z1 < Sa(T0 ) < z 2 ) = Haz(T0 ,z1 ) − Haz(T0 ,z 2 ) (5)

Next, each realization of the epsilons is equally likely, so the rate for each realization of the CS is
just the rate given in eq. (5) divided by the number of realizations of epsilons. If enough realization
are included, then using the CS and their rates will reproduce the hazard curves at all spectral
periods.

Generation of Correlated Epsilon Values

Spectral accelerations of earthquake ground motions are usually obtained from empirical
ground motion models at different spectral periods. The residuals of the ground motion prediction
models are correlated across spectral periods. The correlation of the residuals is needed to develop
conditional mean spectra and to perform vector hazard calculations. Correlation coefficients can be
6
obtained from Abrahamson and Silva (2008) or Baker and Cornell (2006). Table 1 shows the
Abrahamson and Silva (2008) unsmoothed correlation coefficients for the normalized intra-event
residuals. Using these correlation coefficients for the development of the conditional mean spectra
gives the mean epsilon values. Figure 5 shows an example of the relationship between the epsilon
values at two periods (0.1 and 0.2 seconds). The black line in Figure 4 represents the average
correlation between the epsilon values; however, a large scatter is observed around the average.
This scatter needs to be sampled for the development of scenario spectra.
The adopted procedure for the generation of different realizations of correlated epsilon
values is based on the principles of generating spatially correlated random fields described in
Walling (2009). Here we replace space and use spectral period. The goal is to generate different
sets of correlated epsilon values that can reproduce an experimental semi-variogram estimated
using a set of empirical data.

Variogram Analysis
In spatial statistics, a variogram is a function describing the degree of spatial dependence or
continuity of a spatial random field or stochastic process. It can be defined as:

[
2γ (Δh) = E (Z(h + Δh) − z(h))
2
] (6)

where γ (Δh) is the semi-variogram, z(h) is the value of the random variable at location h and Δh is
the location separation or lag. A variogram analysis consists of developing an experimental
variogram from a set of empirical data and then fitting a variogram model to the data. For ground
motion residuals, we find that a spherical model is applicable. The functional form of the spherical
model is given as:
⎡ Δh ⎛ Δh ⎞ 3⎤
c ⎢1.5 − 0.5⎜ ⎟ ⎥ for Δh ≤ a
γ (Δh) = ⎣ a ⎝ a⎠⎦ (7)
c otherwise

where the sill, c, is the value at which the semi-variogram levels off and the range, a, refers to the
lag value at which the semi-variogram reaches the sill value.

Applying the variogram analysis to epsilons, the random field to be modeled is epsilon and
the spatial lag, Δh, is replaced by the difference of the natural logarithm of spectral periods, ln(T1
/T2). Using the unsmoothed correlation coefficients, ρ, of the intra-event residuals given in Table 1
between periods of 0.1 and 5 seconds, the experimental semi-variogram was calculated over all
pairs of observations with different lag periods as:
⎡ ⎛ T ⎞⎤ ⎡ ⎛ T ⎞⎤
γ ⎢ln⎜ 1 ⎟⎥ =1− ρ⎢ln⎜ 1 ⎟⎥ (8)
⎣ ⎝ T2 ⎠⎦ ⎣ ⎝ T2 ⎠⎦

The empirical data were then fitted with the spherical semi-variogram model as shown in Figure

7
6 and the coefficients of the model were determined. Since epsilon values have a mean of zero
and variance of 1, the spherical model was constrained to level off at a value of 1.

Table 1. Unsmoothed correlation coefficients for intra-event residuals (Source:

Abrahamson and Silva, 2008)

PGA T=0.02 T=0.03 T=0.04 T=0.05 T=0.075 T=0.10 T=0.15 T=0.20 T=0.25 T=0.30
PGA 1.000 0.999 0.992 0.982 0.965 0.921 0.898 0.888 0.885 0.879 0.864
T=0.02 0.999 1.000 0.995 0.986 0.970 0.926 0.902 0.889 0.882 0.874 0.858
T=0.03 0.992 0.995 1.000 0.993 0.980 0.936 0.909 0.886 0.872 0.857 0.837
T=0.04 0.982 0.986 0.993 1.000 0.990 0.949 0.917 0.885 0.863 0.840 0.815
T=0.05 0.965 0.970 0.980 0.990 1.000 0.961 0.927 0.883 0.849 0.818 0.788
T=0.075 0.921 0.926 0.936 0.949 0.961 1.000 0.962 0.893 0.837 0.785 0.740
T=0.10 0.898 0.902 0.909 0.917 0.927 0.962 1.000 0.918 0.850 0.789 0.733
T=0.15 0.888 0.889 0.886 0.885 0.883 0.893 0.918 1.000 0.913 0.843 0.780
T=0.20 0.885 0.882 0.872 0.863 0.849 0.837 0.850 0.913 1.000 0.917 0.848
T=0.25 0.879 0.874 0.857 0.840 0.818 0.785 0.789 0.843 0.917 1.000 0.925
T=0.30 0.864 0.858 0.837 0.815 0.788 0.740 0.733 0.780 0.848 0.925 1.000
T=0.40 0.826 0.817 0.792 0.766 0.734 0.668 0.649 0.688 0.749 0.820 0.885
T=0.50 0.785 0.775 0.748 0.718 0.682 0.609 0.582 0.614 0.672 0.744 0.803
T=0.75 0.679 0.669 0.642 0.611 0.573 0.498 0.467 0.479 0.529 0.598 0.651
T=1.0 0.591 0.581 0.554 0.524 0.488 0.416 0.383 0.392 0.427 0.491 0.545
T=1.5 0.501 0.492 0.467 0.439 0.405 0.344 0.307 0.317 0.353 0.409 0.456
T=2.0 0.404 0.396 0.376 0.353 0.325 0.278 0.246 0.250 0.276 0.319 0.355
T=3.0 0.328 0.320 0.301 0.277 0.255 0.217 0.198 0.202 0.223 0.264 0.289
T=4.0 0.276 0.269 0.251 0.228 0.211 0.183 0.170 0.181 0.194 0.229 0.245
T=5.0 0.204 0.197 0.181 0.164 0.149 0.119 0.113 0.124 0.135 0.152 0.171
T=7.5 0.179 0.174 0.164 0.153 0.142 0.117 0.107 0.098 0.103 0.105 0.121
T=10 0.220 0.219 0.211 0.203 0.187 0.158 0.146 0.146 0.123 0.139 0.164

T=0.40 T=0.50 T=0.75 T=1.00 T=1.50 T=2.00 T=3.00 T=4.00 T=5.00 T=7.50 T=10
PGA 0.826 0.785 0.679 0.591 0.501 0.404 0.328 0.276 0.204 0.179 0.220
T=0.02 0.817 0.775 0.669 0.581 0.492 0.396 0.320 0.269 0.197 0.174 0.219
T=0.03 0.792 0.748 0.642 0.554 0.467 0.376 0.301 0.251 0.181 0.164 0.211
T=0.04 0.766 0.718 0.611 0.524 0.439 0.353 0.277 0.228 0.164 0.153 0.203
T=0.05 0.734 0.682 0.573 0.488 0.405 0.325 0.255 0.211 0.149 0.142 0.187
T=0.075 0.668 0.609 0.498 0.416 0.344 0.278 0.217 0.183 0.119 0.117 0.158
T=0.10 0.649 0.582 0.467 0.383 0.307 0.246 0.198 0.170 0.113 0.107 0.146
T=0.15 0.688 0.614 0.479 0.392 0.317 0.250 0.202 0.181 0.124 0.098 0.146
T=0.20 0.749 0.672 0.529 0.427 0.353 0.276 0.223 0.194 0.135 0.103 0.123
T=0.25 0.820 0.744 0.598 0.491 0.409 0.319 0.264 0.229 0.152 0.105 0.139
T=0.30 0.885 0.803 0.651 0.545 0.456 0.355 0.289 0.245 0.171 0.121 0.164
T=0.40 1.000 0.904 0.743 0.629 0.538 0.420 0.335 0.266 0.205 0.172 0.199
T=0.50 0.904 1.000 0.838 0.722 0.619 0.501 0.409 0.332 0.248 0.207 0.220
T=0.75 0.743 0.838 1.000 0.878 0.753 0.634 0.494 0.393 0.291 0.229 0.259
T=1.0 0.629 0.722 0.878 1.000 0.853 0.727 0.581 0.471 0.359 0.287 0.313
T=1.5 0.538 0.619 0.753 0.853 1.000 0.875 0.707 0.601 0.497 0.415 0.450
T=2.0 0.420 0.501 0.634 0.727 0.875 1.000 0.824 0.714 0.638 0.551 0.562
T=3.0 0.335 0.409 0.494 0.581 0.707 0.824 1.000 0.875 0.779 0.658 0.634
T=4.0 0.266 0.332 0.393 0.471 0.601 0.714 0.875 1.000 0.905 0.788 0.720
T=5.0 0.205 0.248 0.291 0.359 0.497 0.638 0.779 0.905 1.000 0.871 0.805
T=7.5 0.172 0.207 0.229 0.287 0.415 0.551 0.658 0.788 0.871 1.000 0.921
T=10 0.199 0.220 0.259 0.313 0.450 0.562 0.634 0.720 0.805 0.921 1.000
8
 3

Epsilon (0.2 sec)

1

0
-2 -1 0 1 2 3
-1

-2
Epsilon (0.1 sec)

Epsilon Values Trendline

Figure 5. Correlation of epsilon values at periods of 0.1 and 0.2 seconds

Figure 6. Semi-variogram model for periods of 0.1 to 5 seconds

9
The procedure for generating one suite of the correlated epsilon is as follows:
1. Compute the covariance function given by 1-γ.
2. Extend the covariance function to negative lags to define an even function
3. Compute the Fourier transform of the extended covariance function
4. Sample random phase angles
5. Compute the inverse Fourier transform
As a test of the method, the semivariogram of the generated epsilons can be compared to the model
(Figure 7).

Figure 7. Comparison of the semi-variogram model and the semi-variogram of the generated
epsilon values

Conditional Spectra Example

As an example, CS are developed for a structure with a fundamental period of 0.5 seconds
for a site in the San Francisco Bay area. Using the procedure described above, suites of CS can be
developed. It s easy to generate a large number of realizations, but there are limits to the number of
scenarios that the engineering analysis can consider. In this example, we limited the number of
realizations of epsilons to 10 for each return period. Figure 8 shows the 10 realizations for a single
return period. The CS have the same value at the reference period of 0.75 seconds, but they have
peaks and troughs in the spectra. Here, the sampling with period has only 20 spectral periods from
0.1 to 5 seconds, limiting the resolution of the peaks and troughs.

As a check, the hazard from the suite of CS can be computed and compared to the original
hazard curves (Figure 9). The hazard for the reference period is reproduced closely as this is used
to set the rates (eq. 5). At T=0.2 seconds, the hazard computed from the small set of CS is similar
to the hazard from the PHSA for return periods from 200 to 5000 years. The underestimation of the
10
hazard at low return periods results from not considering more small return periods in defining the
CS. If a larger number of realizations is used, for example 30 per return period, then the agreement
between the hazard from the CS and the PSHA is much better, but that leads to a larger number of
cases than most projects will be willing to consider. There is a trade-off between accuracy of the
hazard at all spectral periods and the number of realizations that can be considered in the structural
evaluations.

This method can also be used to select time histories for dynamic analyses of structures. If
the scaling approach is used, then the CS can be used to define the range of the peak to trough
variability after scaling to the spectrum at T0. If spectral matching is used, then the CMS can be
used as a target with weak matching criteria to preserve the peak to trough variability given by the
CS.

Figure 8. Example of 10 conditional spectra for a ground motion level of 0.93g at a spectral
period of 0.75 seconds showing peak to trough variability.

11
Figure 9. Comparison of the hazard fro the PHSA and the hazard from the CS using 10
realizations for each ground motion level. The left plot is for the reference period
T0=0.75 sec and the right hand plot is for a period of 0.2 sec.

Conclusions
The conditional mean spectra and the conditional spectra provide methods for breaking the UHS
down into realistic earthquakes. The choice of using a UHS rather than multiple spectra for the
different scenarios should be the decision of the engineering analyst, not the hazard analyst. If it
is worth the additional analysis costs to avoid exciting a broad period range in a single
evaluation, then the engineer should request multiple scenario spectra from the hazard analyst.
For risk calculation for structures that have a response that depends on widely separated spectral
frequencies, then the conditional spectral approach can be used.

References
Abrahamson, N. A. and A. A. Yunatci (2010). Ground motion occurrence rates for scenario spectra, Proc.
Fifth International Conference on Recent Advances in Geotechnical Earthquake Engineering and
Soil Dynamics, May 24-29, 2010, San Diego, CA,

Abrahamson, N. A. and W. J. Silva, 2008. Summary of the Abrahamson and Silva NGA ground motion
relations, Earthquake Spectra, 22 (1), 67 – 98.

Baker, J. W. and C. A. Cornell, 2006. Spectral shape, epsilon and record selection, Earthquake
Engineering and Structural Dynamics, 35, 1077 – 1095.

Baker, J. W. and C. A. Cornell, 2006. Correlation of response spectral values for multicomponent ground
motions, Bulletin of Seismological Society of America, 96 (1), 215 – 227.

Bazzurro, P. and C. A. Cornell, 1999. Disaggregation of seismic hazard, Bulletin of the Seismological
Society of America, 89 (2), 501 – 520.

Walling, M. A., 2009. Non-ergodic probabilistic seismic hazard analysis and spatial simulation of
variation in ground motion, PhD Dissertation, University of California, Berkeley
12