Engineering Structures 28 (2006) 289303 www.elsevier.

com/locate/engstruct

The effect of material and ground motion uncertainty on the seismic vulnerability curves of RC structure
Oh-Sung Kwona,, Amr Elnashaib
a University of Illinois, 205 North Mathews, Urbana, IL 61801, United States b Mid-America Earthquake Center, University of Illinois, 205 North Mathews, Urbana, IL 61801, United States

Received 23 September 2004; received in revised form 17 June 2005; accepted 12 July 2005 Available online 3 October 2005

Abstract Starting from the premise that vulnerability curves are an indispensable ingredient of earthquake loss assessment, this paper focuses on establishing the relative effect of strong-motion variability and random structural parameters on the ensuing vulnerability curves. Moreover, the effect of the selection of statistical models used to present simulation results is studied. A three story ordinary moment resisting reinforced concrete frame, previously shake-table tested, is used as a basis for the fragility analysis. The analytical environment and the structural model are veried through comparison with shaking-table test results. The selection of ground motion sets, denition of limit states, statistical manipulation of simulation results, and the effect of material variability are investigated. No approximations are used to reduce the sample size or minimize the analytical effort, in order that attention is focused on the parameters under investigation. Notwithstanding the limited scope of the study, the results presented indicate that the effect of randomness in material response parameters is far less signicant than the effect of strong-motion characteristics. Therefore, the importance of scrupulous selection and scaling of strong-motion and use of appropriate limit states and statistical models is emphasized. 2005 Elsevier Ltd. All rights reserved.
Keywords: Seismic vulnerability; Ground motion uncertainty; Material uncertainty; Ordinary moment resisting concrete frame

1. Introduction Vulnerability curves relate strong-motion shaking severity to the probability of reaching or exceeding a specied performance limit state. Strong-motion shaking severity may be expressed by an intensity (I ), peak ground parameters (a, v or d) or spectral ordinates (Sa , Sv or Sd ) corresponding to an important structural period. The number of limit states used varies between three and ve. In this paper, three limit states considered as the most signicant are used; serviceability, damage control and collapse prevention. Vulnerability curves play a critical role in regional seismic risk and loss estimation as they give the probability of attaining a certain damage state when a structure is
Corresponding author. Tel.: +1 217 265 5497; fax: +1 217 333 3821.

subjected to a specied demand. Such loss estimations are essential for the important purposes of disaster planning and formulating risk reduction policies. The driving technical engines of a regional seismic risk and loss estimation system are [12]: Seismic hazard maps (i.e. peak ground parameters or spectral ordinates). Vulnerability functions (i.e. relationships of conditional probability of reaching or exceeding a performance limit state given the measure of earthquake shaking). Inventory data (i.e. numbers, location and characteristics of the exposed system or elements of a system). Integration and visualization capabilities (i.e. data management framework, integration or seismic risk and graphical projection of the results). The scope of this study is to present vulnerability curves of a reinforced concrete structure subjected to

E-mail addresses: okwon2@uiuc.edu (O.-S. Kwon), aelnash@uiuc.edu (A. Elnashai). 0141-0296/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.engstruct.2005.07.010

290 Table 1 Categorization of vulnerability curvea Category

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

Characteristics Feature Limitation Based on post-earthquake survey Most realistic Highly specic to a particular seismo-tectonic, geotechnical and built environment The observational data used tend to be scarce and highly clustered in the low-damage, low-groundmotion severity range Include errors in building damage classication Damage due to multiple earthquakes may be aggregated [24] Based on expert opinion The curves can be easily made to include all the factors The reliability of the curves depends on the individual experience of the experts consulted A consideration of local structural types, typical congurations, detailing and materials inherent in the expert vulnerability predictions ATC-13 [1] Based on damage distributions simulated from the analyses Reduced bias and increased reliability of the vulnerability estimate for different structures Substantial computational effort involved and limitations in modeling capabilities The choices of the analysis method, idealization, seismic hazard, and damage models inuence the derived curves and have been seen to cause signicant discrepancies in seismic risk assessments [21,26,9] Compensate for the scarcity of observational data, subjectivity of judgmental data, and modeling deciencies of analytical procedures Modication of analytical or judgment based relationships with observational data and experimental results The consideration of multiple data sources is necessary for the correct determination of vulnerability curve reliability [19]

Empirical vulnerability curve

Sample ref. Feature Judgmental vulnerability curve Limitation

Sample Ref. Feature Analytical vulnerability curve Limitation

Sample Ref. Feature

Hybrid vulnerability curve Limitation Sample Ref.

a Mainly excerpted from [27].

various ground motion sets and to investigate the effects of material uncertainties and selected ground motion sets on the obtained vulnerability curves. In addition, several other aspects of vulnerability curve derivation are investigated such as the selection of ground motion duration, denition of limit states, selection of damping parameter, and statistical manipulation. To establish a framework for the study, a brief review of vulnerability curves and the procedures used in this study are presented. Details of the structural and analytical models, alongside a brief description of the analytical platform and uncertainties in the analysis, are introduced and limit states are dened for the selected demonstration structure. Sensitivity analyses are conducted and vulnerability curves are derived. The results are fully investigated to identify trends and derive conclusions on the relative sensitivity to input motion and the randomness of material properties. 2. Brief review of analytical vulnerability functions Vulnerability functions exhibit considerable variability depending on the approaches used in their derivation. The factors that inuence the vulnerability functions are input ground motion sets, severity indices of ground motions, performance limit states, source of structural

damage data, structural modeling method, analysis platform characteristics, analysis method, consideration of epistemic uncertainty, etc. Based on the sources of data, vulnerability curves may be sub-divided into four categories [27] as summarized in Table 1. A class of the curves are based on observational data from post-earthquake surveys (e.g. [24,27]) while others are based on analytical simulation (e.g. [21,26, 9]). Empirical vulnerability curves should be inherently more realistic than their analytical counterparts should, since they are based on the observed damage of actual structures subjected to real strong motion. They have, however, limitations in general application since the curves are derived for a specic seismic region and a sample that is not necessarily similar to that sought. On the other hand, analytical vulnerability curves can be derived for general purposes, but the choice of analytical model, simulation method, and required computational power pose challenges for the development of the required relationship. One of the main criteria for the selection of the method is the availability of structural damage data; either the observation of post-earthquake losses or the analytical simulation. Observational data are realistic, but are often neither statistically viable nor homogeneous. The data

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

291

Fig. 1. Flow chart for the derivation of analytical vulnerability curves.

from simulation, on the other hand, is constrained by computational power and reliability of analytical tools. With the expansion of computational power and the development of reliable analysis tools, the limitations in the analytical derivation of vulnerability curves are diminishing. For the analytical derivation of vulnerability curves, Mosalam et al. [21] used single-degree-of-freedom (SDOF) systems, representations of the pushover curves of inlled and bare frames. In HAZUS [23], the variability in seismic demand is provided without explicit consideration of the inuence of the structural parameters such as damping, period, and yield strength level. Reinhorn et al. [26] used constant yield-reduction-factor inelastic spectra with the capacity spectrum method to evaluate inelastic response. The above studies used simplied methods because the derivation of vulnerability curves required a large amount of simulation. Thus, the results are approximate as these methods neglect the effect of higher modes, hysteretic damping, and limit states based on local failure. In this study, inelastic dynamic response-history analysis is adopted for the sample building using ZEUS-NL [13] as the analysis platform. Multi-threading techniques are used to deploy ZEUS-NL on large computer clusters, thus reducing very substantially the computational time. 3. Procedure of vulnerability curve derivation Four aspects of the derivation process mainly affect vulnerability curves as shown in Fig. 1. These are structure, hazard denition, simulation method, and vulnerability analysis. Each component can be divided into a number of sub-tasks. By denition, vulnerability analysis is probabilistic since each of the constituent components is uncertain. Some of the uncertainties are inherently random while others are consequences of lack of the knowledge. In this study, only uncertainties in material properties and input motion are considered. Those uncertain components

are indicated by the shaded areas in Fig. 1. The remaining components, such as limit state, analysis method, structural modeling, etc., also affect the derived vulnerability curves. The estimation of these uncertainties, however, inevitably includes subjective judgment. In this study, the latter variables are assumed deterministic. Hence, the derived vulnerability curves are conditional on the assumption that the input, the analysis methods, and the output from the analysis reect reality. A structure representing an important sub-class in the Mid-America region is chosen; a medium rise RC frame structure with no seismic detailing. The ultimate strength of concrete and yield strength of steel are selected as random variables. An analytical model of the structure is constructed and veried through comparison with shaking table experiments. The limit states of the structure are dened based on material response and sectional behavior obtained from pushover analysis. The seismic hazard is dened by selecting several ground motion sets. Full combination of ground motion sets and random material properties are used for the simulation. Notwithstanding previous studies that suggested the use of a few records, the current investigation employs full Monte Carlo simulation under all earthquake records, scaled at small intervals. 4. Selection of reference structure for simulation A three-story ordinary moment resisting concrete frame (OMRCF) is chosen for this benchmark study. It is appreciated that the curves developed may not be generally applicable to the loss estimation of all RC buildings. The sample structure, however, serves ideally the purposes of this study, since experimental results under earthquake loading exist for detailed verication of the analytical model. Moreover, it is postulated that the curves derived could be applicable to the sub-class of medium rise RC buildings with limited ductility and no seismic design provisions; this might be applicable to many areas in the mid-west of the USA and Central/Northern Europe. The prototype structure was originally designed for the purpose of an experimental study [6]. The building has three and four bays in EW and NS directions, respectively. The story height is 3.7 m (12 ft) and the bay width is 5.5 m (18 ft). The total building height is 11 m (36 ft). It is designed for gravity loads since wind loads seldom govern for low-rise buildings, and is non-seismically detailed. The provisions of ACI 318-89 code, with Grade 40 steel [ f y = 276 MPa (40 ksi)] and ordinary Portland cement [ f c = 24 MPa (3.5 ksi)], was employed. The plan and elevation layouts of the structure are given in Fig. 2. The analyzed frame is shaded in the gure. For detailed design information, reference is made to Bracci et al. [6]. 5. Selection of analysis method and environment Several analysis methods have been proposed to determine the seismic demand of structures subjected to

292

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

inelastic material response of the individual bers describing the section. The Eulerian approach towards geometric nonlinearity is employed on the element level. Therefore, full account is taken of the spread of inelasticity along the member length and across the section depth as well as the effect of large member deformations. Since the sectional response is calculated at each loading step from inelastic material models that account for stiffness and strength degradation, there is no need for sweeping assumptions on the momentcurvature relationships required by other analysis approaches. In ZEUS-NL, conventional pushover, adaptive pushover, Eigen analysis, and dynamic analyses are available and have been tested on the member and structural levels ([14,15,7,20,25] amongst others). Recently, ZEUSNL was used to steer a full scale 3D RC frame testing campaign, and the a priori predictions were shown to be accurate and representative of the subsequently undertaken pseudo-dynamic tests [17,18]. In the following section, the verication of the analysis model and environment through comparison with shaking table experiments are undertaken. 6. Verication of analytical model The structural model and analysis environment are veried through comparison of response history analysis with shake table test by Bracci et al. [6]. In response history analysis, damping, mass and stiffness are key parameters that affect the assessment result. The verication is undertaken in terms of structural periods and global displacement time histories since local stressstrain measurements are not available in the published literature. 6.1. Structural period Columns and beams are divided into six and seven elements, respectively, in the numerical model. Mass is deposited at the beam and column connections, as shown in Fig. 3. Material properties are taken from the reported test result of the experimental model. The elastic structural periods from eigenvalue analysis are 0.898, 0.305, and 0.200 s for the rst, second, and third modes, respectively. Bracci et al. [6] conducted a snap-back test before running shaking table experiments to estimate the natural periods of the 1/3 scale specimen and found that the periods, after conversion to full scale using similitude laws, were 0.932, 0.307, and 0.206 s. The experimental values under small amplitude testing are just 3%4% longer than the analytical values which might have resulted from minor cracking in the test specimen. These values give credence to the analytical model. 6.2. Displacement response history verication Fig. 4 depicts the comparison of the 3rd story displacements of the 1/3 scale specimen and analysis using a 1/3

Fig. 2. Plan and elevation of prototype structure [6], (a) plan, (b) elevation.

earthquake loading. Static pushover analysis, conventional, modal, or adaptive, yields the capacity and collapse mechanism of a structure. For seismic response assessment, though, a seismic demand procedure is required, where effective damping or equivalent ductility is accounted for. Such equivalence may introduce approximations in the analysis results. In addition, the more irregular the structure is and more peculiar the strong-motion records are, the less representative are pushover results of dynamic response. This limitation lends weight to the use of dynamic analysis which deals with the coupled demandcapacity problem [11], especially for an irregular structure. Since structures not designed to resist seismic loads usually fail in localized modes, their response is not likely to be well estimated by static methods. Moreover, in order to focus attention on other approximations in the vulnerability functions derivation, it was decided to deploy the most accurate and generally applicable method available for seismic demand and supply evaluation; inelastic dynamic response-history analysis. Use is made of the Mid-America Earthquake Center analysis environment ZEUS-NL [13]. Elements capable of modeling material and geometric nonlinearity are available in the program. The sectional stressstrain state is obtained through the integration of the

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

293

Fig. 3. Analytical model conguration.

scale analytical model. For the analysis, Rayleigh damping is used for small amplitude ground excitation with PGA of 0.05g, in which the damping ratio was taken from the snap-back test, [6]. For moderate and severe ground motions of 0.20g and 0.30g, damping other than hysteretic is not included. Fig. 4(a) does not show good agreement whilst Fig. 4(b) and (c) show very good agreement between experimental and analytical results. This is attributed to the fact that at low level ground motion, it is difcult to accurately estimate the level of damping and to model initial cracks due to curing and experimental set-up. On the contrary, at medium to high earthquake motion, the inelastic response from the fully cracked section mainly governs the behavior, thus reducing the effect of cracking and small amplitude damping. Good agreement in the moderate-tolarge amplitude shaking veries that the analytical model represents the experimented structure well. It is also observed that assuming the same level of damping from the low amplitude to the collapse level ground motion could result in non-conservative vulnerability curves at medium-to-high ground motion intensities. In this study, it is assumed that there is no source of damping other than hysteretic. Thus, for low level ground motions, the vulnerability curves might be on the conservative side. Assuming no viscous damping may cause spurious higher mode oscillation in a structure without any other sources of energy dissipation. In the studied structure, however, this is not a signicant issue because, (a) the shaking table experiment conrmed that the numerical model is of good accuracy, (b) the height of the structure is short, and with no irregularity, hence there is limited possibility for spurious modes, and (c) the inelastic concrete material used in the analysis shows hysteretic damping even

under very small magnitude of vibrations, thus it damps out very short period spurious modes. The current and previous verication of the modeling approach and analysis platform lend weight to the condent use of the analytical tools to investigate the effects of parameter variation within the fragility analysis presented in this paper. 7. Uncertainties in capacity and demand In the derivation of vulnerability functions, a probabilistic approach is adopted owing to uncertainties in the hazard (demand) as well as structural supply (capacity). Some of those uncertainties stem from factors that are inherently random (referred to as aleatoric uncertainty), or from lack of knowledge (referred to as epistemic uncertainty) [30]. In this paper, the effects of aleatory uncertainties from material and ground motion on the vulnerability curves are investigated. Epistemic uncertainty considerations are beyond the scope of this study. 7.1. Material uncertainty 7.1.1. Concrete strength Barlett and MacGregor [3] investigated the relationship between the strength of cast-in-place concrete and specied concrete strength. When concrete is 1 year old, the ratio of the average in-place strength to the specied strength was 1.33 and 1.44 for short and long elements, respectively, with a coefcient of variation of 18.6%. The variation of strength throughout the structure for a given mean in-place strength

294

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

Fig. 4. Comparison of dynamic analysis: (a) 3rd story displacement Taft 0.05g, (b) 3rd story displacement Taft 0.20g, (c) 3rd story displacement Taft 0.30g.

depends on the number of members, number of batches, and type of construction. In this study, it is assumed that there is no variability of concrete strength since the structure is a low-rise building of limited volume that would have been constructed in a relatively short period. Thus, a coefcient of variation of 18.6% is adopted. The specied concrete strength (or design strength) of the considered structure was 24 MPa. In-place concrete strength is assumed 1.40 times larger than the specied strength (33.6 MPa). Normal distribution assumption is adopted for the concrete strength. 7.1.2. Steel strength Mirza and MacGregor [22] reported results of about 4000 tests on Grade 40 and 60 bars. The mean values and coefcient of variation of the yield strength were 337 MPa (48.8 ksi) and 10.7%, respectively. The probability distribution of modulus of elasticity of Grade 40 reinforcing steel followed a normal distribution with a mean value 201,327 MPa (29,200 ksi) and a coefcient of variation of 3.3%. Due to the low level of variability observed, the modulus of elasticity is assumed deterministic

(201,327 MPa) in this study. The structure was designed with grade 40 steel, thus the mean steel strength is assumed 337 MPa. The steel strength is assumed to follow a normal distribution. 7.2. Input motion uncertainty 7.2.1. Selection of ground motion In this study, nine sets of ground motions are used for the derivation of vulnerability curves. Derived vulnerability curves for each set are compared to gain insight into the effect of ground motion variation on fragility analysis. The rst three sets of ground motions are based on the ratio of peak ground acceleration to peak ground velocity (a/v). Zhu et al. [31] discussed these three categories of earthquake ground motions and their engineering and seismological signicance. The a/v ratio implicitly accounts for many seismo-tectonic features and site characteristics of earthquake ground motion records. Sawada et al. [28] concluded that low a/v ratios signify earthquakes with low predominant frequencies, broader response spectra,

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303 Table 2 Properties of the selected ground motions based on a/v ratio a/v ratio Earthquake event/Location Bucharest/Romania Erzincan/Turkey Aftershock of Montenegro/Yugoslavia Kalamata/Greece Kocaeli/Turkey Aftershock of Friuli/Italy Athens/Greece Umbro-Marchigiano/Italy Lazio Abruzzo/Italy Basso Tirreno/Italy Gulf of Corinth/Greece Aftershock of Montenegro/Yugoslavia Aftershock of Montenegro/Yugoslavia Aftershock of Umbro-Marchigiana/Italy Friuli/Italy ML 6.40 Unknown 6.20 5.50 Unknown 6.10 Unknown 5.80 5.70 5.60 4.70 6.20 6.20 5.00 6.30 Date 3/4/1977 3/13/1992 5/24/1979 9/13/1986 8/17/1999 9/15/1976 9/7/1999 9/26/1997 5/7/1984 4/15/1978 11/4/1993 5/24/1979 5/24/1979 11/9/1997 5/6/1976 Soil type Rock Stiff soil Alluvium Stiff soil Unknown Soft soil Unknown Stiff soil Rock Soft soil Stiff soil Rock Alluvium Rock Rock D (km) 4 13 8 9 101 12 24 27 31 18 10 32 16 2 27 A max (m/s2 ) 1.906 3.816 1.173 2.109 3.039 0.811 1.088 0.992 0.628 0.719 0.673 0.667 1.709 0.412 3.500

295

a/v ratio (g/ms1 ) 0.275 0.382 0.634 0.657 0.750 1.040 1.090 1.108 1.136 1.183 1.432 1.526 1.564 1.902 1.730

Low

Intermediate

High

Table 3 Properties of the articial ground motions for Memphis, TN EQ scenario Set ID PGA (g) PGV (m/s) PGD (m) Set ID PGA (g) PGV (m/s) PGD (m) Scenario 1 7.5 @ Blytheville, AR Set L-1 0.1427 0.1520 0.0606 Set U-1 0.1407 0.1290 0.0537 Scenario 2 6.5 @ Marked Tree, AR Set L-2 0.0632 0.0576 0.0202 Set U-2 0.0676 0.0516 0.0178 Scenario 3 5.5 @ Memphis, TN Set L-3 0.0958 0.0665 0.0138 Set U-3 0.1030 0.0609 0.0118

Memphis, TN (Lowlands)

Memphis, TN (Uplands)

longer durations and medium-to-high magnitudes, long epicentral distances and site periods. Conversely, high a/v ratios represent high predominant frequencies, narrow band spectra, short duration and smallmedium magnitudes, short epicentral distance and site periods. Ground motions were classied in the following ranges: Low: a/v < 0.8g/m s1 Intermediate: 0.8g/m s1 a/v 1.2g/m s1 High: 1.2g/m s
1

(1)

< a/v.
Fig. 5. Average response spectrum of selected ground motion sets.

Based on the above categorization, three sets of ground motions are selected in this study (Table 2). The average response spectra of selected ground motion sets (Fig. 5) show distinctive difference among each ground motion set. The remaining six sets used in this study are articial ground motions. Drosos [10] generated the bedrock motion for the Mississippi Embayment in the New Madrid Seismic Zone. Equivalent linear site response analyses were conducted to evaluate the soil surface ground motions. During the site response analysis, shear wave velocity proles were randomized to account for the uncertainties in shear wave velocity and layer thickness. Three of the sets, set L-1, L-2, and L-3, were generated based on a Lowlands soil

prole in the Memphis area. The other three sets, set U-1, U-2, and U-3, were generated based on an Uplands soil prole in the same region. Each of the three sets of ground motions were based on three scenario earthquakes; small, medium, and large, at three epicenter distances, short, medium, and long. Each set contains ten ground motions. Table 3 shows the ground motion parameters for Memphis, TN.

296

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

7.3. Random variables sampling Three different ground motion sets are combined with different material properties. For ground motion set low, normal, and high a/v, ten ultimate concrete strengths, f c , and ten steel yield strengths, Fy , were generated and a full combination of material strengths are used, resulting in a total of 100 frames. The analysis results are used to study the effect of material properties on the structural response. For ground motion set L-1, L-2, and L-3, which are articial ground motions based on Lowlands prole, 50 concrete and steel strengths are generated following the mean and standard deviation given in Section 7.1 resulting in 50 different frames. For ground motion set U-1, U-2, and U-3 based on Uplands prole, 100 concrete and steel strengths are generated. From the analysis result of these frames, the effect of sample size is investigated. 8. Ground motion duration and scale factors 8.1. Ground motion duration The duration of the signicant part of strong motion affects the maximum response when a structure undergoes inelastic deformations. There are many previous studies aimed at dening the duration of strong ground motion. Bommer and Martinez-Pereira [4] assembled 30 denitions of strong-motion duration suggested by previous researchers and divided them into four categories: bracketed, uniform, signicant, and structural response-based denitions. Since the selected motions are to be scaled for vulnerability curve generation, the duration should be dened in a relative manner as in the signicant duration option. Trifunac and Brady [29] used signicant duration concepts based on the integral of the square of acceleration, velocity and displacement where the duration is dened as the interval between the times at which 5% and 95% of the total integral is attained. The latter range of duration is meaningful in characterizing ground motions. From a structural analysis point of view, however, this duration may not be practical. For instance, if the above-mentioned interval of total integral is used as shown in the Fig. 6, the ground motion acceleration could start at very large values, which may apply an unrealistic pulse to the structure. Moreover, since the major part of ground motion energy is skewed to the early part of the motion, using identical margins for the start and the end of the duration is not a reasonable approach. Based on the above argument, for the current study, the interval between 0.5% and 95% of the integrals is used. 8.2. Selection of scale factors The computational demand for derivation of vulnerability curve is very large. Analysis time for 100 frames (combination of 10 concrete and 10 steel strengths)

subjected to ve ground motions (e.g. intermediate a/v ratio ground motion set) that are scaled from PGA of 0.050.5g at 0.05g interval was 121 h on a fast PC (Pentium IV 2.65 GHz, 1 GB of RAM) at the time of the current study. Thus, a reasonable range of PGA levels should be selected to most effectively utilize the computational power, since each ground motion set has different spectral acceleration at the fundamental period of the structure. For example, for the low a/v ratio ground motion set, a PGA of 0.50g was too large because most structures collapsed at a PGA of 0.2g. Conversely, for high a/v ratio ground motion set, 0.50g PGA level was not large enough to derive the vulnerability curve for the collapse limit state. Thus, for low and high a/v ratios, additional selective analyses had to be performed to improve the resolution of the vulnerability curves. For the determination of the range of reasonable PGA scaling, the capacity spectrum method was utilized. The capacity curves were obtained from adaptive pushover analysis, and demand curves were converted from the elastic displacement and acceleration spectra of each ground motion set. For accurate estimation of maximum PGA scaling at which the structure collapses, elastic demand should be decreased to consider inelasticity using effective damping [2,5] or ductility ratio [8]. In this analysis, however, rough estimation of PGA scales is undertaken using elastic demand spectra and inelastic capacity spectra. 9. Limit state denition In ATC 40 [2] and FEMA-273 [16], four limit states are dened based on global behavior (interstory drift) as well as element deformation (plastic hinge rotation). Rossetto and Elnashai [27] used ve limit states for derivation of vulnerability curves based on observational data while Chryssanthopoulos et al. [9] used only two limit states. In the latter studies, the global limit states are independent of the specic response of the structure. For example, the FEMA-273 [16] life safety level limit state of interstory drift (ISD) for non-ductile moment resisting frame is 1.00% regardless of gravity force levels or the details of structural conguration within the sub-class of structure. For rigorous analysis, it is necessary to dene limit states for each individual structure, since the deformational capacity could be affected by many other factors such as gravity force level, irregularity, anticipated plastic hinging mechanism, etc. In this study, three limit states are dened for the prototype structure based on the rst yielding of steel, attainment of maximum element strength, and maximum conned concrete strain during the adaptive pushover analysis. These are termed, serviceability, damage control, and collapse prevention, limit states, respectively. In this study, the local damage of individual structural element, such as beam, column, or beamcolumn joint, is not accounted for. Only interstory drift is used as a global measure of damage. The 1st story drift which corresponds to each limit state, for the prototype

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

297

Fig. 6. Signicant duration of a ground motion.

Fig. 7. Denition of limit states: (a) serviceability state, (b) damage control state, (c) collapse state.

structure, is 0.57%, 1.2% and 2.3% for the selected three limit states, as shown in Fig. 7. In the gure, c01 through c31 represent the bottom element of the 1st

story columns as indicated in Fig. 3. It is assumed that theses limit states are also applicable to the 2nd and 3rd stories.

298

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

Fig. 9. Derived vulnerability curves using various methods normal a/v ratio, limit state = ISD 0.57%. Fig. 8. Distinction of collapse state and lognormal distribution assumption (ground motion normal a/v ratio, PGA = 0.40g).

= P(ISD > ISDLimit | E 1 ) P(E 1 ) + 1.0 P(E 2 ) (2) where P(E 1 ) and P(E 2 ) represents the probability that the structure is in a stable or in a collapse state, respectively. As shown in Fig. 8, a lognormal distribution was assumed for structures with ISD < 2.3%, and the structures were considered to have collapsed if ISD 2.3%. This assumption provides conservative results. Fig. 9 shows a sample vulnerability curve for the 0.57% ISD limit state using intermediate a/v ratio ground motion sets. Without considering collapse state ISD, normal and lognormal distribution assumptions show lower probability of attaining the limit state at larger PGA level, which is due to the misleading average and coefcient of variance of unstable structures. 11. Effect of analysis parameters on vulnerability curves 11.1. Effect of ground motion set Fig. 10 compares vulnerability curves for each limit state derived from nine ground motion sets. For general application of vulnerability curves, it is necessary to perform regression analysis to obtain functional forms that are readily programmable in loss assessment software. For the purpose of comparison amongst different ground motion sets, however, actual data with linear interpolation are plotted in order to insure that differences are not masked by regression smoothing. Fig. 11 shows the coefcient of variation of ISD from each ground motion set vs. ground motion level. In this gure, all response data, including collapse state, are used. From Figs. 10 and 11 the following observations are made: (a) Because of large discrepancies in the coefcient of variation of the maximum interstory drift, each ground motion set shows signicantly different vulnerability curves.

10. Simulation and vulnerability curve derivation The total simulation analysis of the frame using ground motion set low, normal, and high a/v ratio required 456 h using a Pentium IV-2.65 GHz PC for a total of 23,000 response history analyses. For the analysis of frames using ground motion set L-1 through U-3, a mass-simulation environment was developed. The masssimulation approach deploys 32 processes simultaneously on the super-computer IBM p690 at the National Center for Supercomputing Applications (NCSA). Interstory drifts are retrieved from each run and used for statistical analysis. The maximum interstory drifts are assumed to follow a lognormal distribution. The normal distribution is avoided since its abscissa ranges from negative innity to positive innity, which implies that negative maximum interstory drift could exist. The analysis platform used for simulation accounts for geometric nonlinearity as well as material inelasticity. Thus, when the structure is subjected to large seismic demands, instability under gravity loading may ensue. Since the drift of unstable structures depends on convergence criteria of the analysis tool, it is concluded that interstory drift at such response states should not be included in the statistical analysis. Thus, the structure is in the collapse state if the maximum interstory drift is larger than 2.3%, as discussed in Section 9. The value, 2.3% is applicable only to the studied structure, and possibly the sub-class of non-seismically designed medium-rise RC frames. To include the statistics of collapsed frames, the total probability theorem is adopted. The probability where a frame maximum interstory drift will be larger than a certain limit state is calculated as below: P(ISD > ISDLimit) = P(ISD > ISDLimit | E 1 ) P(E 1 ) + P(ISD > ISDLimit | E 2 ) P(E 2 )

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

299

Fig. 10. Vulnerability curves for each ground motion set.

(b) The differences between the vulnerability curves increase for higher damage limits. (c) The difference between the vulnerability curves increases with ground motion intensity until the structure reaches the collapse state. (d) The vulnerability curves do not indicate monotonic increase in damage prediction at large PGA levels due to instability of structure and statistical manipulation. Consequently, it is concluded that the ground motion sets should be selected with great care since they affect very signicantly the outcome from the fragility analysis represented by vulnerability curves.

11.2. Effect of material properties The effect of material properties is investigated using analysis results from a/v ratio ground motion sets for which 10 concrete ultimate strengths and 10 steel yield strengths are combined. It is assumed that ISDmax is a function of ground motion sets, concrete ultimate strength, and steel yield strength; i.e, ISDmax = g(X 1 , X 2 , X 3 ) where X 1 : ground motion X 2 : concrete strength X 3 : steel strength. (3)

300

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

Fig. 11. COV of ISDmax for each ground motion set. Fig. 12. Mean ISDmax from mean material properties and full simulation.

Mean ISDmax can be calculated from the result of full simulation, thus using 100 frames, or from the simulation of single frame using mean material properties. Fig. 12 compares the means values of ISDmax using 100 frames and using a single frame with mean material properties. Up to 0.25g, the means from the two methods are almost identical. The difference becomes large as the ground motion intensity increases. Assuming that all uncertain variables, i.e. ground motion, concrete, and steel properties, are statistically independent and as a rst order approximation, the variance of ISD can be calculated from Eq. (4) Var(ISDmax )
2 2 2 + c2 Var(X 2 ) + c3 Var(X 3 )

second and third terms of Eq. (4), to the variance of ISDmax . In these gures, the rst term of Eq. (4) is zero since the plots are developed for each ground motion, i.e. the input is deterministic. The contribution of concrete to the variance of ISDmax is generally larger than that of steel since the former affects stiffness, hence period and amplication characteristics.

11.3. Effect of sample size The analysis results of ground motion set U-1 are used where 100 concrete and steel properties are combined to assess the effect of the sample size on the variance of ISDmax . From the discussion in the previous section, it is reasonable to state that material variability has very little effect on the variability of ISDmax at low PGA levels. The latter observation, however, is not globally applicable; highly irregular systems with random spatial distribution of materials may cause failure mode switches. In such cases, material variability may exhibit more prominence in fragility analysis. Notwithstanding, based on the observations in this study, the variances of ISDmax from the mean frame, 10, 50, and 100 frames were compared. Variance of the mean frame, Var1 , is calculated as in Eq. (5).
n

(4)

where is the standard deviation of the error term and ci is evaluated at x i = X i . The error term represents the uncertainty of ground motion and has zero mean and a standard deviation of , hence the ground motion set adds uncertainty as expressed in the rst term of Eq. (4). The c2 and c3 values can be calculated numerically in the vicinity of the mean values of concrete and steel properties. Variances, Var(X 2 ) and Var(X 3 ), can be calculated from the coefcient of variation, COV = 0.186, and COV = 0.107, for concrete and steel, respectively. Fig. 13(a), (b), (d), and (e) show ISDmax vs. f c and Fy at small (0.05g) and large (0.35g) PGA levels. From Fig. 13(a) and (b), different ground motions exhibit different slopes, d(ISDmax )/d f c . This is because the elastic modulus of concrete is affected by the ultimate strength, f c , hence, the structural period is also affected. As a result, the relationship between ISDmax and f c does not follow a clear trend, i.e. it is inappropriate to state that higher concrete strength reduces structural response since ISDmax and f c are correlated with spectral displacement, which is random in nature. On the contrary, the ISDmax at low PGA level is rarely affected by the yield strength of steel (Fig. 13(d) and (e)), since the elastic modulus of steel is constant regardless of yield strength. Fig. 13(c) and (f) show the contribution of concrete and steel strengths, which are the

Var1 =
i=1

(x i X i )2 n1

(5)

where n is the number of ground motions. If multiple frames are used, the variance can be calculated as below.
jn

Var2 =
i=1

(x i X i )2 jn 1

(6)

where j is the number of frames. From the assumption that the difference in material has little effect on the outcome of the result, Eq. (6) is rewritten as below.

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

301

2 Fig. 13. Effect of material strength on the maximum ISD (a) f c vs. ISDmax at 0.05g, (b) f c vs. ISDmax at 0.35g, (c) c2 Var(X 2 ) vs. PGA, (d) Fy vs. ISDmax 2 Var(X ) vs. PGA. at 0.05g, (e) Fy vs. ISDmax at 0.35g, (f) c3 3

jn

Var2 =
i=1

(x i X i )2 = jn 1
n i=1

j
i=1

(x i X i )2 jn 1 (7)

n1 = n 1/j

n1 (x i X i )2 = Var1 . n1 n 1/j

In other words, the increase in sample size considering material variability, which does not increase variability of ISDmax , reduces the variance of ISDmax considering only ground motion uncertainty. Fig. 14(a) shows the coefcient of variation from 1, 10, 50, and 100 frames vs. PGA level. From this gure, it is observed that the coefcient of variation does not vary much even if uncertain material properties are used. With the stipulation that material properties have little effect on ISDmax , which is true for low PGA levels, the correction factor in Eq. (7) is applied to the variance and the resulting coefcients of variation are compared in Fig. 14(b). The latter gure shows COV for the mean, 10, 50, and 100 frames that are closer than before. Comparison of Figs. 14 and 11 conrms that there is indeed an effect from material variability, but that such an effect is signicantly smaller, verging on the negligible, compared to the effect of ground motion uncertainty.

The above observations conrm that the variability in materials is much less signicant than the effect of ground motion variation. The conclusion is veried by using SDOF system analysis as shown in Fig. 15 where the spectral displacement is plotted vs. period and ductility. The period range, from 0.85 to 0.95 s, corresponds to the period of the prototype structure with concrete ultimate strength of mean one standard deviation. The used range covers most of the possible concrete strength variation. Since steel yield has practically no effect on the structural period, the period axis represents variability in concrete only. The ultimate strength of concrete and yield strength of steel affect the ductility of structures for a given ground motion level. Since the ductility cannot be estimated without analysis, a ductility range from 1 to 2 is plotted for comparison purposes. For the three ground motion sets, U-1, U-2, and U-3, the mean spectral displacement is calculated after normalizing the ground motion to a PGA of 1g. Fig. 15 clearly shows that the structural response parameters are very signicantly affected by the input motion set, while the effect of material variability is rather insignicant.

302

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

Fig. 15. Response surface of spectral displacement against period and ductility.

the resulting variability is much smaller than that due to ground motion variability. Input motion characteristics have a most signicant effect on vulnerability curves. Therefore, meticulous consideration is required when ground motions are selected.
Fig. 14. Comparison of COV with mean frame and multiple frames: (a) variance before applying correction factor, (b) variance after applying correction factor.

12. Conclusion In this study, vulnerability curves of a three-story RC structure are derived for nine sets of ground motions. The effects of ground motion input and material variability are the main focus of the study, alongside other supporting aspects of vulnerability curve derivation such as selection of representative structure, verication of analytical model and analysis environment, effect of explicit damping, scaling of ground motion sets, signicant duration of ground motion, and limit state denition. Even though each of those parameters affects the derived vulnerability curves, many previous studies have been performed with simplied models and sweeping assumptions. In the current study, focus is placed on the most realistic vulnerability curve derivation through inelastic response history analysis. Conclusions are drawn in the preceding sections. Hereafter, some of the ndings as well as comments on the methods adopted in this study are reiterated below: The concrete ultimate strength affects the structural response under low ground motion intensities since the concrete elastic modulus is related to the ultimate strength. The yield strength of steel has little effect on the structural response at low ground motion levels. At high ground motion levels, material properties contribute to the variability in structural response, but

The conclusions above are robust because of the verication of the analysis platform, the care taken in selecting the reference structure, the wide range of selected and scaled input motion and the rigor of the statistical analysis. An important feature of the study is the extensive simulation undertaken with minimum assumptions. It is noted though that for structures with higher levels of irregularity, and where material variability is spatially uncorrelated, different failure modes may be observed where interstory drift may not be an appropriate measure of the structural damage. In addition, since the denition of limit state and response of the structure under earthquake loading vary with structural conguration, the derived vulnerability curve in this study may not be generally applicable to all RC structures of limited ductility. The procedures and discussions in this study, however, provide an insight into the vulnerability derivation of low ductility structures.

Acknowledgements The paper is an outcome of the Mid-America Earthquake Center project DS-3: Advanced Simulation Tools. The MidAmerica Earthquake Center is an Engineering Research Center funded by the National Science Foundation under cooperative agreement reference EEC-9701785. The authors are grateful to Drs. Tom Prudhomme and Cristina Beldica for their help in having access to the parallel processing facilities at the National Center for Supercomputing Applications at the University of Illinois at UrbanaChampaign.

O.-S. Kwon, A. Elnashai / Engineering Structures 28 (2006) 289303

303

References
[1] ATC-13. Earthquake damage evaluation data for california. Redwood City (Palo Alto, California): Applied Technology Council; 1985. [2] ATC-40. Seismic evaluation and retrot of concrete buildings. Redwood City (Palo Alto, California): Applied Technology Council; 1997. [3] Barlett FM, MacGregor JG. Statistical analysis of the compressive strength of concrete in structures. ACI Material J 1996;93(2):15868. [4] Bommer JJ, Martinez-Pereira A. The effective duration of earthquake strong motion. J Earthq Eng 1999;3(2):12772. [5] Borzi B, Calvi GM, Elnashai AS, Faccioli E, Bommer JJ. Inelastic spectra for displacement-based seismic design. Soil Dyn Earthq Eng 2001;21(1):4761. [6] Bracci JM, Reinhorn AM, Mander JB. Seismic resistance of reinforced concrete frame structures designed only for gravity loads: part Idesign and properties of a one-third scale model structure. Technical report NCEER-92-0027. 1992. [7] Broderick BM, Elnashai AS. Seismic resistance of composite beamcolumns in multi-storey structures, part 2: analytical model and discussion of results. J Construct Steel Res 1994;30(3):23158. [8] Chopra AK, Goel RK. Capacity-demand-diagram methods for estimating seismic deformation of inelastic structures: SDF systems. Report No. PEER-1999/02. Berkeley: Pacic Earthquake Engineering Research Center, University of California; 1999. [9] Chryssanthopoulos MK, Dymiotis C, Kappos AJ. Probabilistic evaluation of behaviour factors in EC8-designed R/C frames. Eng Struct 2000;22(8):102841. [10] Drosos VA. Synthesis of earthquake ground motions for the new madrid seismic zone. MS thesis. Atlanta (GA): Georgia Institute of Technology; 2003. [11] Elnashai AS. Do we really need inelastic dynamic analysis? J Earthq Eng 2002;6(1):12330. [12] Elnashai AS. Next generation vulnerability functions for RC structures. In: Proceeding response of structures to extreme loading. 2003. [13] Elnashai AS, Papanikolaou V, Lee D. Zeus NL a system for inelastic analysis of structures. Mid-America Earthquake Center, University of Illinois at Urbana-Champaign, Program Release September 2002; 2002. [14] Elnashai AS, Elghazouli AY. Performance of composite steel/concrete members under earthquake loading, Part I: Analytical model. Earthq Eng Struct Dyn 1993;22(4):31445. [15] Elnashai AS, Izzuddin BA. Modeling of material nonlinearities in steel structures subjected to transient dynamic loading. Earthq Eng Struct Dyn 1993;22:50932.

[16] FEMA 273. NEHRP guidelines for seismic rehabilitation of buildings. Washington (DC): Federal Emergency Management Agency; 1997. [17] Jeong SH, Elnashai AS. Analytical assessment of an irregular RC frame for full-scale 3D pseudo-dynamic testing, Part I: Analytical model verication. J Earthq Eng 2005;9(1):95128. [18] Jeong SH, Elnashai AS. Analytical assessment of an irregular RC frame for full-scale 3D pseudo-dynamic testing, Part II: Condition assessment and test deployment. J Earthq Eng 2005;9(2):26584. [19] Kappos A, Pitilakis K, Stylianidis K, Mordis K, Asimakopoulos D. Cost-benet analysis for the seismic rehabilitation of buildings in Thessaloniki, based on a hybrid method of vulnerability assessment. In: Proceedings of the fth international conference on seismic zonation, vol. I, Nantes (France): Ouest editions; 1995. p. 40613. [20] Martinez-Rueda JE, Elnashai AS. Conned concrete model under cyclic load. Mater Struct 1997;30(197):13947. [21] Mosalam KM, Ayala G, White RN, Roth C. Seismic fragility of LRC frames with and without masonry inll walls. J Earthq Eng 1997;1(4): 693720. [22] Mirza SA, MacGregor JG. Variability of mechanical properties of reinforcing bars. J Struc Div 1979;105(ST5). [23] NIBS. HAZUS, Earthquake loss estimation technology. Technical manual prepared by the National Institute of Buildings Sciences (NIBS) for the Federal Emergency Management Agency (FEMA). 1999. [24] Orsini G. A model for buildings vulnerability assessment using the parameterless scale of seismic intensity (PSI). Earthq Spec 1999; 15(3):46383. [25] Pinho R. Selective repair and strengthening of RC buildings. Ph.D. thesis. Imperial College, London; 2000. [26] Reinhorn AM, Barron-Corvera R, Ayala AG. Spectral evaluation of seismic fragility of structures structural safety and reliability. In: ICOSSAR 2001. 2001. [27] Rossetto T, Elnashai A. Derivation of vulnerability functions for European-type RC structures based on observational data. Eng Struct 2003;25(10):124163. [28] Sawada T, Hirao K, Yamamoto H, Tsujihara O. Relation between maximum amplitude ratio and spectral parameters of earthquake ground motion. In: Proc. 10th world conf. on earthquake engineering, vol. 2. 1992, p. 61722. [29] Trifunac MD, Brady AG. A study on the duration of strong earthquake ground motion. Bull Seismol Soc Amer 1975;65(3):581626. [30] Wen YK, Ellingwood BR, Veneziano D, Bracci J. Uncertainty modeling in earthquake engineering. Mid-America Earthquake Center Project FD-2 report. January; 2003. [31] Zhu TJ, Heidebrecht AC, Tso WK. Effect of peak ground acceleration to velocity ratio on ductility demand of inelastic systems. Earthq Eng Struct Dyn 1988;16:6379.