Bulletin of the Seismological Society of America, Vol. 85, No. 3, pp.

937-942, June 1995

Short Notes

A Probabilistic Approach to Ground-Motion Selection

for Engineering Design
by Martin C. Chapman

Abstract The solutions of many earthquake engineering problems involve dy-

namic analyses using ground-motion time series. It is often desirable to base the se-
lection of such motions on a probabilistic estimate of the seismic hazard. The hazard
density function evaluated at a chosen hazard level provides the information neces-
sary to determine objectively the most likely earthquake events, defined by magni-
tude and distance, that contribute to seismic hazard. For a wide range of hazard
models it is possible to show that the difference between the median motion at a
site, given the occurrence of the most likely event, and the motion value correspond-
ing to a specified hazard level, is due entirely to the modeling of random error in the
strong-motion data set. This points to a straightforward approach to selecting
ground-motion recordings that represent the most likely time-domain realizations of
the hazard model for a given motion parameter and hazard level. Ground-motion
time series selection and/or synthesis based upon this approach, for various fre-
quency bands of the response spectrum, can provide an optimum basis for seismic
design.

Introduction Background

The method of probabilistic seismic hazard analysis de- An expression for seismic hazard at a given site due to
veloped by Cornell (1968) has been used widely: see Reiter a single earthquake source is as follows (Reiter, 1990):
(1990) and the National Research Council (1988) for sum-
maries of the method and its applications. A common ob- O~ r~
jective of the analysis is the estimation of a "hazard func- E(x) = fR(r)fM(m)G(X > xlm, r)dmdr. (1)
in rain

tion," defining the frequency of exceeding a range of

ground-motion or harmonic-oscillator response amplitudes. Here, E(x) is the expected number of exceedances per unit
This function is a compact representation of the hazard time of ground motion (or oscillator response) amplitude x,
posed by the potential occurrence of all damaging earth- and c~ = the number of earthquakes per unit time with mag-
quakes at all possible locations with respect to a specific site. nitude m between minimum and maximum values m,,in and
Because the hazard function is the result of an integration mm~;fR(r) = probability density of earthquake-site distance;
over magnitude and distance, the relative contributions to fM(m) = probability density of earthquake magnitude; and
total hazard due to earthquakes of given magnitude and dis- G(X > xlm, r) = conditional probability that, given an earth-
tance are obscure. However, the capability of identifying the quake of magnitude m at distance r, the motion intensity X
major contributors to hazard has considerable importance. will exceed a specified value x.
For a given site condition, the earthquake magnitude and The probability densityfR(r), nonzero between limits rmi~
distance are the dominant parameters that determine the and r,n,~, depends upon the geometry of the earthquake
character of the ground motion. Many engineering problems source, which usually takes the form of a point, line segment,
can be addressed optimally if representative ground-motion or bounded surface. The magnitude density fM(m) depends
time series, selected or synthesized on the basis of realistic upon the earthquake recurrence model. Often, fM(m) is as-
earthquake scenarios, are available for dynamic analysis. It sumed to be exponential, truncated at lower and upper limits
is shown below that the probabilistic method, when com- mmin and m,,~: however, that functional form is not a required
bined with frequency dependent ground-motion prediction assumption. In equation (1), distance and magnitude are as-
models, can provide a formal, objective approach to the se- sumed to be statistically independent. In some cases, it is
lection of time series for dynamic analysis. necessary to construct a seismic hazard model wherein these

937
938 Short Notes

random variables are dependent: for example, in the mod- ,I .... i .... I - , - ,-.) . . . . t, ,pj ~--,L, I .... I .... i,
40.5
eling of earthquakes as finite ruptures along a fault (e.g., . " , . 7 0 KM

Bender, 1984). The distance in a fault model could represent

40.0 / o ~ 0
the nearest horizontal distance from a site to the projection I II o o

of the rupture on the ground surface. For a vertical fault, this 39.5
o 81 0
0 o o 0
distance would depend on the length of the rupture segment, @
¢-~
o o 0o°
MAGNITUDE
oo ~ o
which in turn is a function of magnitude. In that case, fe(r) 39.0
o
.0 o ~ . oo 07
0) o o-I- • 06
in equation (1) is replaced with the conditional density func- "O 0 o o o 0 ~.~ * o 0 * 5

tion fea~(rlm), interpreted to indicate the probability density 38.5 tD

0 0 0 * 0 o o
of distance R, given that magnitude M has taken on the value 0 o
0 o
~ *
._1
m. Some further implications of statistical dependence in- 38.0 ~ o 0 ~ o

volving distance and magnitude will be discussed below. " o

Finally, if more than one earthquake source can affect the 37.5
0 ~ 0 C~ o~ J

site, and if the sources are independent, the total seismic o ~ 0

J I i , ) L I , = , , I I ,~,'i I-PJ ~J I , ~',i)ililllnlii~Jl~
hazard is the sum of hazard functions for the individual -79 -78 -77 -76 -75
sources. Longitude (Deg.)
The remainder of this article addresses an allied prob-
lem: objective determination of the magnitudes m and dis- Figure 1. Hypothetical epicenter map for the ex-
tances r to associate with a given hazard level. For example, ample discussed in the text. The map represents a 50-
a user may wish to select an actual strong-motion recording yr period.
or create a synthetic-motion time series, that is in some sense
"compatible" with a specific value of E(x). This is a com-
plicated problem that in general has no unique, deterministic magnitude, at that distance, is the most likely event that can
answer. As indicated by equation (1), seismic hazard is the result in motion exceeding x.
result of potential earthquakes that may exhibit a range of For the more general case in which there are n statisti-
magnitudes and may occur over a range of distances from a cally independent sources of earthquakes, we have
site. Therefore, the problem of selecting one or more specific
earthquake events, as required for some engineering appli- (3)
U(m, rlx) = ~ c(ifM,(m)fR,(r)G(X> xlm, r),
cations, is fundamentally probabilistic in nature. An appro- i=1
priate approach is to base this selection on a formal proba-
bilistic model of the seismic hazard. where the subscript i refers to the ith source. This can result
As shown below by example, it is possible to identify potentially in multiple local maxima of U, in which case the
the most likely (i.e., most frequent) events, defined in terms consideration of multiple design events is indicated.
of magnitude and distance, that contribute to seismic hazard
E(x). Further, as demonstrated in the Appendix for a large
class of seismic hazard models, the difference in motion in- Example
tensity Xzo~, predicted at a site given the occurrence of the As an example, consider the hypothetical problem sug-
most likely or "modal" event, and x, the motion intensity gested by Figure 1. The hazard is posed by two independent
corresponding to the hazard level E(x), is due entirely to the sources: a line source with nearest approach to the site of 30
modeling of random error in the strong-motion data set. This km, and a "background" source area, enclosed by a circle of
indicates a straightforward approach to the selection of radius 200 km. The example represents an active fault, em-
strong-motion recordings representing the most likely mo- bedded in a relatively less seismic region. For both sources,
tions for a given hazard level. we will assume the following recurrence relationship:

T h e M o d a l Event log N = 4.101 - 0.8 m, (4)

Let the integrand of equation (1) be represented by

where N is events per year, implying two magnitude 6 or
greater earthquakes per decade within each source. We will
U(m, rlx) = o~fM(m)fR(r)G(X > xlm, r). (2)
assume a truncated exponential density function for magni-
tude of the form
Note that U is not a probability density. However, U can be
considered the joint "hazard density" for the frequency of
b'e-b'm
exceeding motion x. In analogy to the definition of the mode =
fM(m) e -b'mm~ _ e_b,m,,~ , m,.in < m < mm,~, (5)
of a random variable, let the modal event for a specified
motion parameter value x be defined as magnitude m and
distance r such that U is maximum. An earthquake of that where b' = 0.8 In (10). For the line source, let minin = 5.0
ShonNo~s 939

| i ! i i i i i i | i i i i i i i

'L

•. - 0.1
l " = IItllUlIilIIIIIIlIII
v
__~~ /
1111111
-.-.-~IIIIIMIII~
¢j
r- \ \\ i
q~ 10-2
-.I \\ \ \

o" \ \

i,
1 Hz
\ \
10 -3
r-
10 Hz '\\\\\\ """~,,

O,,) 10-4
X \ .,.
i,i
\ ", 5 Hz
\ '.
lO-S i n n ~ ~ Ill I I I I I"11 !

lO lO~
"~
PSV (cm/sec) ~. ~

Figure 2. Seismic hazard functions (PSV re-

sponse) for the example discussed in the text.

o
and m,,a~ = 7.7. In the background, let mm~, = 5.0 and mmax
= 6.5. The expected number of earthquakes per year with
magnitudes between m,,, and rnma~is

cz = 104-1°1-°-8. . . . . l04-101-0-8m~. (6)

~| x = 9.3 cm/sec
The probability density of epicentral distance for the back-
ground source is
N~ -.:,~lllHIIIlllllllllllllN.--,o
2?"
f~(r) = ~ , rm,. = 0 <=r <- rmo~, (7)
?"max

where r , ~ is the radius of the source area (200 km). For the
line source, assuming that earthquakes can occur anywhere
along the line with uniform probability, independent of mag- -~..,~,=.~, , . , . - ~ ~
~,o -,~~-
nitude, we have
Figure 3. Hazard density functions U(m, rlx) for
2r the example discussed in the text. (Top) 1-Hz PSV;
fR(r) - 30 < r < ~/302 + LV4, (8) (middle) 5-Hz PSV; (bottom) 10-Hz PSV. The three
30 ' functions are for a hazard level of 2.0 × 10-3 1/yr.

where L is the length of the line (387 km).

Finally, assume that the motion parameter X is a log- ard density functions U(m, r Ix) for the three oscillator fre-
normally distributed random variable. Let X be pseudorela- quencies are shown in Figure 3, where x is PSV response
tive velocity (PSV) response, given an earthquake of mo- such that E(x) = 2.0 X 10 -3 (500-yr return period). The
ment magnitude m at distance r. The mean logarithm of X total volume represented in each plot is equal to 2.0 X 10 -3.
is derived from the prediction equations of Joyner and Boore The three hazard density functions are bimodal, clearly re-
(1982) for 5% damping and rock site conditions: the con- flecting the contributions to hazard from the two sources.
ditional probability G ( X > xlm, r) is obtained from the cu- The line source (fault) dominates the hazard for the
mulative lognormal distribution function, with specified 1-Hz oscillator. The maximum of U corresponds to magni-
standard deviation ~ (see Appendix). tude 6.9 at 30 km. From Figure 2, the response amplitude,
Figure 2 shows the hazard functions for the above ex- x, corresponding to the 500-yr return period is 28 cm/sec for
ample, computed for 1-, 5-, and 10-Hz oscillators. The haz- the 1-Hz oscillator. However, the median oscillator response
940 Sho~ No~s

for the modal event, Xmode, is 14 cm/sec, which is only 1/2 statistically dependent may not satisfy this condition in all
the response amplitude for the chosen exceedance fre- cases (see Appendix). The second condition, implied by
quency. In our example, this difference between the median equation (3), is that G ( X > xlm, r) remain invariant among
motion prediction for the modal event, Xmode, and x arises the sources contributing to hazard. This amounts to using
because the predicted oscillator response is treated as a ran- the same attenuation model H(m, r) and probability distri-
dom variable. This element of the seismic hazard model bution to predict ground motion for each source. Finally, the
complicates the problem of design event selection, but is third condition is that H(mmi,, r) <=log x for rmin <= r <=rmax.
very necessary: it is observed that motion intensity at a given AS in the Appendix, H(m, r) here represents the attenuation
distance from an earthquake exhibits statistical variation or model that is used to predict median values of motion X. All
"scatter," here represented by the standard deviation a. Al- three of these conditions are satisfied by a broad range of
though the scatter associated with a particular motion pre- useful seismic hazard models, in addition to the simple ex-
diction model can, in principle, be reduced by modeling ad- ample discussed previously.
ditional information on the earthquake source, propagation The approach outlined above provides the user with a
path, and site response, it cannot be eliminated entirely. A time-domain estimate of the most likely ground motions at
significant reduction in the scatter is particularly difficult a given site corresponding to a predefined hazard level and
when the locations and magnitudes (and associated source oscillator frequency. However, complex structures may ex-
and path effects) of future earthquakes are uncertain. It is hibit several response modes. This, combined with the fact
shown in the Appendix that in the special case where o- = that both the shape and amplitude of the ground-motion
O, Xmode = X in a large class of seismic hazard models obeying spectrum change as functions of distance and magnitude,
equations (1) and (3). This implies that in many cases, the means that several time series may be necessary to represent
quantity Xmode -- X is due entirely to the modeling of random properly, for engineering purposes, the most likely motions
scatter in the observational strong-motion data set, and in- indicated by the hazard analysis. These points are further
dicates the following approach to time-series selection. addressed below.
In the example, the base 10 logarithm of oscillator re-
sponse is assumed to be normally distributed with a = 0.33. Discussion
The logarithm of oscillator response corresponding to E(x)
= 2 × 10 -3 is approximately 1 standard deviation above The shape of the ground-motion spectrum and, there-
the predicted mean logarithm of response for the modal fore, the response spectrum is strongly dependent upon dis-
event. Therefore, given the occurrence of the modal event, tance and magnitude. The maximum of U will vary depend-
there is approximately a 16% probability that the resulting ing upon the frequency of the oscillator, as well as upon the
PSV response at the site will exceed x = 28 cm/sec, for E(x) response amplitude x. This means that for a fixed frequency
= 2 × 10 -3. For dynamic analyses at frequencies near 1 of exceedance E(x), the modal event for a high-frequency
Hz, a ground-motion time series consistent with the results oscillator will generally differ from that of a low-frequency
of the example hazard analysis would be selected at the 84% oscillator. The same can be said for any motion parameter
percentile from the population of time series recorded at 30 that is frequency dependent. Typically, the modal events for
km from magnitude 6.9 earthquakes. Because this popula- the higher frequencies will tend to be of smaller magnitude
tion is small, a more practical approach is to select or syn- at smaller distances, compared to the modal events for the
thesize a "best estimate" ground-motion time series repre- lower frequency motion parameters. Generally, a unique
sentative of the modal event, and scale the amplitude of the modal event cannot be defined for the entire response spec-
time series such that the PSV response is x corresponding to trum at a fixed exceedance frequency: i.e., a single modal
the design E(x). event will not generally match the uniform hazard response
It is important to note that this approach is strictly valid spectrum. For this reason, multiple design time series should
only for a narrow frequency band. Further, as shown in the be developed for the specific frequency band(s) of engi-
Appendix, the interpretation of Xmode -- X as due entirely to neering concem in regard to structural response.
the modeling of random scatter in the observational data is These points are illustrated in Figure 3, which shows
strictly valid only for hazard models wherein three condi- that the contribution to seismic hazard shifts to smaller earth-
tions are satisfied. First, the partial derivative with respect quakes, at smaller distances, as oscillator frequency in-
to magnitude of the joint magnitude-distance probability creases. This is a consequence of attenuation of the higher
density function of each individual source contributing to frequency ground motion, and the shape and magnitude scal-
hazard at a given site must be negative: i.e., fR(r)fM(m) and/ ing of the earthquake source spectrum. In the example, there
or fRM(r Im)f~(m) must decrease with increasing magnitude. is clearly a need to consider two design events for the 10-
This condition is always satisfied for the common situation Hz oscillator: a magnitude 6.5 shock at 30 kin, due to the
where distance and magnitude can be treated as statistically line source, and a magnitude 5.4 event at 8 km in the back-
independent and the magnitude density functions of the var- ground. The events contribute equally to seismic hazard at
ious sources are assumed exponential. However, a subset of 10 Hz, yet their time series can be expected to be very dif-
the group of models wherein distance and magnitude are ferent: e.g., the larger, more distant shock would generate
Short Notes 941

strong motions of much longer duration. Thus, the two

G ( X > xlm, r) = 1 - - -
events could have very different consequences for certain 2~-~
types of construction.
In some cases, particularly those in which nonlinear be- - (u -
havior of structures or soils must be considered, the duration exp[-2~Y)Z]du, (A2)
of shaking as well as the amplitude of motion can be an
important consideration for design. The approach described
where y = H(m, r) is the mean logarithm of the motion
above does not take duration into account because the ran-
parameter at distance r from an earthquake of magnitude m;
dom variables serving as the basis for the hazard estimates
In x = the logarithm of the specified motion parameter cor-
(oscillator response) are essentially independent of the du-
responding to hazard level E(x); u = dummy variable of
ration of ground shaking. In principle, a duration-dependent
integration; and a = standard deviation of the logarithm of
parameter such as a measure of energy dissipated by a linear
the motion parameter.
(or nonlinear) system might be used as the basis variable for
From equation (A2), for y > > In x, G approaches 1.0,
the hazard analysis. An approach similar to that discussed
and for y < < In x, G approaches 0. However, for a = 0, G
previously could be performed to identify the modal events
= 1 if y > In x and G = 0 if y < In x, because in that limit,
and select appropriate time series, provided that the duration-
the integrand is the Dirac delta function 6 (x - y). Therefore,
dependent parameter is predictable as a function of magni-
from equation (A1) and for o- = 0, we have
tude and distance.

U(m, rlx) = afR(r) fM(m) y > In x,

Acknowledgments
= 0 y < In x. (A3)
I sincerelythank William Joyner and Leon Reiter for their review efforts.
Their comments greatly improved this article. I also thank J. Arthur Snoke
and ShaosongHuang for helpful suggestions and assistance. The work was From equation (A3), values of m and r defining the nonzero
sponsored by the Federal EmergencyManagementAgency and the Virginia domain of U must satisfy
Department of Emergency Services.
y = H ( m , r) > In x, (A4)
References
where H(m, r) is an attenuation function, usually of the form
Bender, B. (1984). Seismic hazard estimation using a finite-fanlt rupture
model, Bull Seism. Soc. Am. 74, 1899-1923.
y = C1 + C2m + C 3 1 n r + Ca r . (A5)
Cornell, C. A. (1968). Engineering seismic risk analysis, Bull Seism. Soc.
Am. 58, 1583-1606.
Joyner, W. B. and D. M. Boore (1982). Prediction of earthquake response The domain of U is further bounded by minimum and max-
spectra, U.S. Geol. Surv. Open-File Rep. 82-977. imum distances r,,in and rm~ and magnitudes minin and mmo~,
National Research Council (1988). Probabilistic Seismic Hazard Assess- as shown schematically in Figure A1.
ment, National AcademyPress, Washington, D.C.
We are now in position to examine the relative values
Reiter, L. (1990). Earthquake Hazard Analysis: Issues and Insights,
Columbia Univ. Press, New York. of U. Partial differentiation of U with respect to magnitude
m gives

Appendix
aU dfu(m )
-- = c~fR(r)- (A6)
The interpretation of the quantity Xmode -- X given in the Orn dm
text can be justified by considering in some detail the hy-
pothetical case of no random error in the strong-motion data Note that the arbitrary distance density fR(r) is assumed in-
set. In the text, it is asserted that in many cases of practical dependent of magnitude, and, like all probability densities,
value, Xmode = X if a = 0. Below, I demonstrate the basis of is a positive function. In the case of an exponential magni-
this assertion, and define the conditions under which Xmoae = tude density function, as considered previously in the text,
X for o- = 0. I also identify a subset of hazard models in we have
which Xmoae = x may not hold in all cases.
Recall that, for a single earthquake source, OU
- o~b'2Afn(r)e -b'm, (A7)
Orn
U(m,rlx) = o~fR(r)fM(m)G(X > xlm, r) (A1)
where c~, b', and A are positive constants. From equation
represents the hazard density function. At this point, the only (A7), we note that U is a decreasing function of m, for any
restrictive assumption is that r and m are statistically inde- fixed value of r. Therefore, values of r and m corresponding
pendent. We represent the conditional probability G using to the maximum value of U (i.e., the modal event) must
the complementary Gaussian probability distribution as satisfy
942 Short Notes

y = H(m, r) = In x, (A8)

and therefore, Xmode = X, throughout the domain of nonzero

U, provided that

H(mmin, r) < In x, for r~i, <_ r <= r ~ . (A9)

7 An example of the violation of this condition is shown in

0 Figure A1, where the domain of U is artificially bounded at
"0 m = mmin such that n(mmin, r) ~> In x over a distance range
Z~
rmi, to r. In practice, the condition represented by equation
¢-
(A9) amounts to ensuring that the minimum magnitude trun-
cation value mmin is consistent with the selected ground-mo-
tion parameter values x at which hazard is to be estimated.
Failure to meet this condition results in underestimation of
hazard E(x) for the smaller values of x.
The above argument can be extended to other models
5 in which the functional form offu(m) is other than expo-
nential, as well as to cases where distance depends upon
magnitude. In general, Xmode = X for a single source when
0 20 40 equation (A9) is satisfied and when
8
OU
- - < O. (AIO)
Om

In the case where fRM (rim) replaces fR(r), equation (A10)

7 becomes
0
"0
OU d
.,l,,..a - - = o~feu(rlm) - z - f , ( m )
Om elm'-
t""
o
+ ctf~(m) w ' - f e ~ ( r l m ) < O. (A11)
z; om

In this case, O/Omfr~(rlm) can be positive, and may lead to

a situation where O/Om U > 0, for negative d / d m f M ( m ) . This
5
indicates that Xmode ~ X f o r a subset of the models wherein r
and m are dependent.
If multiple, statistically independent sources contribute
0 20 40 to seismic hazard at a given site, then U(m, rJx) is given by
Distance (Km) equation (3) in the text as the sum of hazard densities for
the individual sources. Provided that the conditions repre-
Figure A1. A schematic representation of the sented by equations (A9) and (A10) are satisfied for each
nonzero domain of U(m, fix) for a single source is source, U also satisfies both conditions. Therefore, as before,
shown as the shaded area. The heavy line shows the X~oae = x. This assumes, of course, that the same ground-
locus of points satisfying H(m, r) = In x, where H(m, motion-prediction model H ( m , r) is used for all sources.
r) is an attenuation function. Top: Situation where
Xmode = X, provided that fR(r)fM(m) or fR~t(rlm)fu(m)
Department of Geological Sciences
is a decreasing function in the direction of increasing
magnitude. Bottom: The case where Xmode = X may Seismological Observatory
Virginia Polytechnic Institute and State University
not hold because of the lower magnitude limit m,,i,.
Blacksburg, Virginia 24061-0420

Manuscript received 1 June 1994.