You are on page 1of 20

EARTHQUAKE ENGINEERING & STRUCTURAL DYNAMICS

Earthquake Engng Struct. Dyn. (2012)


Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/eqe.2231

Development of earthquake vulnerability functions for tall buildings

Nirmal Jayaram*,†, Nilesh Shome and Mohsen Rahnama


1
Risk Management Solutions Inc., Newark, CA 94560, USA

SUMMARY
This study focuses on the development of vulnerability functions for tall buildings. A systematic simulation
approach based on the Pacific Earthquake Engineering Research loss assessment framework is used to develop
building vulnerability functions that provide estimates of losses to buildings under ground motions of various
intensities. The steps involved in the procedure are: quantifying ground-motion hazard using a vector of spectral
accelerations; predicting building response parameters such as story drifts, floor accelerations, and residual drifts
under the quantified hazard; accounting for structural collapse and demolition; and predicting story-wise losses
and total building loss using the building response information. Emphasis is placed on capturing the effects of
epistemic and aleatory uncertainties in random variables, such as ground motions, structural response para-
meters, loss costs, etc., to quantify the uncertainty in the final loss estimate. The risk assessment approach is used
for developing vulnerability functions for six tall buildings, namely, 20-story and 40-story steel moment
resisting frame buildings based on 1973 and 2006 codes, a 42-story concrete core wall building and a 42-story
concrete dual system building. The vulnerability functions are used to perform loss assessments for individual
buildings assumed to be located in Los Angeles. The vulnerability and the loss assessment procedures are
illustrated in detail for the 2006 20-story steel moment frame building, and a summary of the final loss estimates
are provided for all other buildings. It is seen that epistemic uncertainties in both ground motion hazard and
building vulnerability cause significant epistemic uncertainties in the loss assessment results. Copyright ©
2012 John Wiley & Sons, Ltd.

Received 28 March 2011; Revised 11 April 2012; Accepted 14 May 2012

KEY WORDS: tall buildings; vulnerability functions; seismic risk assessment; epistemic uncertainty
quantification

1. INTRODUCTION

As of October 2010, there are over 200 tall buildings (over 15 stories tall) in both Los Angeles and San
Francisco. The mixed performance of tall buildings during past earthquakes including the 2010 M8.8
Chilean earthquake has cast doubts on the performance of tall buildings in the United States, and
consequently, researchers have recently started focusing on evaluating the seismic capacity of tall
buildings, for example, in [1, 2]. The current study evaluates the risk of earthquake losses to tall
buildings in the Western United States, and in particular in California.
The risk assessment approach used throughout the insurance industry primarily involves
probabilistically quantifying the seismic monetary losses to a building in terms of the average annual
loss and various return period losses. The computation of these monetary risk measures requires
information about the seismic performance of the building under earthquake ground motions, which is
often quantified using a vulnerability function (also known as damage function) that provides the ratio
of the building loss to the building value during ground motions of different intensities. The
vulnerability function can either be developed using observations of building performance during past

*Correspondence to: Nirmal Jayaram, Risk Management Solutions, Inc., Newark, CA 94560, USA.

E-mail: nirmal.jayaram@rms.com

Copyright © 2012 John Wiley & Sons, Ltd.


N. JAYARAM, N. SHOME AND M. RAHNAMA

earthquakes, or by estimating the performance of an analytical model of the building and probabilistically
sampling its losses under possible future earthquakes. Because of the scarcity of tall buildings, there is
generally a lack of tall building damage data even in seismically-active regions such as California.
Hence, this study uses the second approach of sampling losses to an analytical model.
The building vulnerability functions are developed using a systematic simulation approach based on the
Pacific Earthquake Engineering Research (PEER) loss assessment framework. The steps involved in the
procedure are quantifying ground-motion hazard using a vector of spectral accelerations; predicting
building response parameters such as story drifts, floor accelerations, and residual drifts under the
quantified hazard; accounting for structural collapse and demolition; and predicting story-wise losses
and total building loss using the building response information. Emphasis is placed on capturing the
effects of epistemic and aleatory uncertainties in random variables, such as ground motions, structural
response parameters, loss costs, etc., to quantify the uncertainty in the final loss estimate. The risk
assessment approach is used for developing vulnerability functions for six tall buildings, namely,
20-story and 40-story steel moment resisting frames (SMRFs) based on 1973 and 2006 codes, a
42-story concrete core wall building and a 42-story concrete dual system building. The 42-story core-
wall and dual-system building are expected to be representative future tall-building constructions in
California. These buildings have been designed and analyzed by the PEER Tall Building Initiative
(TBI) project [3] for reviewing the performance of future tall buildings in California. The 20-story and
40-story SMRFs are representative of a majority of tall building constructions in California.
The vulnerability functions are used to perform loss assessments for individual buildings assumed to be
located in Los Angeles, using a stochastic catalog of possible future earthquakes. The monetary loss
measures presented here are the average annual loss and the 500-year return period loss. In addition, the
manuscript also presents the epistemic uncertainties in these loss measures as a result of the epistemic
uncertainties in both ground-motion hazard and building vulnerability.
This manuscript focuses on a vulnerability function assessment methodology for tall buildings. While
quantifying the vulnerability functions for the structural models considered in this study, numerical values
are used for the medians and dispersions of fragility and loss functions, correlations between component
capacities and between incurred losses of building subsystems, etc. Some of these values are well-founded
on engineering principles and available data, whereas others are based on judgment that depends on the
current state of knowledge. For this reason, the results presented at the end are representative but cannot
be considered as final products.

2. BUILDING MODELS

In this study, we consider six different tall buildings, namely, 20-story and 40-story SMRF buildings
designed based on 1973 and 2006 seismic codes, a 42-story concrete core-wall hotel tower and a
42-story concrete dual-system building, as representative tall buildings in California. The four SMRF
buildings were designed by the authors, while the two concrete buildings were designed as part of the
PEER TBI. Design details and analysis results for the latter two buildings are available in [3]. The
design and the analytical modeling of the SMRF buildings are briefly summarized in the following
paragraphs.

2.1. Steel moment resisting frame structures


In the 1970s, when most tall buildings in the Western United States were built as steel moment resisting
frame structures, it was customary to use moment resisting connections at most or all beam to column
intersections. This resulted in ‘complete steel moment frames (MFs)’, implying that essentially all
frames in the two orthogonal directions are moment resisting. This practice was abandoned in the
1980s for economic reasons and was replaced with the use of ‘perimeter frame’ structures in which
only two frame lines in each direction were made moment resisting and the interior framing was used
primarily to resist gravity loading. Because this change in practice caused a considerable change in
member sizes, it was decided in this study to develop designs according to the 1973 Uniform Building
Code (complete MFs) and the 2006 International Building Code (IBC; perimeter MFs with reduced

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

beam section connections). Complete MFs have many more beams and columns that contribute to lateral
load resistance than perimeter MFs. Beams are often much smaller, but most columns are controlled by
gravity loading and may be of sizes comparable to those of perimeter frame columns. A basic
difference between complete and perimeter MFs, however, is that the columns in complete MFs are
usually stocky W14 sections, whereas they are often deep and relatively slender W30 or lW36 sections
in perimeter MFs.
Tall steel MFs designed in accordance with recent code provisions are controlled mostly by stiffness
(drift control) and not strength considerations. Drift limitations require large beam sizes, and strong
column–weak beam and overturning moment requirements demand that column sizes also be
increased. Thus, the code design base shear, which is used for strength design, is often a poor
measure of performance or even strength of the structure. In the 1970s, the implementation of drift
control was voluntary. Its neglect decreases beam stiffness and strength considerably, which can be
detrimental to seismic performance. It also increases story drifts (perhaps the most important
engineering demand parameter or EDP) considerably. Consultation with engineers who designed
steel MFs in the 1970s led to the conclusion that it was good practice, even then, to implement drift
control by providing an allowable story drift of about 0.0025 under code seismic and wind loads.
Therefore, both the 20-story and the 40-story 1973 steel MFs designed in this study comply with
this drift limitation.

2.2. Analytical model of steel MFs


Simplified analytical models are used in this study, given the numerous uncertainties in architectural and
structural design decisions. Torsional effects and biaxial effects in columns are neglected, and therefore,
only two-dimensional analytical models are employed. All moment-resisting bays are lumped into a
single-bay moment frame, and P-Delta effects that are tributary to the gravity system are accounted
using a leaning column. Overturning moment effects on strength (column axial forces) and stiffness
(flexural mode of displacement) are simulated by adjusting the bay width of the single bay frame. Such
a model greatly reduces the computational effort and often facilitates interpretation of global results.

2.3. Component model, analytical platform, and representative results


The analytical models of the tall buildings are analyzed using the latest version of the OPENSEES platform
(http://opensees.berkeley.edu). The models for the flexural behavior of steel components are based on the
point hinge concept and utilize a backbone curve that is discussed in detail in the PEER/ATC-72-1 report
[4]. The deterioration parameters of the backbone curve are based on regression equations developed
using experimental results [5], and are functions of the geometric section and material properties that
control deterioration in strength and stiffness because of local and lateral torsional buckling. Modeling
deterioration in component properties is important particularly at large ground-motion intensities where
dynamic instability (collapse) is approached. In all response history analyses, a Rayleigh damping of
2.5% is assigned at the first mode period T1 and at T = 0.2T1.
For illustration, Figure 1 shows median peak story drift ratios and floor accelerations for the 20-story
2006 IBC SMRF building, under five subsets of ground motions described in Section 3.2.1. These
responses are essentially identical to those obtained using a more refined analytical model in which all
moment frame bays were modeled individually, which serves as a validation of the single-bay frame
model used in this study.

3. VULNERABILITY FUNCTION DEVELOPMENT METHODOLOGY

Vulnerability functions are developed in this study using a simulation procedure based on the PEER
loss analysis framework [6, 7]. In the PEER framework, the probability of exceedance of a loss
measure (referred to as decision variable, DV) is estimated as follows:

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

(a) (b)

Figure 1. (a) Medians of peak story drifts and (b) peak floor accelerations at five performance levels for the
20-story 2006 IBC design. The ground motions are selected based on the uniform hazard spectra at the Civic Center
in Los Angeles at five different return periods, namely, 25 years (SLE-25), 43 years (SLE-43), 475 years (design
basis earthquake), 2475 years (MCE), and 4975 years Over design earthquake (OVE), as described in Section 3.2.1.

Z Z Z
PðDV > dvÞ ¼ GðdvjDMÞjdGðDM jEDPÞjjdGðEDPjIM ÞjjdGðIM Þj (1)

where P(DV > dv) is the probability of exceedance of the decision variable; G(dv|DM) denotes the
probability of exceedance of the decision variable given a damage measure (DM) (i.e., a damage state
associated with a specific repair action, such as excessive cracking of concrete in columns that requires
epoxy injection); dG(DM|EDP) is the derivative of the probability of exceedance of the damage
measure given an EDP (e.g., story drift ratio, peak floor acceleration); dG(EDP|IM) is the derivative of
the probability of exceedance of the EDP given an intensity measure (IM) (e.g., spectral acceleration);
and dG(IM) is the derivative of the probability of exceedance of the intensity measure.
Often, the spectral acceleration at a structure’s fundamental period, denoted Sa(T1), is chosen as the
intensity measure. This is generally a good choice when the structural performance is reasonably
elastic and dominated by the first mode, but is not the best choice when estimating the losses for a
tall building whose performance is nonlinear and determined by multiple frequencies.1
Consequently, in this study, a vector of intensity measures, comprised of spectral accelerations at
multiple periods, is used in place of a single intensity measure. The corresponding loss assessment
equation is shown below.

Z Z Z
PðDV > dvÞ ¼ GðdvjDMÞjdGðDMjEDPÞjjdGðEDPjSa ðTÞjjdGðSa ðTÞj (2)

where Sa(T) is a collection of spectral accelerations at multiple periods. The choice of periods is structure-
specific and does not necessarily correspond to the modal periods of the structure. The approach used for
selecting the periods is described subsequently in this manuscript.
The objective of this study is to develop vulnerability functions for tall buildings that provide estimates
of the damage ratio (ratio of loss to building value) subject to ground motions of various intensities. In
particular, the mean vulnerability function provides estimates of the expected value of DV (i.e., damage
f (where IM
ratio) given IM f does not have to match the IM used in Equation (2)). The choice of IM f used
in this study is the spectral acceleration (equivalently, the spectral displacement) at a reference period
such as the structure’s code-based period. Note that although the structural response is predicted based
on a vector of spectral accelerations, the final vulnerability function uses a single intensity measure for
simplicity of use. It is also possible to develop a vulnerability function based on a vector of intensity
measures, but this is not done in this study.
The reference period, denoted T0 in this manuscript, is not necessarily the same as the fundamental
period of the analytical model of the structure under consideration. Real buildings may not necessarily
have the period of the considered analytical model, and hence, it is more appropriate to use a
1
This is true unless the computation of the responses is carried out using ground-motion records that not only have the
correct intensity at T1, but the ‘correct’ intensity at all other periods, as discussed by Baker [8].

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

standardized code period for reference. Furthermore, this is reasonable while assessing losses to building
portfolios with diverse sets of buildings. Shome and Luco [9] provide some additional discussion on the
choice of the reference period.
On the basis of Equation (2), the mean vulnerability function (mean damage ratio at Sa(T0)) can be
estimated as follows:
Z Z Z Z
EðDV jSa ðT0 ÞÞ ¼ DV jdGðDV jDM ÞjjdGðDM jEDPÞjjdGðEDPjSa ðTÞÞjjdGðSa ðTÞÞjSa ðT0 Þj
(3)

The variance of the vulnerability function can be estimated as follows:

VarðDV jSa ðT0 ÞÞ


Z Z Z Z
¼ ðDV

E ðDV jSa ðT0 ÞÞ2 jdGðDV jDM ÞjjdGðDM jEDPÞjjdGðEDPjSa ðTÞÞjjdGðSa ðTÞÞjSa ðT0 Þj
(4)

The above integrals can be evaluated using numerical integration under several simplified
conditions, for example, Ref. [10]. When a vector of intensity measures is used however, numerical
integration becomes infeasible, particularly when the intensity measure vector is of a large
dimension. Therefore, in this study, Monte Carlo Simulation (MCS) is used to evaluate the integrals
shown in Equations (3) and (4).
The MCS approach involves simulating all the random variables in Equations (3) and (4) (Sa(T),
EDP, DM, DV) and then computing the mean and the variance of DV for a wide range of Sa(T0)
values. The steps involved in the MCS approach are listed and discussed in detail below.
• Step 1: Simulate realizations of Sa(T) conditioned on a specific Sa(T0) using dG(Sa(T)|Sa(T0)).
• Step 2: Simulate realizations of EDP corresponding to the Sa(T)s simulated in Step 1 using dG
(EDP|Sa(T)).
• Step 3: Simulate DMs corresponding to the EDPs using dG(DM|EDP).
• Step 4. Adjust the DMs accounting for structural collapse and demolition.
• Step 5: Simulate DVs corresponding to the DMs using dG(DV|DM).
• Step 6: Repeat steps 1–5 for a wide range of Sa(T0) values to obtain the vulnerability function.
The MCS approach provides flexibility and allows the consideration of more accurate but complex
models for EDP, DM, and DV. It also allows for the rigorous consideration of collapse and demolition,
as described subsequently.
The following sections describe the steps in more detail, and show illustrative results for the
20-story, 2006 IBC steel moment frame building.

3.1. Step 1. Simulation of Sa(T) conditioned on a specified Sa(T0) using dG(Sa(T)|Sa(T0))


3.1.1. Simulation of Sa(T) conditioned on Sa(T0) for a given earthquake scenario and location. This
section describes the approach that is used for simulating Sa(T) conditioned on Sa(T0) for a given
earthquake scenario (specific magnitude, distance, etc.) and location (specific local site condition).
For a given scenario and location, the univariate distribution of Sa(T) is lognormal, and its
parameters are obtained directly from ground-motion models, for example [11]. The ground-motion
models take the following form:


lnSa ðT Þ ¼ lnSa ðT Þ þ sðT ÞeðT Þ (5)

where lnSa(T) denotes the logarithmic spectral acceleration at period T, lnSa ðT Þ denotes the mean
predicted logarithmic spectral acceleration, which depends on parameters, such as magnitude,
distance, and local site conditions; e(T) denotes the normalized (total) residual and s(T) denotes the

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

logarithmic standard deviation that is estimated as part of the ground-motion model. The ground-
motion models are fitted using large databases of recorded ground motions (such as the PEER Next
Generation Attenuation database). It is important to note that e varies as a function of period [13],
and the properties of the joint distribution of e (equivalently of the logarithmic spectral
accelerations) at multiple periods are described below.
Jayaram and Baker [12] studied the joint distribution of spectral accelerations using recorded time
histories, and concluded that it is reasonable to assume a multivariate normal distribution for
logarithmic spectral accelerations (denoted lnSa(T)) at multiple periods at a given location during a
given earthquake. On the basis of the properties of the multivariate normal distribution, it can be
concluded that the distribution of lnSa(T) (say, [lnSa(Ta), lnSa(Tb), ⋯, lnSa(Tn)], where Ta, Tb, ⋯,Tn
do not necessarily correspond to modal periods) conditioned on lnSa(T0) during a given earthquake
will also follow a multivariate normal distribution with the following mean (m) and covariance (Σ)
matrices:
2   3
m ln S a ðTa Þ þ r Ta; T0 eðT0 Þs ln S a ðTa Þ
6 m ln S ðT Þ þ rðTb ; T0 ÞeðT0 Þs ln S a ðTb Þ 7
6 a b 7
m¼6 6 : 7
7 (6)
4 : 5
m lnSa ðTn Þ þ rðTn ; T0 ÞeðT0 Þs lnSa ðTn Þ
 
where eðT0 Þ ¼ lnSa ðT0 Þ  lnSa ðT0 Þ =sðT0 Þ, and r(Ti,Tj) denotes the correlation between lnSa(Ti )
and lnSa(Tj) provided by, for instance, Baker and Jayaram [13]. It is important to note that the
unconditional marginal distribution of lnSa(T0) is completely specified by the ground-motion

equation through lnSa ðT0 Þ and s(T0). The equations discussed in this section are intended at
deriving the distribution of [lnSa(Ta), lnSa(Tb),. . ., lnSa(Tn)] conditioned on lnSa(T0).
Let Σ denote the covariance matrix of [lnSa(Ta), lnSa(Tb),. . ., lnSa(Tn)] conditioned on lnSa(T0). Let
Σ0 denote the unconditional covariance matrix of [lnSa(Ta), lnSa(Tb),. . ., lnSa(Tn)] and Σ1 denote the
covariance between [lnSa(Ta), lnSa(Tb),. . ., lnSa(Tn)] and lnSa(T0).
2 3
s2ln Sa ðTa Þ rðTa ; Tb Þs ln Sa ðTa Þ s ln Sa ðTb Þ . . . rðTa ; Tn Þs ln Sa ðTa Þ s ln Sa ðTn Þ
X 6 7
6 rðTb ; Ta Þs ln Sa ðTb Þ s ln Sa ðTa Þ s2ln Sa ðTb Þ . . . rðTb ; Tn Þs ln Sa ðTb Þ s ln Sa ðTn Þ 7
¼6 7
0 4 : : ... : 5
rðTn ; Ta Þs ln Sa ðTn Þ s ln Sa ðTa Þ rðTn ; Tb Þs ln Sa ðTn Þ s ln Sa ðTb Þ ... s ln Sa ðTn Þ
2

(7)

2 3
rðTa ; T0 Þs ln Sa ðTa Þ s ln Sa ðT0 Þ
X 6 rðTb ; T0 Þs ln Sa ðTb Þ s ln Sa ðT0 Þ 7
6 7
¼ 6 : 7 (8)
6 7
1
4 : 5
rðTn ; T0 Þs ln Sa ðTn Þ s ln Sa ðT0 Þ

Σ can be computed as follows [14]:

X X 1 X X’
¼  (9)
0 s2ln Sa ðT0 Þ 1 1

where Σ’1 denotes the transpose of Σ1.


Figure 2, for instance, shows the exponential of the conditional mean, exp[m], and the conditional
standard deviation (square root of the diagonals of Σ ) of Sa(T) at 20 different periods ranging from
0 (i.e., PGA) to 10 s for a magnitude 7 earthquake at a distance of 10 km corresponding to a Vs,30
(average shear wave velocity in the top 30 m of the soil, often used in ground-motion models as an
indicator of local site conditions) of 360 m/s. The Sa(T) values are conditioned on a spectral

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

Figure 2. (a) Exponential of the conditional mean of lnSa(T), (b) the conditional standard deviation of lnSa
(T). The mean and the standard deviation are conditioned on Sa(T0) = 0.22g at T0 = 2.5 s.

acceleration of 0.22g at 2.5 s (the reference period T0 of the 20-story SMRF building), which
corresponds to an e(T0) of 1.0. The unconditional mean and the standard deviation values, m lnSa ðT Þ
and s lnSa ðT Þ respectively, are obtained using the Boore and Atkinson [11] ground-motion model.
For any given event and location, Sa(T) values can now be simulated conditioned on Sa(T0) by sampling
from a multivariate normal distribution with mean m and covariance Σ. Figure 3 shows 100 Monte Carlo
realizations of the conditional Sa(T) values for the earthquake scenario described earlier. The values of T at
which Sa(T) has been computed are 0, 0.5, 1,. . .,9.5, and 10 s.
3.1.2. Simulation of Sa(T) conditioned on Sa(T0) considering multiple earthquake scenarios. The
previous section described the simulation of Sa(T) conditioned on Sa(T0) for a given earthquake
scenario. In general however, a given Sa(T0) is caused by several possible scenarios (i.e., M–R
combinations). This section describes the procedure used for sampling Sa(T) conditioned on Sa(T0)
considering multiple M–R scenarios. To do that, consider the following expression for the distribution
dG(Sa(T)|Sa(T0)):

Z
dGðSa ðTÞjSa ðT0 ÞÞ ¼ dGðSa ðTÞjSa ðT0 Þ; M; RÞdGðM; RjSa ðT0 ÞÞ (10)
M;R

Figure 3. 100 Monte Carlo simulations of Sa(T) conditioned on Sa(T0) = 0.22g.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

It is inferred from Equation (10) that Sa(T) can be simulated by first simulating M and R conditioned
on Sa(T0) and then by simulating Sa(T) conditioned on Sa(T0), M and R as described in Section 3.1.1.
The distribution dG(M, R|Sa(T0)), which is required for simulating M and R conditioned on Sa(T0), can
be obtained by hazard deaggregation for the site of interest and is provided by the United States
Geological Survey (USGS) [15]. The simulation procedure first involves sampling M and R pairs in
proportion to the likelihoods obtained using deaggregation, and subsequently simulating Sa(T)
conditioned on Sa(T0), M and R as described in Section 3.1.1. In this study, the hazard
deaggregation is carried out at the civic center in Los Angeles, the same location used in the PEER
TBI. It is to be noted that the vulnerability function calculations using hazard deaggregation results
at the Civic Center in San Francisco, using the Vs,30 values typical of San Francisco, did not result
in considerably different vulnerability functions.
In summary, Sections 3.1.1 and 3.1.2 focus on identifying appropriate M–R pairs based on the
deaggregation performed by USGS (using all relevant ground-motion models) and sampling response
spectra corresponding to these pairs. The simulation of the response spectra for any scenario uses
information about the spectrum mean and uncertainty for that scenario as quantified by the ground-
motion models. The simulation process essentially involves sampling from the multivariate Normally
distributed e(T)s. Overall, the process helps determine the appropriate shape of the ground-motion
spectrum (i.e., the correct values of Sa(T) for a given Sa(T0)) to be expected at the site for the
multifrequency dominated tall buildings (where T0 is just a reference period, which can be the
code-based fundamental period or any other reasonable period).

3.2. Step 2. Simulate realizations of EDP conditioned on noncollapse corresponding to the Sa(T)s
simulated in Step 1 using dG(EDP|Sa(T))
The approach used in this work to simulate EDP given Sa(T) is similar to that used by the PEER Ground
Motion Selection and Modification (GMSM) group while evaluating various ground motion selection
methodologies [16] and by Shome and Bazzurro [2] as part of their tall building loss assessment
methodology. As part of this approach, regression equations are developed to predict the EDPs as
functions of Sa(T), and are subsequently used to predict the EDPs for the simulated Sa(T)s. The
procedure is described in detail in the next section.
In theory, it is possible to evaluate dG(EDP|Sa(T)) without using any regression. This can be carried out
by first selecting ground motions corresponding to every simulated Sa(T) and performing response history
analyses to derive dG(EDP|Sa(T)). The primary advantage of the regression approach is to reduce the
computational demand that would be incurred if we were to analyze the building performance under
time histories matching all the Sa(T)s simulated earlier. In fact, it might even be difficult to select time
histories that match all the simulated Sa(T)s. Furthermore, the PEER TBI buildings were analyzed by
PEER using a set of 75 ground motions (described in the next section) and the proposed regression
allows us to utilize the results of these 75 analyses to estimate EDP for the Sa(T)s simulated in this study.
3.2.1. Estimate the EDPs corresponding to a selected training set of 75 ground motions using
OPENSEES. The 75 ground motions used for the response history analyses are those selected by
PEER for the TBI [3]. The ground motions are selected based on the uniform hazard spectra at
the Civic Center in Los Angeles at five different return periods, namely, 25 years (SLE-25), 43 years
(SLE-43), 475 years (design basis earthquake), 2475 years (maximum considered earthquake, MCE),
and 4975 years (OVE). It is to be noted that these ground motions may be conservative because they
have been chosen based on uniform hazard spectra for the selected return periods [8]. As discussed by
Baker [8], this is because the uniform hazard spectrum is a compilation of rare spectral accelerations at
multiple periods, which is unlikely to be seen in real ground motions. Although the EDPs are computed
for these records, it is to be noted that the purpose of computing these EDPs is primarily to develop
regression models to predict EDPs as functions of Sa(T). The use of the appropriate distribution for
Sa(T) conditioned on Sa(T0) as described in Section 3.1 will ensure that realistic EDPs are used while
developing the vulnerability function.
The EDPs of interest in this study are the story drift ratio (SDR) at each story and the peak floor
acceleration (PFA) in each floor and the maximum residual story drift ratio (ResDR) over all the
stories. Only the peak absolute values of these parameters with time are considered. The EDPs

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

mentioned here are all necessary for a robust assessment of building losses. The SDR values are used to
predict the damage to structural subsystems and nonstructural drift-sensitive (NSD) subsystems in each
story, for example, Refs. [17, 18]. The PFA values are used to predict the damage to nonstructural
acceleration-sensitive (NSA) subsystems on the floor at the bottom of each story. Recent research by
Ramirez and Miranda [18] discussed the significance of the maximum ResDR values in determining
whether a structure would have to be demolished after the earthquake.
An important note is that the building loss assessment carried out in this study is ‘story based’ [18].
Here, the SDR in each story is used to predict the damage measure of the corresponding story structural
and NSD subsystems, and the PFA in each floor is used to predict the damage measure of the
corresponding story NSA subsystems. Once all the story losses are computed for each simulation, they
are summed to obtain the total building loss.
The story loss concept is not implemented for consideration of collapse and demolition. Total loss is
assumed if any one story experiences dynamic instability or the maximum residual drift exceeds a
threshold given by the fragility curve that relates maximum residual drift to demolition (discussed in
Section 3.4).
3.2.2. Develop regression equations to predict EDPs as functions of Sa(T). For each of the 75 ground
motions, spectral accelerations are computed at 21 periods from 0 to 10 s in multiples of 0.5 s. The use of
the 0.5 s interval results in spectral accelerations that do not have a strong correlation between each other,
thereby avoiding multicollinearity issues in the regression. Motivated by a regression fit of the type
lnEDP = lna + blnSa(T) [10] and the functional form chosen in the PEER GMSM study [16], a set of
available predictors (X) for the EDPs is then created as follows:
h i
X ¼ lnSa ð0:0Þ lnSa ð0:5Þ⋯lnSa ð10:0Þ ðlnSa ð0:0ÞÞ2 ðlnSa ð0:5ÞÞ2 ⋯ðlnSa ð10:0ÞÞ2 (11)

The parameter to be predicted, Y, is the logarithm of the EDP (denoted lnEDP). Not all the predictors in
X are used in the regression for Y. In other words, although X provides the entire set of available predictors,
only a subset of X is used for predicting each Y. This is carried out to primarily avoid over-fitting the EDPs
and reduce the possibility of selecting a cluster of highly-correlated predictors that can lead to
multicollinearity issues as discussed in [19]. Second, the benefit of using an additional predictor
typically reduces with increase in the number of predictors. The ideal number of predictors used is the
one that results in a small mean squared error (or an alternate measure of the fit error) that does not
reduce much with the addition of an extra predictor.
The subsets of predictors used in the regressions for the EDPs are chosen based on the exhaustive subset
selection approach [19]. In this approach, for each EDP, 10 (adopted based on a sensitivity study)
predictors are chosen out of the 42 available predictors from X that predict the EDP with the least mean
squared error (MSE). In other words, the MSE is computed using all possible combinations of 10
predictors of the available 42, and the combination that results in the smallest MSE is selected.
Although the initial 42 periods are the same for any building considered in the study, the final set of 10
predictors will be structure specific. The automatic selection of the predictors using the subset selection
approach eliminates the need to manually identify predictors or choose spectral accelerations at the
structure’s modal periods that may not be appropriate for highly nonlinear structures. If spectral
accelerations at the modal periods are the best predictors, they will form part of the final subset because
they will minimize the MSE the most.
Figure 4 shows example comparisons of median EDPs estimated using the regression to values
obtained directly from the response history analyses. It is to be noted that this procedure is used
separately for each structure of interest in this study to estimate the relevant EDP predictors for that
structure.
It is assumed in this approach that the response spectrum, Sa(T), can effectively predict the EDP.
Past studies have explored the impact of several parameters in addition to Sa(T ), such as magnitude,
distance and duration, on the EDP, and found generally only a weak dependence on some of these
parameters [20, 21]. Hence, these additional parameters are not considered in this regression. This is
also consistent with the regressions carried out in the PEER GMSM study [16] and the research work

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

Figure 4. Predicted (continuous line) and calculated median (circles) EDP values for the 20-story 2006 IBC
SMRF building under the 2475-year return period ground motions. (a) SDR (b) PFA.

of Shome and Bazzurro [2]. It is possible that the structural response of some highly nonlinear structures
may also depend on other parameters such as duration, in which case, it will be important to ensure that the
training data set has representative duration values.
The regression equations provide both the mean logarithmic EDPs ( m ln EDP) and the standard
deviation of the logarithmic EDPs (sln EDP). Therefore, the distribution of the EDPs can be described
as follows (using the properties of classical linear regression [19]):

lnEDPeNðm ln EDP ; s lnE DP Þ (12)

where N denotes the Normal/ Gaussian distribution.


Although the above equation provides the univariate distribution of the EDPs, the EDPs need to be
jointly simulated to account for possible inter-EDP correlations. For instance, a higher than average
SDR in story 1 will likely imply a higher than average SDR in story 2. The extent of these correlations

can be computed using the regression residuals. Let ln EDPij and ln EDPij (= m lnE DPij ) denote the
computed (using response history analysis) and the predicted (using the regression) values of the ith
logarithmic EDP corresponding to the jth ground motion. The residual corresponding to lnEDPij can

then be estimated as ~eij ¼ lnEDPij  ln EDPij . The covariance between ln EDPij and lnEDPi0 j can
then be calculated as the covariance between ~eij and ~ei0 j . In the study, we consider the covariances
between lnSDR’s in two different stories, the lnPFA’s in two different floors and the cross-covariances
between all possible pairs of lnSDR in story m, lnPFA in floor n and ln(maximum ResDR) (where m
varies from 1 to number of stories and n varies from 1 to number of floors).
Figure 5 shows, for illustration, the correlations between the lnSDR’s and the cross-correlations
between the lnSDRs and the lnPFAs. We found that the correlations are relatively constant across
ground-motion intensity levels. On the basis of this observation combined with the fact that the
structure has only been analyzed using a reasonably small number of ground motions (75), we do not
model the correlations (a matrix of K by K correlation values for each EDP type, where K is the
number of stories) as a function of intensity, because that will not result in robust correlation predictions.
Denoting the vector of all the logarithmic EDPs of interest by lnEDP = [lnEDP1, lnEDP2, . . . ,
lnEDP0], the mean (mln EDP) from the regression equation and the covariance (Σ lnEDP) computed using
the residuals can be used to define the joint distribution of EDP, namely, N(m ln EDP, Σ ln EDP), where N
denotes the multivariate Normal distribution. The EDPs can then be simulated using the multivariate
normal distribution. These simulations should be performed for all the Sa(T) vectors simulated in Step 1.
3.2.3. Accounting for epistemic uncertainties in the engineering demand parameters. As mentioned
earlier, an EDP can be modeled as lnEDP ~ N(mln EDP, sln EDP). sln EDP is the standard deviation of the

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

Figure 5. (a) Correlation between lnSDR’s in two different stories (b) cross-correlation between lnSDRs and
lnPFAs. The correlations are estimated using the 75 PEER ground motions.

aleatory uncertainty variable corresponding to the EDP at a given Sa(T0) (i.e., the uncertainty around the
regression prediction). Apart from aleatory uncertainties, there are also various sources of epistemic
uncertainty (i.e., the uncertainties that arise out of lack of knowledge) in the EDP, including,
uncertainties because of modeling, software and analysis procedures, damping, and material properties.
Incorporating these epistemic uncertainties, lnEDP can be modeled as follows:

lnEDPij ¼ m ln EDPij þ ~eij þ ij (13)

where m ln EDPij is the mean logarithmic value of the ith EDP during the jth ground motion predicted by the
regression, ~eij is the aleatory uncertainty estimated from the regression (as described earlier), which is a
Normal random variable with zero mean and standard deviation s ln EDPij , and ij is the epistemic
uncertainty term, which is a Normal random variable with zero mean and standard deviation t lnEDPij .
The extent of epistemic uncertainty as reflected by t ln EDPij can either be measured on the field, obtained
from past research studies that have estimated the extent or using engineering judgment. This study
uses a combination of the last two sources mentioned to obtain representative values that are reasonable
while developing fragility functions to be used for a large number of buildings.
After simulating the EDPs as described, the epistemic uncertainty term ij is also simulated from a
univariate normal distribution. It is assumed that the value of ij is the same (i.e., perfect correlation) for
a particular EDP type across all the stories of the building (e.g., SDRs at all the stories). The ij values
corresponding to two different types of EDPs are assumed to be independent. The normality
assumption for epistemic uncertainty is based on standard convention, and is usually justified by the
central limit theorem.

3.3. Step 3. Simulate DMs corresponding to the EDPs simulated in Step 2 using dG(DM|EDP)
The DM refers to the potential damage state of the structure/subsystem after an earthquake. In this study,
three subsystems are considered in each story: (a) structural, (b) NSD, and (c) NSA. These are aggregated
subsystems. This study does not consider individual components such as beams, columns, ceilings,
partition walls, etc., one at a time.
The damage measures considered in this discussion are no damage, minor damage, moderate damage,
extensive damage, and complete damage [22]. The damage measures defined in newer research works
such as the ATC-58 [23] project are more informative and will be explored in future work. The
probability of a subsystem attaining any of these damage measures for a given EDP is provided by a
‘fragility function’. A fragility function typically models the distribution of DM conditioned on EDP
as follows.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

The damage measure of a subsystem equals or exceeds i (1 to 5, corresponding to no damage,


minor damage, moderate damage, extensive damage and complete damage respectively, or more
informatively-defined damage measures such as those proposed in ATC-58 [23]) if the EDP exceeds
a certain threshold, which is modeled as a lognormal random variable with a specified mean and
variance.

 
ln EDPijk e N mijk ; bijk (14)

where lnEDPijk is the threshold logarithmic EDP that causes at least the damage measure i (1–5) for
subsystem j (1: structural, 2: NSD, 3: NSA) located in the kth story, m ijk denotes its mean and bijk its
dispersion. The epistemic uncertainty in lnEDPijk is handled using the approach employed for modeling
the epistemic uncertainty in the EDP (Section 3.2.3).
Several past professional and research projects have used experimental data and observed damage
levels to develop fragility functions for both structural and nonstructural subsystems. The fragility
functions used in the results presented later come from a variety of sources including HAZUS [22],
Ramirez and Miranda [18], ATC-58 [23], FEMA 355F [24] and FEMA-273 [25]. In particular, the
structural component fragilities are directly obtained from ATC-58 [23] for steel structures and
FEMA-273 [25] for concrete structures. The medians for the nonstructural components are obtained by
equally weighting the medians from HAZUS [22] and Ramirez and Miranda [18]. The dispersion of the
epistemic uncertainty term for the structural subsystems is assumed to be 0.25, as recommended by
FEMA 355F [24], while the aleatory uncertainty dispersion equals 0.35 [23]. For nonstructural
subsystems, the total dispersion of 0.5 from HAZUS [22] is split into aleatory and epistemic
uncertainties in the proportion of 2 : 1 (assumed based on engineering judgment). More advanced data
on fragilities are being developed in the ATC-58 project but are not utilized here because the ATC-58
fragility functions are not yet in the public domain. It is to be noted that that the differentiation between
epistemic and aleatory uncertainties is still somewhat subjective. In this study, the definition of
epistemic uncertainty is tied to that component of uncertainty, which is not ergodic, as defined in
Section 4.3.
A fragility function only defines the univariate distribution of lnEDPijk. The performance of subsystems
tend to be correlated as well, and it is reasonable to expect that a higher than average damage in story 1 will
imply a higher than average damage in story 2 because of comparable quality of construction and materials
and other common factors. In other words, if a smaller than expected EDP causes a particular damage
measure in story 1, we would expect a smaller than expected EDP to cause that damage measure in
story 2 as well. Similarly, it is also reasonable to assume that EDPs that cause the exceedance of two
different damage measures for a single subsystem are correlated. In the results presented later, the
correlations between lnEDPijk and lnEDPi’j’k’ are assumed to be 0.8 when j = j’ (same subsystem type),
0.5 when j = 1 and j’ = 2 or 3, and 0.67 when j = 2 and j’ = 3. The correlations are assumed to be valid
for all combinations of k and k0 and i and i0 . This approach of focusing on the capacity term (lnEDPijk)
rather than the discrete damage measure (DM), originally proposed by Jalayer and Cornell [26] and
Baker [27], allows us to consider the correlation between the damage measure occurrences. It was
found as part of extensive sensitivity analysis calculations that small to moderate changes in these
correlation estimates do not significantly impact the vulnerability functions.
With information about the medians, dispersions and correlations, the logarithmic EDPs that cause
different damage measures to be reached or exceeded for all the subsystems can be jointly simulated by
drawing realizations from a multivariate normal distribution. The damage measure of each subsystem
can then be determined based on the predicted EDPs from the regression equations. This procedure
needs to be carried out for each of the Sa(T) vectors simulated in Step 1 (using the corresponding EDPs
predicted in Step 2).

3.4. Step 4. Adjust the DMs accounting for structural collapse and demolition
During an earthquake, there is a possibility of collapse of the structure or of excessive residual drifts
necessitating demolition. Two different types of structural collapses are considered in this study. The
first type of collapse, simulated collapse, is modeled using a collapse probability curve, which provides

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

the probability of collapse of the structure as a function of Sa(T0). The collapse probability curve is derived
based on the collapse occurrences because of excessive lateral drifts and deterioration leading to a loss of
the gravity load resistance of the structure as indicated during the OPENSEES analyses using the 75 ground
motions. The collapse probability curve is developed using logistic regression, which is commonly used
while regressing probabilities [19]. On the basis of the logistic regression approach, if p denotes the
probability of collapse,
 
p
ln ¼ b0 þ b1 S a ðT0 Þ (15)
1p

It is possible to use a vector of spectral accelerations here, but we found that Sa(T0) was
sufficient in the cases considered here. For illustration, the collapse probability curve obtained for
the 20-story 2006 IBC steel moment frame building is shown in Figure 6. For each of the
simulated Sa(T) vectors, collapse of a structure is determined by drawing a uniform random
variable between 0 and 1, and determining whether the realization falls below the collapse
probability curve at the corresponding Sa(T0).
In general, it is found that the contribution of collapse to building loss is about 10% under MCE ground
motions (2475-year return period), and hence, small changes to the collapse probability curve will not
significantly affect the developed vulnerability functions, except at very large Sa(T0) values.
Assuming that the structure does not suffer simulated collapse under a particular Sa(T) vector, the
regression equations described in Section 3.2.2 are now used to estimate the EDPs for this Sa(T)
vector. In simulation, it is observed sometimes that a structure undergoes a very large drift without
collapse. This is referred to as nonsimulated collapse, which is determined to occur when the
maximum (over all stories) SDR exceeds a threshold value. The threshold is assumed to be a
lognormal random variable with a median of 7.5% (50% more than the complete damage fragility
median), an aleatory dispersion of 0.35 and an epistemic dispersion of 0.25 (identical to those for
structural subsystems).
Finally, even if a structure does not collapse during an earthquake, demolition is deemed necessary if
the maximum residual drift ratio exceeds a threshold, which is defined to be a lognormal random
variable with median 1.50% for concrete structures and 1.85% for steel structures, an aleatory
dispersion of 0.3 and an epistemic dispersion of 0.25 [18]. Ramirez and Miranda [18] originally
recommended a median of 1.50% for concrete buildings. We increased this median by 25% for steel
structures.

3.5. Step 5. Simulate DV corresponding to the DMs using dG(DV|DM)


Loss functions provide estimates of the decision variable, namely, loss costs (alternately, the loss ratio,
which is the ratio of the loss cost to the subsystem value), associated with each damage measure for

Figure 6. Collapse probability curve for the 20-story 2006 IBC SMRF building (T0 = 2.5 s).

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

each subsystem [17, 22]. Loss functions typically model the loss ratio as a lognormal random variable
with a certain median and dispersion. It is to be noted that the loss functions provide the univariate
distribution of the loss ratio for each subsystem and damage measure. Defining the multivariate
distribution of the loss ratios of multiple subsystems across multiple damage measures requires
information about the inter-loss-ratio correlations. The loss ratios for different subsystems across
different damage measures are expected to be correlated because of common factors involved in
repairs, such as the contractor hired for the repairs and the material costs. These correlations are
accounted for in the manner described in Section 3.2.2. Ramirez and Miranda [18] have compiled
information about repair costs charged by multiple contractors to estimate the extent of these
correlations. On the basis of their observations, the loss cost correlation between two subsystems of
the same type used in the simulations is 0.75. The correlation between loss costs of structural and
NSA/NSD subsystems is 0.25, and the correlation between loss costs of NSA and NSD subsystems is
0.33. The logarithmic loss ratios are simulated from a multivariate normal distribution with the above
mentioned medians, dispersions, and correlations.
The epistemic uncertainty in the loss cost is mathematically handled using the same approach that is
used for modeling the epistemic uncertainty in the EDP (Section 3.2.3). The total dispersion of the loss
cost is assumed to be 0.7, as recommended by Aslani and Miranda [17], and is split into aleatory and
epistemic dispersions in the proportion of 1 : 2.
On the basis of the simulated values of the damage measures and the loss ratios, the total loss can be
computed by summing the subsystem losses in each story (calculated as the product of the value of the
subsystem and the loss ratio). For the examples illustrated later, the total values of the structural and the
NSD subsystems are assumed to be 0.19 and 0.33 times the total building value, respectively, and are
assumed to be equally distributed across all the stories. The value of the NSA subsystems at the ground
level is assumed to be 0.1 times the building value. The remaining NSA subsystem value (0.38 times
the building value) is assumed to be equally distributed across all the stories. These assumptions are
based on those in HAZUS [22] for a commercial building. The total loss normalized by the building
value is the damage ratio for the building. This damage ratio computation needs to be performed for
each Sa(T) vector simulated in Step 1.

3.6. Step 6. Repeat the procedure for all the Sa(T0) values of interest to obtain the
vulnerability function
3.6.1. Estimate the mean and the variance of the damage ratios as functions of Sa(T0). In summary,
Steps 1 and 2 involve the selection of appropriate ground motions for a given Sa(T0) and the estimation
of the associated EDPs. The EDPs are used to identify whether the structure collapses or needs to be
demolished (Section 3.4). If the structure survives, the EDPs are used to estimate the damage measures
as discussed in Step 3. Finally, the damage information is used to estimate the corresponding damage
ratio as discussed in Section 3.5.
At the end of Step 5, we have estimated the damage ratios for the building for all the simulated Sa(T)
vectors corresponding to a specific Sa(T0) value. The entire procedure is now repeated for a wide range
of Sa(T0) values (0 to 2g, for instance). At the end of the procedure, a collection of damage ratios
corresponding to all the considered Sa(T0) values is obtained, as shown in Figure 7. Even for a given Sa
(T0) value, the damage ratios vary significantly for the Sa(T) vectors considered on account of the
uncertainty in parameters such as EDP, DS, subsystem loss and ground-motion intensity at periods
other than T0. The data points shown in Figure 7 are then binned into Sa(T0) bins, and the mean and the
coefficient of variation of the damage ratios are then computed for each bin.
For the illustrative vulnerability function calculation, a set of 5000 Sa(T) vectors is generated to quantify
the ground-motion hazard. For each of the 5000 simulations, 25 additional simulations of EDPs, damage
measures, and loss costs are performed. For each of these 25 simulations, the epistemic uncertainty term is
kept constant across the 5000 simulations, while the aleatory uncertainty terms are varied (mimicking
the behavior that the epistemic uncertainty variable does not change from one earthquake to another,
while the aleatory term does). This approach helps quantify both the epistemic and the aleatory
uncertainties in the developed vulnerability function.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

Figure 7. Damage ratios for the 20 story 2006 IBC SMRF for all the simulated Sa(T) vectors corresponding
to a wide range of Sa(T0) values.

The computational time required for estimating the damage ratios shown in Figure 7 is approximately
30 s on a 2.83 GHz, 4 GB RAM quad core processor.

3.6.2. Account for the impact of ground-motion uncertainty on the vulnerability function. When
estimating potential building losses during future earthquake events, it is first necessary to estimate a Sa
(T0) value and then obtain the corresponding damage ratio mean (MDR) and variance (VDR). There
exists considerable uncertainty in the Sa(T0) value even for a given earthquake and site, and it will be
incorrect to use a deterministic estimate of Sa(T0), such as a median value, to obtain MDR and VDR
from the developed vulnerability function. It is correct, however, to simulate a set of possible Sa(T0)
values for the earthquake and site in consideration, and estimate the MDRs and VDRs corresponding to
these Sa(T0) values. Alternately, it is possible to directly modify the vulnerability function mean and
variance so that it will be correct to use the median Sa(T0) to estimate these parameters. The
mathematical basis for doing this is as follows:

 
ð
MDR Sa ðT0 Þ ¼ E DRj Sa ðT0 Þ Þ

¼ E ð
E DRj Sa ðT0 Þ; Sa ðT0 Þ
Sa ðT0 Þj Sa ðT0 Þ Þ Using the tower property
¼ E EðDRjSa ðT0 ÞÞ
(16)
Sa ðT0 Þj Sa ðT0 Þ

¼E  MDRðSa ðT0 ÞÞ
Sa ðT0 Þj Sa ðT0 Þ 
R 
¼ MDRðSa ðT0 ÞÞdF Sa ðT0 Þj Sa ðT0 Þ


where Sa ðT0 Þ is the median Sa(T0), DR denotes damage ratio, E(.) denotes the expectation operator, MDR

(Sa(T0)) denotes the mean damage ratio at Sa(T0), dF ðSa ðT0 Þj Sa ðT0 ÞÞ is the derivative of the cumulative
 
distribution function of Sa(T0) conditioned on Sa ðT0 Þ. From Equation (5), lnSa ðT0 Þj lnSa ðT0 Þ follows

a Gaussian (Normal) distribution with mean lnSa ðT0 Þ and standard deviation s(T0). Using this
information, the integral in Equation (16) can beevaluated using numerical integration.

In essence, Equation (16) shows that MDR Sa ðT0 Þ can be computed by accounting for all the

probable Sa(T0) and the corresponding MDR values associated with the Sa ðT0 Þ . The MDR values
corresponding to the Sa(T0)s are then multiplied by the likelihood of observing the Sa(T0)s given the
 
Sa ðT0 Þ value and summed
  It is now appropriate to use Sa ðT0 Þ to estimate the expected damage
up.
ratio using the MDR Sa ðT0 Þ curve. A similar approach can also be used to estimate the variance of
the vulnerability function with consideration of the ground-motion uncertainty.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

  
VDR Sa ðT0 Þ ¼ Var DRj Sa ðT0 Þ
 
¼E Var DRjSa ðT0 Þ; Sa ðT0 Þ
Sa ðT0 Þj SaðT Þ
0
 
þVar E DRjSa ðT0 Þ; Sa ðT0 Þ (17)
Sa ðT0 Þj SaðT Þ
R 0  
¼ VDRðSa ðT0 ÞÞdF Sa ðT0 Þj SaðT0 Þ
Rh i2  

þ MDRðSa ðT0 ÞÞ  MDR Sa ðT0 Þ dF Sa ðT0 Þ SaðT0 Þ

where VDR(Sa(T0)) denotes the variance of the vulnerability function at Sa(T0) and Var(.) denotes the
variance operator.

4. LOSS RESULTS AND DISCUSSION

4.1. Loss assessment methodology for a single building


In this study, loss assessments are carried out for individual buildings under a stochastic catalog of
earthquakes. The stochastic catalog used in this study is a large set of earthquakes of different
magnitudes occurring on different faults that could affect the site of interest (assumed to be Los
Angeles) and whose annual rates are determined using seismicity information (e.g., magnitude–
recurrence relationship) from the USGS [28]. The following steps summarize the loss assessment
methodology for individual buildings:
• For each event in the catalog, generate the median ground-motion intensity at the site corresponding
to the rupture scenario of the considered earthquake [29]. The intensity used is the (weighted) mean
of the median intensities from all the ground-motion models used by USGS [28] while generating
their hazard maps.
• Use the median intensities to predict the MDR and the VDR of the damage ratio using the functions
developed in Section 3.6.2 for each earthquake in the catalog.
• Assume that the loss at the site for any event i (denoted Li) in the catalog follows a Beta distribution
defined by the mean (MDR multiplied by the total building value) and the variance (VDR multiplied
by the square of the total building value) estimated above.
• The average annual loss (AAL) and the loss exceedance rates (ER) are now computed as follows
(using the Poissonian assumption for event occurrences) [e.g., 29]:

X
N
AAL ¼ ni L i (18)
i¼1

PN
n PðL i >lÞ
ERðlÞ ¼ 1  e i¼1 i (19)
where ni is the annual rate of occurrence of event i, N is the total number of events in the catalog, l is the
loss value whose exceedance rate is being calculated and P(Li > l) is the probability that Li exceeds l,
which is calculated using the Beta cumulative distribution function.

4.2. Loss assessment results for a single building


Vulnerability functions based on the methodology discussed in Section 3, combined with the loss
assessment methodology summarized in Section 4.1, are used to perform loss assessments for
individual tall buildings. The assumed location is the civic center in Los Angeles. Figure 8 shows
the AAL and the 500-year exceedance losses obtained for the six buildings summarized in Section
2. Error bars shown in the figures indicate the plus-minus one standard deviation bounds of the loss
estimates corresponding to the epistemic uncertainty in the developed vulnerability function. The
procedure used for computing these bounds is described in the next section.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

Figure 8. (a) AAL and (b) 500-year return period loss.

4.3. Impact of epistemic uncertainties in ground-motion hazard and building vulnerability on


loss estimates
Both aleatory and epistemic uncertainties in ground-motion hazard, building performance (EDPs),
damage measures and loss costs are accounted for in the development of vulnerability functions.
When evaluating the loss of a single building, most epistemic uncertainties are nonergodic, meaning
that the values of the epistemic random variables do not change every time an earthquake occurs.
For instance, if a building performs better than an average building of the same type, its
performance will exceed the average performance every time an earthquake occurs, ignoring factors
such as deterioration. Therefore, it is not reasonable to assume that the time-average (equivalently,
the average over all earthquake occurrences) of an epistemic random variable, such as the deviation
of a building performance from the average performance, will equal zero. As a result, if a building
always performs better than average, its AAL and return period losses will be lower than those for a
corresponding average building. Similarly, there might be a different building whose performance is
always below average. Thus, the epistemic uncertainties introduce variability in the AAL and the
return period losses (in addition to impacting the mean values of these parameters). This section
quantifies this variability for the IBC 2006 20-story SMRF building for illustration.
First, to estimate the impact of epistemic uncertainties in the ground-motion hazard, several hazard
curves are sampled for the civic center in Los Angeles, factoring in the uncertainty in parameters such
as source rates, ground-motion models used, etc. [30]. The vulnerability function is then integrated with
the sampled hazard curves (Figure 9(a)) to compute the AAL and the return period losses (as described
in Equations (18) and (19)). Subsequently, the mean of the above hazard curves is used in conjunction
with several vulnerability functions, each obtained by sampling epistemic uncertainty terms related to

Figure 9. (a) Sampled hazard curves for spectral acceleration at 2.5 s. (b) Sampled vulnerability functions for
the IBC 2006 20-story SMRF building considering the epistemic uncertainty.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

Figure 10. Impact of epistemic uncertainty in hazard and vulnerability on (a) AAL and (b) 500-year return
period loss. These results are for the IBC 2006 20 story SMRF building.

vulnerability (e.g., EDP, loss cost) several times (Figure 9(b)). This is carried out to quantify the impact of
epistemic uncertainty in vulnerability on the variability in the loss results.
Finally, the AAL and the return period losses are derived considering the epistemic uncertainties in both
hazard and vulnerability. Figure 10 shows the normalized mean plus–minus one standard deviation of the
AAL and the 500-year return period loss for illustration. There is significant variance in the loss results,
and the standard deviations are approximately 30% of the mean value when the epistemic uncertainties
in both hazard and vulnerability are considered.

4.4. Future work: loss assessment for a building portfolio


In future studies, we intend to assess risks to tall building portfolios in San Francisco and Los Angeles
using vulnerability functions for all relevant construction classes developed based on the methodology
presented in the manuscript. An additional challenge in portfolio loss assessment relative to the loss
assessment of an individual building is the correlation between losses to buildings at two different
locations, which needs to be quantified to obtain accurate portfolio loss exceedance probabilities.
Rigorous methods for computing the spatial correlation and the portfolio losses considering spatial
correlation, such as that proposed by Jayaram and Baker [31], are available in the literature. It is
intended to use these methods to compute the portfolio losses under stochastic earthquake catalogs and
historical/scenario earthquakes.

5. SUMMARY AND CONCLUSIONS

This study focused on the development of vulnerability functions for tall buildings. A systematic
simulation procedure based on the PEER loss assessment framework was used to develop building
vulnerability functions that provide estimates of losses to buildings subject to ground motions of
various intensities. The steps involved in the procedure are quantifying ground-motion hazard using a
vector of spectral accelerations, predicting building response parameters such as story drifts, floor
accelerations, and residual drifts under the quantified hazard, accounting for structural collapse and
demolition, and predicting story-wise losses and total building loss using the building response
information. A total of six tall buildings, four of them designed by the authors, and two designed as
part of the PEER TBI program, were evaluated in this study. These buildings are 20-story and 40-story
steel moment frame buildings designed based on the 2006 IBC and the 1973 Uniform Building Code, a
42-story concrete core wall building and a 42-story concrete dual system building. The performance of
each of these buildings was studied under a set of 75 ground motions expected in the Los Angeles
region. Regression equations were then developed separately for each building using these training sets
to predict the necessary structural response parameters given a vector of ground-motion intensity
measures. Finally, a modified version of the PEER performance assessment approach was used to

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
EARTHQUAKE VULNERABILITY FUNCTIONS FOR TALL BUILDINGS

develop vulnerability functions that provide estimates of losses to these buildings subject to ground
motions of various intensities.
The developed vulnerability functions were used for the story-based loss assessments of six tall
buildings assumed to be located in Los Angeles, using a stochastic catalog of probable future
earthquakes. For the 2006 designs, the average annual losses vary from 0.2% to 0.26% of the building
value and the 500-year return period losses vary from 20% to 27% of the building value. For the 1973
steel moment frame designs, the AAL and the 500-year return period losses are about 40% higher than
for the 2006 steel moment frames. Throughout this study, emphasis was placed on capturing various
epistemic and aleatory uncertainties in ground motions, building performance, loss costs, etc. It was
seen that the epistemic uncertainties in both ground-motion hazard and building vulnerability have
significant impacts on the loss assessment results. The variability in the loss assessment results because
of this epistemic uncertainty is large enough that the standard deviations of the loss results are close to
30% of the mean value.
In future studies, we intend to assess risks to tall building portfolios in San Francisco and Los
Angeles using vulnerability functions of the type presented in this paper.

ACKNOWLEDGEMENTS
The authors thank Patricia Grossi and the two anonymous reviewers for their helpful reviews of the manuscript.
Thanks to Helmut Krawinkler for significant contributions through research discussions and for designing the
four steel moment frame buildings used in this study. Also thanks to Yousef Bozorgnia and Farzin Zareian
for sharing the analysis details about the two concrete buildings considered in this study, and to Hesaam Aslani
and Pasan Seneviratna for discussions on this research topic. Any opinions, findings and conclusions or recom-
mendations expressed in this material are those of the authors and do not necessarily reflect those of the Risk
Management Solutions.

REFERENCES
1. Krishnan S, Ji C, Komatitsch D, Tromp J. Case studies of damage to tall steel moment-frame buildings in Southern
California during large San Andreas earthquakes. Bulletin of the Seismological Society of America 2006; 96(4):
1523–1537.
2. Shome N, Bazzurro P. Comparison of vulnerability of a new high-rise concrete moment frame structure using
HAZUS and nonlinear dynamic analysis. The 10th International Conference on Structural Safety and Reliability,
ICOSSAR, Osaka, Japan, 2009.
3. Moehle J, Bozorgnia Y, Jayaram N, Jones P, Rahnama M, Shome N, Tuna Z, Wallace J, Yang T, Zareian F. Case studies
of the seismic performance of tall buildings designed by alternative means. Task 12 Report for the Tall Buildings
Initiative, Pacific Earthquake Engineering Research Center Report 2011/05, California, USA, 2011.
4. ATC-72. Modeling and acceptance criteria for seismic design and analysis of tall buildings. Applied Technology
Council, 2010.
5. Lignos DG, Krawinkler H. Sidesway collapse of deteriorating structural systems under seismic excitations. Rep. No.
TB 172, The John A. Blume Earthquake Engineering Center, Stanford University, Stanford, CA, 2009.
6. Cornell CA, Krawinkler H. Progress and challenges in seismic performance assessment. PEER Center News 2000; 3(2):1–4.
7. Deierlein GG. Overview of a comprehensive framework for earthquake performance assessment. Technical report,
International Workshop on Performance-Based Seismic Design Concepts and Implementation, Bled, Slovenia, 2004.
8. Baker JW. The conditional mean spectrum: A tool for ground-motion selection. Journal of Structural Engineering
2011; 137(3):322–331.
9. Shome N, Luco N. Loss estimation of multi-mode dominated structures for a scenario of earthquake event. The 9th
U.S. National and 10th Canadian Conference on Earthquake Engineering, 2011.
10. Luco N, Cornell CA. Effects of random connection fractures on the demands and reliability for a 3-story pre-Northridge
SMRF structure. Proceedings of The Sixth US National Conference on Earthquake Engineering, Seattle, Washington,
1998.
11. Boore DM, Atkinson GM. Ground-motion prediction equations for the average horizontal component of PGA, PGV
and 5%-damped PSA at spectral periods between 0.01s and 10.0s. Earthquake Spectra 2008; 24(1):99–138.
12. Jayaram N, Baker JW. Statistical tests of the joint distribution of spectral acceleration values. Bulletin of the Seismological
Society of America 2008; 98(5):2231–2243.
13. Baker JW, Jayaram N. Correlation of spectral acceleration values from NGA ground motion models. Earthquake
Spectra 2008; 24(1):299–317.
14. Johnson RA, Wichern DW. Applied Multivariate Statistical Analysis. Prentice Hall: Upper Saddle River, NJ, 2007.
15. USGS. Interactive Deaggregations (Beta), 2008. Available form: https://geohazards.usgs.gov/deaggint/2008/ (last
verified November, 2011).

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe
N. JAYARAM, N. SHOME AND M. RAHNAMA

16. Haselton CB, Baker JW, Bozorgnia Y, Goulet CA, Kalkan E, Luco N, Shantz TJ, Shome N, Stewart JP, Tothong P,
Watson-Lamprey JA, Zareian F. Evaluation of ground motion selection and modification methods: Predicting median
interstory drift response of buildings. Technical report, Report 2009/01, Pacific Earthquake Engineering Research
Center, University of California, Berkeley, CA, 2009.
17. Aslani H, Miranda E. Probabilistic earthquake loss estimation and loss disaggregation in buildings. Technical Report
# 157, John A. Blume Earthquake Engineering Center, Stanford University, CA, 2005.
18. Ramirez CM, Miranda E. Building-specific loss estimation methods and tools for simplified performance-based
earthquake engineering. Technical Report # 171, John A. Blume Earthquake Engineering Center, Stanford University,
CA, 2009.
19. Kutner MH, Nachtsheim CJ, Neter J, Li W. Applied Linear Statistical Models. The McGraw-Hill Companies Inc.:
New York, 2005.
20. Katsanos EI, Sextos AG, Manolis GD. Selection of earthquake ground motion records: A state-of-the-art review from
a structural engineering perspective. Soil Dynamics and Earthquake Engineering 2010; 30(4):157–169.
21. Shome N, Cornell CA, Bazzurro P, Carballo JE. Earthquakes, records and nonlinear responses. Earthquake Spectra
1998; 14(3):469–500.
22. HAZUS. Multi-hazard Loss Estimation Methodology. Technical Manual, FEMA, 1999.
23. ATC-58. Guidelines for Seismic Performance Assessment of Buildings. ATC-58 50% Draft. Applied Technology
Council, 2009.
24. FEMA 355F. State of the art report on performance prediction and evaluation of steel moment-frame buildings.
Federal Emergency Management Agency, Washington, D.C., 2000.
25. FEMA-273. NEHRP guidelines for the seismic rehabilitation of buildings. Federal Emergency Management Agency,
Washington, D.C., 1997.
26. Jalayer F, Cornell CA. A technical framework for probability-based demand and capacity factor design (DCFD) seismic
formats. Technical Report 2003/08, Pacific Earthquake Engineering Research Center, University of California, Berkeley,
CA, 2004.
27. Baker JW. Introducing correlation among fragility functions for multiple components. The 14th World Conference
on Earthquake Engineering, Beijing, China, 2008.
28. Hu Y, Liu SC, Dong W. Earthquake Engineering. E&FN SPON, Chapman & Hall: New York, 1996.
29. Field EH, Dawson TE, Felzer KR, Frankel AD, Gupta V, Jordan TH, Parsons T, Petersen MD, Stein RS, Welson II
RJ, Wills CJ. The uniform California earthquake rupture forecast, Version 2 (UCERF 2). USGS Open File Report
2007–1437, 2008.
30. Cramer CH, Petersen MD, Reichle MS. A Monte Carlo approach in estimating uncertainty for a seismic hazard
assessment of Los Angeles, Ventura, and Orange Counties, California. Bulletin of the Seismological Society of America
1996; 86(6):1681–1691.
31. Jayaram N, Baker JW. Efficient sampling and data reduction techniques for probabilistic seismic lifeline risk assessment.
Earthquake Engineering and Structural Dynamics 2010; 39:1109–1131.

Copyright © 2012 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. (2012)
DOI: 10.1002/eqe

You might also like