Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake

Similar documents
Empirical Green s Function Analysis of the Wells, Nevada, Earthquake Source

Effects of fault geometry and slip style on near-fault static displacements caused by the 1999 Chi-Chi, Taiwan earthquake

Effects of Fault Dip and Slip Rake Angles on Near-Source Ground Motions: Why Rupture Directivity Was Minimal in the 1999 Chi-Chi, Taiwan, Earthquake

Slip Partition of the 26 December 2006 Pingtung, Taiwan (M 6.9, M 6.8) Earthquake Doublet Determined from Teleseismic Waveforms

Rupture directivity and source-process time of the September 20, 1999 Chi-Chi, Taiwan, earthquake estimated from Rayleigh-wave phase velocity

Seismic Source Mechanism

RELATION BETWEEN RAYLEIGH WAVES AND UPLIFT OF THE SEABED DUE TO SEISMIC FAULTING

RELOCATION OF THE MACHAZE AND LACERDA EARTHQUAKES IN MOZAMBIQUE AND THE RUPTURE PROCESS OF THE 2006 Mw7.0 MACHAZE EARTHQUAKE

SOURCE MODELING OF RECENT LARGE INLAND CRUSTAL EARTHQUAKES IN JAPAN AND SOURCE CHARACTERIZATION FOR STRONG MOTION PREDICTION

Directivity of near-fault ground motion generated by thrust-fault earthquake: a case study of the 1999 M w 7.6 Chi-Chi earthquake

Widespread Ground Motion Distribution Caused by Rupture Directivity during the 2015 Gorkha, Nepal Earthquake

High-Frequency Ground Motion Simulation Using a Source- and Site-Specific Empirical Green s Function Approach

27th Seismic Research Review: Ground-Based Nuclear Explosion Monitoring Technologies

SUPPLEMENTARY INFORMATION

Figure Locations of the CWB free-field strong motion stations, the epicenter, and the surface fault of the 1999 Chi-Chi, Taiwan earthquake.

Earthquake Stress Drops in Southern California

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 109, B08309, doi: /2002jb002381, 2004

CHARACTERIZATION OF DIRECTIVITY EFFECTS OBSERVED DURING 1999 CHI-CHI, TAIWAN EARTHQUAKE

Synthetic Near-Field Rock Motions in the New Madrid Seismic Zone

SPATIAL DISTRIBUTION OF STRONG GROUND MOTION CONSIDERING ASPERITY AND DIRECTIVITY OF FAULT

Grid-Based Moment Tensor Inversion Technique by Using 3-D Green s Functions Database: A Demonstration of the 23 October 2004 Taipei Earthquake

The Chi-Chi,Taiwan Earthquake: Large Surface Displacements on an Inland Thrust Fault

Seismic and aseismic processes in elastodynamic simulations of spontaneous fault slip

ESTIMATES OF HORIZONTAL DISPLACEMENTS ASSOCIATED WITH THE 1999 TAIWAN EARTHQUAKE

Development of Procedures for the Rapid Estimation of Ground Shaking Task 7: Ground Motion Estimates for Emergency Response Final Report

Crustal deformation in Taiwan: Results from finite source inversions of six M w > 5.8 Chi-Chi aftershocks

The combined inversion of seismic and geodetic data for the source process of the 16 October,

ON NEAR-FIELD GROUND MOTIONS OF NORMAL AND REVERSE FAULTS FROM VIEWPOINT OF DYNAMIC RUPTURE MODEL

Waveform inversion of the rupture process of the 2008 Wenchuan earthquake based on the near field strong motions and surface rupture data

Numerical Modeling for Earthquake Source Imaging: Implications for Array Design in Determining the Rupture Process

Inversion of Earthquake Rupture Process:Theory and Applications

by A. Ozgun Konca, Sebastien Leprince, Jean-Philippe Avouac, and Don V. Helmberger

Slip distributions of the 1944 Tonankai and 1946 Nankai earthquakes including the horizontal movement effect on tsunami generation

Earthquakes and Seismotectonics Chapter 5

EARTHQUAKE LOCATIONS INDICATE PLATE BOUNDARIES EARTHQUAKE MECHANISMS SHOW MOTION

The Combined Inversion of Seismic and Geodetic Data for the Source Process of the 16 October 1999 M w 7.1 Hector Mine, California, Earthquake

Topography on Earthquake Motions in Sedimentary Basins

Source rupture process of the 2003 Tokachi-oki earthquake determined by joint inversion of teleseismic body wave and strong ground motion data

Outstanding Problems. APOSTOLOS S. PAPAGEORGIOU University of Patras

Chapter 2. Earthquake and Damage

CHARACTERIZING EARTHQUAKE SLIP MODELS FOR THE PREDICTION OF STRONG GROUND MOTION

A Teleseismic Study of the 2002 Denali Fault, Alaska, Earthquake and Implications for Rapid Strong-Motion Estimation

SOURCE PROCESS OF THE 2003 PUERTO PLATA EARTHQUAKE USING TELESEISMIC DATA AND STRONG GROUND MOTION SIMULATION

Triggering of earthquakes during the 2000 Papua New Guinea earthquake sequence

ESTIMATION OF NEAR-FAULT STRONG GROUND MOTIONS FOR KEY ENGINEERING STRUCTURES

Rupture Process of the Great 2004 Sumatra-Andaman Earthquake

Teleseismic waveform modelling of the 2008 Leonidio event

Data Repository: Seismic and Geodetic Evidence For Extensive, Long-Lived Fault Damage Zones

Earthquake stress drop estimates: What are they telling us?

Modelling Strong Ground Motions for Subduction Events in the Wellington Region, New Zealand

Earthquake Focal Mechanisms and Waveform Modeling

AVERAGE AND VARIATION OF FOCAL MECHANISM AROUND TOHOKU SUBDUCTION ZONE

Ground displacement in a fault zone in the presence of asperities

EARTHQUAKE SOURCE PARAMETERS OF MODERATELY EARTHQUAKE IN THE SOUTH EASTERN IRAN BASED ON TELESEISMIC AND REGIONAL DISTANCES

The Three-Dimensional Dynamics of a Nonplanar Thrust Fault

Constraining earthquake source inversions with GPS data: 2. A two-step approach to combine seismic and geodetic data sets

of other regional earthquakes (e.g. Zoback and Zoback, 1980). I also want to find out

Bulletin of the Seismological Society of America, Vol. 81, No. 5, pp , October 1991

Source Complexity of the 1994 Northridge Earthquake and Its Relation to Aftershock Mechanisms

Simulation of earthquake rupture process and strong ground motion

Observation of shear-wave splitting from microseismicity induced by hydraulic fracturing: A non-vti story

Coulomb stress changes due to Queensland earthquakes and the implications for seismic risk assessment

Comparisons of Ground Motions from the 1999 Chi-Chi Earthquake with Empirical Predictions Largely Based on Data from California

TELESEISMIC SOURCE PARAMETERS AND RUPTURE CHARACTERISTICS OF THE 24 NOVEMBER 1987, SUPERSTITION HILLS EARTHQUAKE

Reconstruction of fault slip of the September 21st, 1999, Taiwan earthquake in the asphalted surface of a car park, and co-seismic slip partitioning

Mechanics of Earthquakes and Faulting

Building up Seismsic Models for Ground Motion Prediction of Taiwan: Problems and Challenges

Part 2 - Engineering Characterization of Earthquakes and Seismic Hazard. Earthquake Environment

BROADBAND STRONG MOTION SIMULATION OF THE 2004 NIIGATA- KEN CHUETSU EARTHQUAKE: SOURCE AND SITE EFFECTS

FULL MOMENT TENSOR ANALYSIS USING FIRST MOTION DATA AT THE GEYSERS GEOTHERMAL FIELD

Preliminary slip model of M9 Tohoku earthquake from strongmotion stations in Japan - an extreme application of ISOLA code.

Earthquakes and Earthquake Hazards Earth - Chapter 11 Stan Hatfield Southwestern Illinois College

Synthetic Seismicity Models of Multiple Interacting Faults

Rupture process of the 2007 Chuetsu-oki, Niigata, Japan, earthquake Waveform inversion using empirical Green s functions

Strong Ground Motion Prediction of Future Large Earthquake from Niavaran Fault in Tehran, Iran by Finite Fault Method

Source Description of the 1999 Hector Mine, California, Earthquake, Part I: Wavelet Domain Inversion Theory and Resolution Analysis

GPS Strain & Earthquakes Unit 5: 2014 South Napa earthquake GPS strain analysis student exercise

Rupture behavior of the 1999 Chi-Chi, Taiwan, earthquake slips on a curved fault in response to the regional plate convergence

PEAT SEISMOLOGY Lecture 12: Earthquake source mechanisms and radiation patterns II

Short Note Source Mechanism and Rupture Directivity of the 18 May 2009 M W 4.6 Inglewood, California, Earthquake

FINITE FAULT MODELING OF FUTURE LARGE EARTHQUAKE FROM NORTH TEHRAN FAULT IN KARAJ, IRAN

Estimation of Peak Ground Acceleration for Delhi Region using Finsim, a Finite Fault Simulation Technique

Elastic rebound theory

The Size and Duration of the Sumatra-Andaman Earthquake from Far-Field Static Offsets

Tohoku-oki event: Tectonic setting

Developing ENA GMPE s Using Broadband Synthe=c Seismograms from Finite- Fault Simula=ons

Challenges in earthquake physics and source imaging

Songlin Li 1, Xiaoling Lai 1 Zhixiang Yao 2 and Qing Yang 1. 1 Introduction

C2.2 The physics of Earthquakes

Dangerous tsunami threat off U.S. West Coast

SIMULATING STRONG GROUND MOTION FROM COMPLEX SOURCES BY RECIPROCAL GREEN FUNCTIONS ABSTRACT

STUDYING THE IMPORTANT PARAMETERS IN EARTHQUAKE SIMULATION BASED ON STOCHASTIC FINITE FAULT MODELING

To Editor of Bulletin of the Seismological Society of America

Supplementary Materials for

Source of the July 2006 West Java tsunami estimated from tide gauge records

Di#erences in Earthquake Source and Ground Motion Characteristics between Surface and Buried Crustal Earthquakes

Lower Bounds on Ground Motion at Point Reyes During the 1906 San Francisco Earthquake from Train Toppling Analysis

Dynamic Triggering Semi-Volcanic Tremor in Japanese Volcanic Region by The 2016 Mw 7.0 Kumamoto Earthquake

Earthquake Doublet Sequences: Evidence of Static Triggering in the Strong Convergent Zones of Taiwan

Source model of the 2005 Miyagi-Oki, Japan, earthquake estimated from broadband strong motions

Transcription:

Bulletin of the Seismological Society of America, 91, 5, pp. 1069 1087, October 2001 Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake by Kuo-Fong Ma, Jim Mori, Shiann-Jong Lee, and S. B. Yu Abstract We investigated the rupture process of the 1999 Chi-Chi, Taiwan, earthquake, using high-quality near-source strong-motion records, broadband teleseismic displacement waveforms, and well-distributed Global Positioning System (GPS) data. The near-source strong-motion displacement waveforms recorded significant static offsets of up to 8 m. The teleseismic displacement records show a significant pulse with duration of about 18 to 20 sec. Taking into account the surface displacements observed along the Chelungpu fault, we considered two fault geometries: a single planar fault and a two-segment fault with a northeast-striking section near the northern end. Using the finite-fault model with variable slip vectors, we derived two models of the temporal and spatial slip distribution of the earthquake. The GPS data provided good surface displacement constraints for the slip-distribution determination. The spatial slip distribution is generally consistent with field observations. The results for the simple fault model show a large asperity located in the region about 25 to 55 km north of the hypocenter with maximum slip of about 15 m. When we use the two-segment model, the asperity further extends to the region where the fault bends toward the northeast with a maximum slip of up to 20 m. A large amount of right-lateral slip beneath station TCU068 is necessary to explain its observed large west movement. It implies a local converging slip at the corner where the fault bends to the northeast. The slip amplitude near the hypocenter is about 3 to 6 m. The seismic moments determined from the various data sets are within the range of 2 to 4 10 27 dyne cm. Most of the slip concentrated at shallow depths (less than 10 km). The total rupture duration is about 28 sec, and the rupture velocity is 75% to 80% of the shear-wave velocity. The slip vector shows a clockwise rotation during the fault rupture. The static stress drop of the large asperity region is comparable with the dynamic stress drop, as observed directly from the slip velocity at the station near the large slip region. Introduction The 20 September 1999 Chi-Chi, Taiwan, earthquake (M w 7.6) was the largest earthquake to strike Taiwan in the twentieth century. This earthquake ruptured about 85 km of the Chelungpu fault with complicated surface faulting (Ma et al., 1999). The faulting was mainly thrust movement on a plane that dipped shallowly (20 30 ) to the east. This earthquake produced some of the largest fault displacements (up to 8 m) and largest ground velocities (up to 3 m/sec) observed for a modern earthquake (Chung and Shin, 2000). The distribution of instrumentally determined intensity (Teng et al., 1997) showed shaking levels of 4 or greater over most of the Taiwan region (Ma et al., 1999). There was extensive damage to structures over a large region of central Taiwan, especially in the towns and cities of Chung Liao, Nantou, Fengyuan, and Dengshi. In general, the hangingwall region of the thrust fault showed much stronger shaking than the footwall region. Within 100 m of the fault, the severe damage was often caused more by the ground deformations than by strong shaking. This earthquake was well recorded by the dense strongmotion network of Taiwan (Shin et al., 1999), providing important information on the fault behavior for earthquake source studies, geological investigations, and engineering evaluations. Determinations of earthquake slip distributions and rupture time histories have been important for providing insights into the earthquake rupture process and for prediction of strong ground motions (Wald, 1992). In this study, we combine all of the high-quality data generated by this earthquake to explore the rupture behavior of faulting during the earthquake. We use teleseismic P- wave data, near-field strong-motion recordings, and Global Positioning System (GPS) results in our analysis. This com- 1069

1070 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu bination of data sets enhances the overall spatial sampling of the studied region (Wald and Heaton, 1994). The teleseismic body waves enhance the vertical resolution of the modeling and help constrain the total moment. The nearsource strong-motion data provides time resolution for the faulting process history. And, the GPS data provides the constraints on the final slip distribution. The Chelungpu fault strikes in a nearly north south direction for most of the length of the rupture, but at the northern end the trend of the surface faulting bends toward the east, and the fault splinters into more complicated traces. The largest dislocation and fault-slip velocity were observed near the location where the fault curves to the east. The large fault movement and velocity might be associated with the dynamic behavior of faulting (Ma et al., 2001) and/or may be due to the fault geometry. In our finite-fault modeling, we consider both a simple planar fault model and a twosegment model to examine the spatial and temporal slip distribution of the earthquake. In our models we divided the fault into smaller subfaults. The slip in each subfault is allowed to rupture in multiple time windows, following the methodology of Hartzell and Heaton (1983). Although the sparse slip distribution of the aftershocks make the determination of fault geometry difficult, and no absolute timing on the strong-motion records also yields the difficulty of calibrating the Green s function calculations, with the density of the combined data sets used in this study, we should be able to understand the spatial and temporal rupture behavior of this large earthquake. From our models we can see the heterogeneous distribution of slip on the fault plane and use these results to infer information about the rupture process of the Chi-Chi earthquakes. Data Strong Motion One of the most significant aspects of the Chi-Chi earthquake is the well-recorded strong-motion data along the fault (Fig. 1). The stations along the Chelunpu fault recorded the time history of the ground shaking providing direct observations of the fault ruptures. In the analysis, we use threecomponent accelerograms from 21 strong-motion stations of the Taiwain strong-motion stations (TSMIP) (Liu et al., 1999), as shown in Figure 1. We detrended the accelograms and integrated once to velocity, then we detrended the records again and integrated a second time to produce the displacements. The strong-motion stations TCU68 and TCU052 show large static offsets compared to other stations. These displacement records might be affected during the data processing. These large static offsets were differed by about 20 50 cm to compare with the records obtained from other studies (e.g., Chung and Shin, 1999) using different data processing techniques. Compared to stations TCU068 and TCU052, the waveforms for other strong-motion stations located near the fault trace and in the hanging-wall region are plotted with a 3.4 times larger scale factor. The remaining stations are plotted with a 6.8 times larger scale factor. The numbers on the right of the waveforms indicate the maximum displacement in centimeters. The stations were chosen by considering the quality of data, allowance of computation time, and azimuthal coverage of the fault area. Unfortunately, no near-source stations to the east of the fault area were available, thus, the results for the slip distribution for deeper portions of the fault might be less reliable. Table 1 lists the station abbreviations and locations, as well as site surface conditions (Kuo, 1994). Some of the site conditions are still not well characterized. The stations TCU052 and TCU068, which is located near the location where the fault bends toward the east, recorded large fault displacements and ground velocities. An interesting feature is that the stations with the largest fault displacements and velocities were not accompanied by large accelerations. The stations to the north are dominated by a smooth, large amplitude pulse in velocity, whereas the southern sites have higher frequency energy with lower velocities. Brodsky and Kanamori (2000) explained this feature as the possible lubrication of the fault, which is related to the dynamic behavior of faulting (Ma et al., 2001). In addition, the stations near the fault show significant static offsets. The displacement records to the north, especially for the station TCU068 and TCU052, are dominated by a large static offsets indicating that there were larger amounts of slip in the northern region of the fault. In modeling these data, we did not use any filtering on the displacement records. The displacement data has most of the signal in the frequency less than 1 Hz. Teleseismic Waveforms The teleseismic station locations for the broadband data used in this analysis are listed in Table 2. The station distribution, shown in Figure 2, provides good azimuthal coverage of the source. The instrument responses have been deconvolved from the original recordings to obtain ground displacements (in microns). We modeled the first 90 sec of the P-wave trains, as also shown in Figure 2. The teleseismic P waves show a significant pulse with a duration of about 15 20 sec, indicating that there is a large asperity involved in the rupture process. Global Positioning System The geodetic data for the Chi-Chi earthquake consist of a number of GPS measurements from several institutions in Taiwan. For detailed description on the data collection and processing, see Yu et al. (2000). There are 131 three-component GPS coseismic displacements, which provided dense coverage of the fault area, as shown in Figures 3a and 3b. These measurements were taken before and following the 20 September 1999 Chi-Chi event. The coseismic deformation exhibits large amounts of horizontal slip (Fig. 3a) on the hanging wall, from 1.5 to 8.5 m. The vertical deforma-

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1071 Z N-S E-W Z N-S E-W TCU068 TCU052 458 415 874 732 720 296 TCU046 TCU039 38 47 48 72 61 TCU067 97 103 186 TCU128 32 92 TCU075 TCU129 CHY101 TCU072 15 44 19 142 75 79 254 172 TCU059 128 TCU123 58 TCU141 226 25 73 62 30 72 70 21 19 27 TCU039 TCU046 TCU128 TCU059 TCU068 TCU052 TCU067 TCU123 TCU072 TCU075 TCU074 TCO129 TCU089 TCU079 TCU141 TCU078 CHY024 CHY101 CHY006 CHY074 CHY041 TCU089 37 146 256 CHY024 29 34 56 TCU074 116 139 194 CHY006 13 15 24 TCU079 219 CHY074 10 13 26 TCU078 108 178 CHY041 5 12 10 80 sec Figure 1. Taiwan strong-motion station (TSMIP) distribution (open triangles). The solid line indicates the surface breaks of the Chelungpu fault during the 1999, Chi-Chi earthquake. The solid triangles indicate the 21 stations used in the analysis. The waveforms for strong-motion stations located near the fault trace and in the hanging-wall region are scaled with a factor of 3.4 larger, and other stations are scaled with a factor of 6.8 larger than the scale of station TCU068 and TCU052. The numbers on the right of the waveforms indicate the maximum displacement in centimeters. tion (Fig. 3b) on the hanging wall is from 2.2 to 4.5 m. The amount of deformation increases from south to the north along the fault. The deformation on the footwall is significantly less than that on the hanging wall, which is characteristic of a low-angle thrust fault. The direction of horizontal movement is generally in a northwest direction, but there is a clockwise rotation of the slip from south to north, which is also seen in the static displacements of the near-fault strong-motion records (Ma et al., 2001; Oglesby and Day, 2001). Possible afterslip a few days after the mainshock may be contained in the coseismic displacements, which may cause discrepancies when comparing with the results from the seismic data. The results from the GPS inversion might be less reliable for the spatial slip distribution determination along the fault, which will be discussed later. Using the large amounts of GPS data provides good constraints on the total slip distribution of the fault. However, for stations very close to the fault, the slip vectors may be affected by local variations of the fault strike and dip. Since our fault model does not include these local features in the fault geometry, there is some difficulty in modeling some of the GPS displacements close to the fault. Fault Parameterization Fault Model The surface breaks observed along the Chelungpu fault exhibit a strike of about 3 to 5 and a dip of about 29 to the east (Lee et al., 2000). Near the region of the large surface displacements in the north, the fault trace splinters into several segments and bends about 60 toward the east, as shown in Figure 1. We first try to model the Chi-Chi earthquake with a single planar fault, as shown in Figure 4a. The single fault model has a strike of 3 and dip of 29 to the east, which coincides with most of the observed surface breaks, except for the areas near the end of the fault. The strike of the fault was carefully chosen to match the polarities of the static offsets for stations neat the fault. To consider the bending of the observed fault trace, we also consider a

1072 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu Table 1 Locations and Site Conditions of the Strong-Motion Stations Used in This Study Station Code Latitude ( N) Longitude ( E) Soil Type TCU039 24.50 120.79 S2 TCU046 24.47 120.86 S1 TCU052 24.20 120.74 TCU059 24.27 120.56 TCU067 24.09 120.72 TCU068 24.28 120.77 TCU072 24.04 120.85 TCU074 23.96 120.96 TCU075 23.98 120.68 TCU078 23.81 120.85 TCU079 23.84 120.89 TCU089 23.90 120.86 TCU123 24.02 120.54 TCU128 24.42 120. TCU129 23.88 120.68 TCU141 23.83 120.46 CHY006 23.58 120.55 S2 CHY024 23.78 120.61 CHY041 23.44 120.60 S2 CHY074 23.51 120.81 S2 CHY101 23.69 120.56 S1, rock; S2, dense soil composition similar to rock. Table 2 Parameters of the Teleseismic Stations Used Station Longitude Latitude Delta ( ) Azimuth Backazimuth ADK 1.68 51.88 54.6 42.29 264.30 AFI 171.78 13.91 75.5 112.25 299.25 BORG 21.32 64.75 87.1 344.75 34.160 CMB 120.38 38.04 95.7 43.96 306.41 COLA 147.85 64.879 69.2 27.12 281.89 COR 120.30 44.59 90.1 39.92 304.70 CTAO 146.26 20.09 50.0 148.28 329.19 DUG 112.81.20 99.0 38.55 311.88 ESK 3.21 55.32 88.1 331. 49.28 FFC 101.98 54.73 93.5 23.20 321.57 KBS 11.94 78.93 70.4 348.81 66.92 KDAK 152.58 57.78 68.3 35.06 280.63 KIEV 29.21 50.69 73.1 318.38 73.11 KIP 158.02 21.42 73.7 73.27 289.78 KONO 9.60 59.65 79.9 331.30 60.03 NEW 117.12 48.26 91.4 34.42 309.24 PAB 4.35 39.55 99.0 320.34 49.06 PAS 118.17 34.15 99.5 46.00 307.44 PMG 147.15 9.41 44.4 139.01 322.52 RAYN 45.50 23.520 68.4 287.00 72.54 SFJ 50.62 67.00 89.3 356.65 7.82 TAU 147.32 42.91 70.6 159.65 344.30 two-segment fault model as shown in Figure 4b. This second model has a change in strike that turns 60 toward the east, near the northern end of the fault rupture. The dip angle is kept at 29. This two-segment model results in a nonplanar fault model, and we treat each of the two segments as an independent fault plane. For a single fault model, we discretized the fault plane into a total of 168 subfaults (Fig. 4a). For a two-segment fault model, the fault plane was discretized into 164 subfaults (Fig. 4b). Each subfault has dimensions of 5 km 5 km. The number of subfaults was chosen visually by considering the resolvable spatial distribution of the variable slip along the fault for the station distribution. Synthetic Green s Functions The point-source responses for the strong-motion synthetics are computed using a layered velocity structure (Table 3) with a frequency-wavenumber (f -k) integration scheme (Bouchon, 1981; Takeo, 1987) for frequencies between 0 and 1 Hz. The velocity model is adopted from the tomography study of Ma et al. (1996) for the southwestern Taiwan region. A thin, slow layer is included near the surface in this model, which is important to better approximate elastic properties near the strong-motion stations. Although the site conditions of some of the strong-motion stations are not well characterized (Table 1), for the relatively lowfrequency band used in this study, the observed waveforms of most of the chosen stations seem to be quite adequate for the analysis without considering the possible site effects. To allow for the variations in direction of the fault slip during the rupture, we calculated the Green s functions for a pure thrust and left-lateral mechanisms. Calculations were done for sample intervals of 0.1 sec and then decimated to 0.4 sec for the inversion. For the subfaults near the stations TCU068 and TCU052, where large static offsets were observed, the subfault motions were obtained by summing the responses of 100 point sources equally distributed over the 5 km 5 km subfault, as shown in Figure 4. The travel-time differences resulting from the varying source-to-station positions and simulating of the propagation of the rupture front across each subfault are considered in each Green s function calculation. The generalized ray method of Langston and Helmberger (1975) is utilized to compute the point-source response of the teleseismic P wave. The responses of all rays with up to two internal reflections in a layered velocity model (Table 4), plus the free surface and internal phase conversations, are included in the calculation. A Q operator (Futterman, 1962) is applied with the attenuation time constant t* equal to 1.0 sec for P waves. For the geodetic modeling, we utilized the analytic expressions of Okada (1985) to calculate the horizontal and vertical static displacements for surface deformation resulting from a uniform slip over each subfault. A Poissonian half-space model is assumed in the calculation. Source Time Function Wald and Heaton (1994) modeled the 1992 Landers, California, earthquake using multiple segments and considered multiple time window source time functions to allow for the possibility of delayed or triggered rupture initiating. The number of time windows depends on the overall dislo-

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1073 Z-component ADK 0.63 KDAK 0.22 CTAO 1.30 PAS 0.17 AFI 0.67 KIEV 0.66 DUG 0.01 PMG 1. BORG 0.15 KIP 0.61 ESK RAYN 0.17 0.21 CMB 0.22 KONO 0.59 FFC 0.10 SFJ 0.35 COLA 0.69 NEW 0.03 KBS 0.65 TAU 0.31 COR 0.30 PA B 0.17 90 sec. Figure 2. The distribution of the stations and teleseismic displacement waveforms used in the analysis. The number on the right of the seismograms indicate the amplitude in microns. cation duration of the earthquake. For the surface rupture of about 100 km of M W 7.6 Chi-Chi earthquake, we thought it was reasonable to allow each subfault to slip in any of 12 adjacent 1-sec time windows. The time of the initial window is given by the arrival of the rupture, which is assumed to be 2.5 km/sec. The source time function of each time window is represented by a triangle with a duration of 1 sec. Inversion Method We applied a positive constrained, damped, linear leastsquares inversion procedure as used by Hartzell and Heaton (1983) to obtain the subfault dislocations that give the best fit to the displacement waveforms and geodetic data. We utilized a minimum curvature technique on the inversion to consider the minimum difference of dislocation values between adjacent subfaults in each time window. We solve an overdetermined system of linear equations, AX B (1) where A is an N M matrix of subfault synthetics (Green s functions), X is the solution vector of slip amplitude with

(a) (b) 5m Figure 3. Coseismic (a) horizontal and (b) vertical displacements from geodetic data. These measurements were taken before and following the 20 September 1999 Chi-Chi event (Yu et al., 2000). The solid lines indicate the surface break of the Chelungpu fault. The asterisk indicates the epicenter. 5m 1074

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1075 (a) Single fault model (b) Two-segment Fault Model 2 2 2 2 2 2 Figure 4. (a) Single-fault and (b) two-segment fault models used in the anlalysis. Total of 168 and 164 subfaults are used in the analysis, respectively. The subfault has the dimension of 5 km 5 km. The points indicate the center of the subfault. The solid triangles indicate the strong-motion stations used in the analysis. The subfaults near stations TCU068 and TCU052 are divided into 100 subfaults as shown by the dense solid dots for the Green s functions calculations. The solid lines indicate the surface break of the Chelungpu fault. The asterisk indicates the epicenter. Table 3 Velocity Structure Used in the Strong-Motion Waveforms Analysis V P (km/sec) V S (km/sec) Thickness (km) 3.50 2.00 1.0 3.78 2.20 3.0 5.04 3.03 5.0 5.71 3.26 4.0 6.05 3.47 4.0 6.44 3.71 8.0 6.83 3.99 5.0 7.28 4.21 Half-space Table 4 Velocity Structure Used in the Teleseismic Waveforms Analysis V P (km/sec) V S (km/sec) Density (g/cm 3 ) Thickness (km) 5.60 2.90 2.50 6.1 6.00 3. 2.60 12.9 6.80 4.00 3.00 30.0 length M, B is the data vector containing the displacement seismograms of strong motion, the teleseismic data, and the three-component GPS data. The size of N is the product of number of stations, number of components, and the number of points in each seismogram. The number of GPS data points is one per component, while the seismic data has over 100 time points for each component seismogram, so extra weighing of the GPS data is need for it to contribute significantly to the solution. The variable M is the product of the number of subfaults, components of slip directions, and number of time windows. We considered both left-lateral strike-slip and thrust movements in the slip-vector determination, which allows for spatial and temporal changes of rake angle during the fault rupture. The variance defined as (AX B)/(n 1), where n is the number of unknowns, is used to quantify the fit of the synthetics to the data. In the inversion, we used 80 sec of 21 three-component strong-motion displacement seismograms, 22 vertical-component teleseismic seismograms, and 131 three-component GPS displacements. The time interval of the seismograms is 0.4 sec. In total, the data dimension is 17,085 for seismic data and 393 for GPS data. The seismic data at each station is normalized to have equal weight for each station. The relative weighting of GPS data to seismic data is difficult to set a priori, and we tested several weightings in the joint inversions. The number of subfaults is 168 for the simple single-fault model and 164 for two-segment model. For 12 time windows and two components of slip vector, the number of unknowns in the model space is 32 and 3936, respectively, for the single-fault and two-segment fault models. To yield a smooth-slip model, a damping parameter is included in the linear equations as A b x. (2) kh 0 H is a matrix of smoothing constraints, which minimizes the slip difference between adjacent subfaults both along strike and downdip. The damping ratio k controls the trade-off

10 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu between the model roughness and fitting of the data. In general, a rougher model fits the seismic data better. As discussed by Cohee and Beroza (1994), the smoothing parameter controls the simple trade-off between model roughness and the fit to the seismograms. Several damping parameters were tested in the calculations for comparison, and we discuss the possible effects of the damping on the results. Also, as mentioned by Wald and Heaton (1994), there are many parameters in our models, and it is not practical to explicitly invert for all of them. These parameters include such variables as the number of subfaults, number of time windows, rise time, rupture velocities, and GPS weighting and smoothing weighting. Determining all of these parameters will result in nonuniqueness of the solution due to the limitation on the density of the data and computer speed. However, numerous inversions were performed during the study, and the slip distributions shown subsequently are representative of the general features common to most of the results. Since we had large amount of GPS data over the faulting region, we did not use the observed surface displacements on the fault as data in the inversion. The GPS data and static offsets of the near-source strong-motion displacement waveforms provided very good constraints for the slip near the surface. The static data is also very crucial in establishing the time dependence of the rupture. We will show that the derived slip distribution results are in very good agreements with field observations of the surface breaks along the fault. Results Figure 5. (a) Trade-off curves of the variance (solid line) and seismic moment (dashed line) for various damping ratios (k), (b) spatial slip distributions for k 0, and (c) k 0.1 from the inversion of strong-motion and teleseismic waveforms. The red arrows indicate the slip vectors on the fault plane. The contour lines indicate the slips of every 2 m. The color bar indicates the amount of slip within the fault plane. The asterisk indicates the hypocenter. Single-Fault Model For the single-fault model, we first consider the joint inversion using the strong-motion and teleseismic displacement waveforms. Figure 5a shows the curve for the variance and total seismic moment over the fault region for various smoothing constants. A larger smoothing constant reduces the amount of seismic moment but increases the data variance. For smoothing constants of 0.1, we obtain an optimum solution. Figure 5b and c shows the slip distributions without (k 0) and withsmoothing constraints (k 0.1), respectively. The results without smoothing constraints have a slightly rougher slip distribution with maximum slip of about 16 m but show a similar slip distribution pattern to the results with smoothing constraints (k 0.1) with maximum slip of about 15 m. The seismic moment is about 11% different for the results for k 0 0.1. This implies that the effect of the damping ratio does not significantly control the general conclusions. The spatial slip distribution over the fault plane shows that most of slip concentrates in the region shallower than the hypocentral depth of 8 km (downdip distance about 15 km). Most of the slip occurred to the north of hypocenter with lesser amount of slip to the south of the hypocenter. The slip near the hypocenter region is about 3 m. The amount of slip increases substantially as the rupture propagates to the north, 20 30 km from the hypocenter. A large asperity is found in the region 25 km to 55 km north of the hypocenter. The maximum slip is about 15 m. The slip vector shows a clockwise rotation of the fault movement during the rupture, consistent with the results of Oglesby and Day (2001). The overall slip distribution is generally consistent with the field observations of the surface break, as shown in Figure 5c (Lee et al., 2000). The slip distribution also shows

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1077 significant displacement in the left corner and bottom of the fault, which might be due to inversion instability because of less station coverage over the eastern part of the fault area. However, the slips at the deeper portion and left corner of the fault did not have significant contribution (less than 1%) to the fitting of the waveforms. The seismic moments from these regions contribute about 20% of the total seismic moment. Figure 6 shows the comparison of observations to the synthetics for the 21 three-component strong-motion and 22 vertical-component teleseismic data. In general, the synthetics match the observation fairly well, however, a number of the east west horizontal components for the stations near the faults are not well matched by the synthetics. Some stations, such as TCU067, TCU075, and TCU129 are located near the kink of the observed fault trace, as shown in Figure 4, and the discrepancy in polarity between the observed and synthetics might be due to the change of fault strike near the station. As mentioned earlier, this study intends to obtain the general features of the fault-slip distribution over the fault area, and local variations of fault strike are not considered in our fault model. Most other stations show synthetic waveforms that are similar to the observations. The stations to the north of the fault (TCU046, TCU039, and TCU128) show less consistency of synthetics to the observations at the north south component. This discrepancy might also be due to changes in strike of the of the fault trace near the northern end of the fault. This feature is improved when we use the two-segment fault model, as discussed later. For the teleseismic waveforms the synthetics match the data well, both in amplitudes and polarities. Next, we consider a combined inversion using the strong-motion and teleseismic seismic data, along with the GPS data. Figure 7a shows the slip distribution for the com- TCU068 Z N-S E-W Z N-S E-W Z Z 874 458 TCU046 38 47 48 0.63 720 ADK KDAK 0.22 TCU052 415 732 296 TCU039 72 61 AFI 0.67 KIEV 0.66 TCU067 97 103 186 TCU128 32 92 BORG 0.15 KIP 0.61 TCU075 15 75 172 TCU059 25 73 62 CMB 0.22 KONO 0.59 TCU129 44 128 TCU123 30 72 70 COLA 0.69 NEW 0.03 CHY101 TCU072 19 142 79 254 58 226 TCU141 21 19 27 COR CTAO 0.30 PAB 1.30 PAS 0.17 0.17 TCU089 37 146 256 CHY024 29 34 56 DUG 0.01PMG 1.4 TCU074 116 139 194 CHY006 13 15 24 ESK 0.17RAYN 0.21 TCU079 219 CHY074 10 13 26 FFC 0.10 SFJ 0.35 TCU078 108 178 CHY041 5 12 10 KBS 0.65 TAU 0.31 80 sec 90 sec Figure 6. Comparison of synthetics (dashed lines) and observations (solid lines) for 21 three-component strong-motion and 22 vertical teleseismic P waveforms from the spatial-slip distribution of strong-motion and teleseismic data for the single-fault model. The numbers on the right of the waveforms indicate the maximum displacement amplitudes in centimeters for strong-motion stations and in microns for teleseismic stations. The station names are shown on the left of the waveforms.

1078 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu Figure 7. The spatial slip distribution for (a) k 0 and (b) k 1 from the combined inversion of strong-motion, teleseismic, and GPS data. The red arrows indicate the slip vectors on the fault plane. The contour lines indicate the slips of every 2 m. The color bar indicates the amount of slip within the fault plane. The asterisk indicates the hypocenter. bined inversion for smoothing parameters k 0 and 1, respectively. The values of linear weighting on the GPS and smoothing parameters (k) were determined on a trial-anderror basis. To fit the observed geodetic data, a GPS weighting of 20 times the weighting of the seismic data is applied in the combined inversion. The combined results without smoothing constraints show a rougher slip distribution than the result from using only the seismic data. This might be due to the denser distribution of GPS data in the fault area. The slip distribution of the combined inversion for a smoothing constant of 1 shows more continuous slip distribution over the fault plane. The slip does not concentrate within shallow region less than the hypocentral depth, but continuously extends to greater depths. The slip near the hypocenter is greater than 6 m. A large asperity is still found in the region 25 to 45 km north of the hypocenter. The maximum slip is about 14 m. Figure 8 shows the comparison of the synthetics to the observed displacement waveforms for the strong-motion and teleseismic data. In general, the synthetics from the combined inversion show waveforms that do not fit as well to the observed data, compared to the results of using only the seismic data. This is not surprising since the GPS data produces more rigid constraints on the slip distribution. However, it does indicate there might be some discrepancy between the static deformation field and the elastic slip distribution that produces the seismic waves. In general, the synthetic waveforms from the combined inversion show larger amplitudes than the results from seismic data. The synthetics are significantly larger than the observed displacement waveforms for vertical component at stations located on the hanging wall of the fault, such as TCU072, TCU089, TCU079, and TCU078. Figure 9a and b shows both the horizontal and vertical displacement vectors, predicted from the slip model of the combined inversion. Except for the stations near the northern end of the fault, the amplitudes of the displacement vectors generally fit the observation very well. The directions of the displacement vectors have some discrepancies for the stations near the fault trace. This might be due to local variations of fault strike near the stations, which are not included in our modeling. The synthetic displacement vectors on the hanging wall show very good agreements to the observations, however, the seismic data for the stations on the hanging wall, as described earlier, were not well explained by the slip distribution of the combined inversion. One possibility for the discrepancy between the results of seismic and combined inversion might be due to inadequate modeling of the GPS data using a half-space model and also to the effect of local fault geometry for the dense GPS data close to the fault. Two-Segment Fault Model Figure 10 shows the slip distribution from the inversion of strong-motion and teleseismic data for the two-segment fault model with k 0.1. Compared with the single-fault model, this slip distribution extends the large slip region through the northeastern bend of the fault. The large asperity is located about 30 to 65 km north of the hypocenter region with width of about 10 km. The average displacement in that region of large slip is about 9 m. The maximum slip is about 12 m. If we consider the overlap region of the two fault segments, the total slip at that region exceeds 12 m. Another asperity is found in the hypocenter region with average slip of about 3 m. Significant slip near the bottom of the fault and at the left corner of the fault are also found. Figure 11 shows the comparison of the synthetics to the observations. The synthetics, in general, can explain the observation very well both for strong-motion and teleseismic waveforms. The synthetics for the strong-motion stations to the north of the fault region (TCU046, TCU039, and TCU128) now fit the observations much better in terms of the amplitudes and polarities. In addition, the static offsets observed at stations TCU068 and TCU052 can also be explained better, especially for the north south component. For the stations near

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1079 TCU068 Z N-S E-W Z N-S E-W 458 874 720 TCU046 47 48 38 ADK Z 0.63 KDAK Z 0.22 TCUO52 415 732 296 TCU039 72 61 AFI 0.67 KIEV 0.66 TCU067 TCU075 TCU129 97 15 44 103 75 186 172 128 TCU128 TCU059 32 25 92 73 62 BORG CMB COLA 0.15 KIP 0.22 KONO 0.69 NEW 0.61 0.59 0.03 TCU101 TCU072 19 142 79 254 58 226 TCU123 TCU141 30 21 72 19 70 27 COR CTAO 0.30 PAB 1.30 PAS 0.17 0.17 TCU089 37 146 256 CHY024 29 34 56 DUG 0.01PMG 1.4 TCU074 116 139 194 CHY006 13 15 24 ESK 0.17RAYN 0.21 TCU079 219 CHY074 10 13 26 FFC 0.10 SFJ 0.35 TCU078 108 178 CHY041 5 12 10 KBS 0.65 TAU 0.31 80 sec Figure 8. Comparison of observations (solid lines) and synthetics (dashed lines) for 21 three-component strong motion and 22 vertical teleseismic P waveforms from the spatial slip distribution of strong motion, teleseismic, and GPS data for the single-fault model. The numbers on the right of the waveforms indicate the maximum displacement amplitudes in centimeters for strong-motion stations and in microns for teleseismic stations. The station names are shown on the left of the waveforms. the fault, the synthetics of the east west components still underestimate the observed data. Figure 12 shows the slip distribution from the joint inversion of strong-motion, teleseismic, and GPS data for the two-segment fault model. Again, compared with the single fault model, the two-segment fault model extends the large asperity to the northeastern-striking segment. Most of large slip still concentrates at shallow depths ( 10 km). The slip distribution from the combined inversion shows large slip in the region of 25 km to 70 km north of the hypocenter. The maximum slip in north south striking fault segment is about 12 14 m. The northeastern-striking segment has a maximum slip of about 20 m, which is much larger than the slip obtained in previous calculation without GPS. It implies that the large slip in the northeast segment is necessary to explain the observed GPS data. The results from the inversion including the GPS data show continuous slip that extends to deeper portions of the fault. This might be because the GPS data have more station coverage in the hanging-wall region than the strong-motion stations, thus providing more constraints than the strong-motion data in that region. Figure 13 shows the comparisons of synthetics to the observed for the strong-motion and teleseismic data. Figure 14 shows the comparisons of the geodetic data. The synthetics from the combined seismic and GPS inversion do not seem to match the data very well for the east west components of stations near the fault, especially for station TCU068. This station is located near the corner of the two fault segments, so the observed displacements may be very sensitive to the local geometry. However, the synthetic geodetic vectors explain the observed geodetic vector rather well. Compared to the results from the single fault, the model can now explain the

1080 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu Figure 9. Comparisons of the synthetics (red arrows) and observations (blue arrows) for (a) horizontal and (b) vertical coseismic displacement from GPS data for the singlefault model. The solid lines indicate the surface break of the Chelungpu fault. The asterisk indicates the epicenter. large vectors near the northeastern segment of the fault. However, there are still some discrepancies for the GPS stations located near the northeastern end of the fault. Hsu et al. (2000) modeled the GPS data and suggested that another fault segment with right-lateral slip is necessary to explain the observed GPS data. This result suggests the possible complicated faulting when the fault bends toward northeast. The discrepancy on the polarities of synthetic to the observed at east west component of station TCU068 might also suggest the possible right-lateral slip behavior near the station. Discussions Figure 10. Spatial slip distribution from the inversion of strong motion and teleseismic for twosegment fault model for k 0.1. The red arrows indicate the slip vectors on the fault plane. The contour lines indicate the slips of every 2 m. The color bar indicates the amount of slip within the fault plane. The asterisk indicates the hypocenter. The dashed line indicates the fault segment, which strikes to the northeast. Temporal Slip Distribution For the temporal slip distribution analysis we consider the results of the joint inversion of teleseismic and strongmotion data for the two-segment fault model. These results show that most of the slip occurred with the initial passage of the rupture front, so that most of the slip is in the first time window. There are lesser amounts of slip in the middle of the fault at about 10 to 20 km north of the hypocenter. The large asperity 30 to 65 north of the fault has a duration of about 6 sec. The large amount of total slip and a 6-sec

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1081 Figure 11. Comparison of observations (solid lines) and synthetics (dashed lines) for 21 three-component strong motion and 22 vertical teleseismic P waveforms from the spatial slip distribution for the model derived from strong motion and teleseismic data for two-segment fault model. The numbers on the right of the waveforms indicate the maximum displacement amplitudes in centimeters for strong motion stations and in microns for teleseismic stations. The station names are shown on the left of the waveforms. duration yields an average slip velocity of about 2.5 m/sec for that region. Although the slip near the hypocenter is not large compared with the asperity on the northern part of the fault, the slip duration near the hypocenter is longer with duration of almost 12 sec, indicating a much slower slip velocity. This result suggests different rupture behaviors for at the northern and southern portions of the fault. Considering an average slip of 9 m in the large asperity region with dimension of 30 10 km as shown in Figure 10, the static stress drop in the large asperisty region is about 225 bars for a dip-slip model for k l 3 10 11 dyne/cm 2. We also estimated the slip velocity of the fault directly from the three-component velocity records of the near-field strong-motion stations. For the stations very close to the fault, the near-source motions dominate, and recorded ground velocities can be used as a direct measure of the movement of the nearby portion of the fault. The analysis shows that the peak slip velocities increase from south to north consistent with our temporal slip distribution. The slip velocities at the stations of TCU052 and TCU068 are 2.3 m/ sec and 4 m/sec, respectively. The slip velocity at station TCU052 is very comparable with the estimated slip velocity of about 2.5 m/sec of the large slip region. Considering the l dynamic stress drop as given by Drd U, where l and b b are the rigidity, 3 10 11 dyne/cm 2, and S-wave velocity, 3 km/sec, respectively, and U is the slip velocity, the slip velocities at station TCU052 and TCU068 yield the dynamic

1082 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu Figure 12. Spatial slip distribution from the inversion of strong motion, teleseismic and GPS data for two-segment fault model for k 1. The red arrows indicate the slip vectors on the fault plane. The contour lines indicate the slips of every 2 m. The color bar indicates the amount of slip within the fault plane. The asterisk indicates the hypocenter. The dashed line indicates the fault segment, which strikes to northeast. stress drops of about 230 and 0 bars, respectively. The station TCU068 is located at the corner of the bending fault. The observed waveforms are more difficult to model as shown in previous calculations. Detail discussion about this station will be addressed in the later session. However, the dynamic stress drop of 230 bars estimated from the station TCU052 is very comparable with the estimated static stress drop of 225 bars at the large slip region from the inversion. This similarity implies a drop of stress to the frictional stress level in the large slip region. In addition, the increase of slip velocities from the hypocenter to the region of large slip in the north implies an increase of the tectonic stress or a decrease in dynamic friction during the rupture. Brodsky and Kanamori (2001) proposed that a fault can be lubricated by a process similar to the mechanical lubrication in bearings, where the viscous resistance of a fluid in a thin zone between opposing surfaces generates a significant pressure that reduces the normal stress and therefore reduce the friction. The dynamic rupture behavior of the slip at the large slip region might be able to be explained by this lubrication behavior. The large slip velocity in the northern portion of the fault increased the lubrication pressure within the fault surface. As the lubrication pressure increases, the fault gap widens and results in a drop of friction stress to generate large slip velocity. The behavior of the large slip velocity and lubrication pressure, thus, acts as a self-controlled system (Kanamori and Heaton, 2000). The widening of the fault gap results in a smaller amount of surface area for the asperities in contact. The reduced number of asperities in contact might, thus, result in a smoother fault, generating a smaller amount of high-frequency acceleration, large slip velocity, and displacement, as observed at stations TCU052 and TCU068. Figure 15 shows the time slice of the fault rupture at every 1 sec. The initiation of the fault rupture is rather quiet during the first few seconds. The fault ruptures bilaterally in the beginning, then after about 12 sec, the fault ruptures more unilaterally to the north. At about 19 sec after the initiation, the slip is concentrated in the region of 20 to 65 km north of the hypocenter and lasts for about 7 8 sec without propagation. This implies a slowing of the rupture velocity near the end of the fault rupture. The consistency of the overall spatial slip distribution and the field observations suggests the reliability of the rupture velocity used in this study. The average rupture velocity is about 2.5 km/sec. The rupture front as shown in Figure 15 shows the variation of the rupture velocities. However, it is difficult to quantitatively estimate the rupture velocity from the visual inspection. Chen et al. (2000) suggests a variation of the rupture velocity from 2.7 km/sec to 2.28 km/sec. Huang et al. (2000) also suggested a slow rupture velocity of about 2.0 km/sec at the northern portion of the fault. Seismic Moments The seismic moment determined from Harvard CMT solution is about 4.1 10 27 dyne-cm, while the solution from the Earthquake Research Institute, University of Tokyo (ERI), and the U.S. Geological Survey (USGS) are 2.5 10 27 dyne cm and 2.4 10 27 dyne cm, respectively. Table 5 lists the seismic moments, variances of the synthetic to the data, and the maximum slip of each fault model. The seismic moment determined from our results of the inversion of teleseismic and strong motion is about 2.2 10 27 dyne cm for the single fault model and 2.3 10 27 dyne cm for the two-segment fault model. If we include the GPS data, the seismic moment increases to about 4.6 4.7 10 27 dyne cm. The fit of the synthetics to the data did not have too much difference for single or two-segment fault models. However, if the GPS data were included in the calculation, the data became less well explained by the synthetics. The maximum slip in each model ranged from 12 to 15 m. For the two-segment fault model, a maximum slip of 20 m was obtained at the northeast fault segment, if the GPS data were included in the calculation. The amount of seismic moment will also vary with a different damping parameter. As shown in Figure 5a, the trade-off between seismic moment and damping is about 0.5 10 27 dyne cm for damping ratio of 0 to 1. The discrepancy between the seismic moments with and without the GPS data might be due to the overestimate of the slip because of our use of half-space modeling for GPS data. Wald and Graves (2000) suggest that using a half-space to model the deformation data might generate models with larger amount of slip. Lee and Ma (2000) modeled the GPS data and showed that the dense distribution of the GPS data is very sensitive to the 3D structure of the fault geometry, including the variation of fault strike and dip on the fault plane. The simple 2D fault geometry used in this

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1083 TCU068 Z N-S E-W Z N-S E-W Z Z 458 874 720 TCU046 47 48 ADK 0.63 KDAK 38 0.22 TCUO52 415 732 296 TCU039 72 61 AFI 0.67 KIEV 0.66 TCU067 TCU075 TCU129 TCU101 TCU072 97 15 44 19 142 103 75 79 254 186 172 128 58 226 TCU128 TCU059 TCU123 TCU141 32 25 30 21 92 73 72 19 62 70 27 BORG CMB COLA COR CTAO 0.15 KIP 0.22 KONO 0.69 NEW 0.30 PAB 1.30 PAS 0.61 0.59 0.03 0.17 0.17 TCU089 37 146 256 CHY024 29 34 56 DUG 0.01PMG 1.4 TCU074 116 139 194 CHY006 13 15 24 ESK 0.17RAYN 0.21 TCU079 219 CHY074 10 13 26 FFC 0.10 SFJ 0.35 TCU078 108 178 CHY041 5 12 10 KBS 0.65 TAU 0.31 80 sec Figure 13. Comparison of observations (solid lines) and synthetics (dashed lines) for 21 three-component strong-motion and 22 vertical teleseismic waveforms from the spatial-slip distribution of strong-motion, teleseismic, and GPS data for two-segment fault model. The numbers on the right of the waveforms indicate the maximum displacement amplitudes in centimeters for strong-motion stations and in microns for teleseismic stations. The station names are shown on the left of the waveforms. study might not be sufficient enough to well explain the observed GPS data. For further studies, for the GPS data, a layer structure for Green s function calculations and detail 3D fault geometry might be necessary to explore this possibility. Mismatch of Displacements Close to the Fault For our models, there is not a very good match of the synthetic displacements for strong-motion stations close to the fault. Part of the explanation may be because these records are sensitive to the local fault geometry close to the station, which is not accounted for in our rather simple models, especially for the effect of 3D fault geometry. There is a significant mismatch of displacement close to the fault at station TCU068 in our modeling. Ji et al. (2001) modeled the static displacement close to the fault and GPS data using layer velocity structure with dense Green s function calculations. Their results show good consistency with the GPS data and strong-motion static displacements. These results imply the necessity of the layer structure calculation of GPS data, especially for the near-source data modeling. In addition, they also show that the slip of the earthquake was concentrated on the surface of a wedge-shaped block. The static displacement at the station TCU068 was dominated by the extended northeastern-striking fault with large component of right-lateral slip. Figure 16 shows our calculation, for considering a right-lateral slip instead of left-lateral slip in our previous modeling for the block near the corner of the two-segment fault, the static displacements at the stations TCU052 and TCU068 can now be explained very well. It suggests the significant requirement of the right-lateral slip in that block. Thus, as the fault bending to the northeast, there s a local converging of the slip at the corner of the bending fault. This behavior might be associated with local stress and tectonic environment. However, the detailed rupture behavior of the earthquake over the fault plane still need

1084 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu Figure 14. Comparisons of the synthetics (red arrows) and observations (blue arrows) for (a) horizontal and (b) vertical coseismic displacement from GPS data for the two-segment fault model. The solid lines indicate the surface breaks of the Chelungpu fault. The asterisk indicates the epicenter. a detail 3D geometry to explain the overall observed data. Nevertheless, the well explanations of the observed waveforms further from the fault surface suggest that we had obtained a general feature of the fault rupture; while the stations close to the fault make significant contributions on the understanding of the detail fault rupture, especially at the complex surface rupture region. Comparison with Aftershock Distribution Figure 17 shows the distribution of aftershocks for six months following the mainshock. An interesting feature is that most of the aftershocks appear not to have occurred directly on the Chelungpu fault. There are six aftershocks with magnitude larger than 6.0 located to the east of the epicenter. To compare the aftershock distribution to the spatial slip distribution determined from the combined data sets for the two-segment fault model, we projected the slip distribution to the surface as also shown in Figure 16. Most of the aftershocks surround the areas of large slip. The asperity near the epicenter has less slip compared with the region north of the epicenter. Most of large aftershocks (M 6) occurred near the epicentral region rather than to the north. The spatial slip distribution shows areas of relatively low slip about 10 20 km north of the epicenter. The aftershock distribution cuts through this area and it is also the only cluster of the aftershocks near the fault trace. This suggests that most of the stress was released in the asperity regions, so that no aftershocks occurred after the mainshock. This is similar to the pattern of aftershocks in relation to the areas of large slip observed in other large earthquakes, (e.g., the 1992 Landers earthquake) (Wald and Heaton, 1994). Conclusions The 1999 Chi-Chi, Taiwan, earthquake is the best recorded large damaging earthquake in terms of the local strong-motion and GPS data. We investigated the rupture process of the earthquake, using the combination of three independent data sets, high-quality near-source strongmotion records, broadband teleseismic displacement waveforms, and well-distributed GPS displacements. We estimated the spatial and temporal distribution of slip on the fault plane using finite source inversions that used several combinations of the data and several different parameterizations for the model. Common features of our various results show that most of the slip was concentrated at shallow depths (less than 10 km) with a large asperity north of the hypocenter with maximum displacements of 12 to 15 m. The slip near the hypocenter ranges from about 3 to 6 m. The slip vector shows a clockwise rotation during the fault rupture The seismic moments determined from the various data sets are within the range of 2.0 to 4.5 10 27 dyne cm. The total rupture duration is about 28 sec, and the rupture velocity is about 75% to 80% of the shear-wave velocity.

Spatial and Temporal Distribution of Slip for the 1999 Chi-Chi, Taiwan, Earthquake 1085 Figure 15. Snapshot of the slip propagation at every 1 sec. The numbers on the left of each panel indicate the time of the rupture. The color indicates the amount of slip as shown in the color bar. The dashed lines indicate the grids at a 5-km interval. The asterisk indicates the hypocenter. Table 5 Seismic Moments, Variances, and Maximum Slips of the Models Fault Model Single Fault Two-Segment Fault Data* SM Tele SM Tele GPS SM Tele SM Tele GPS Moment (dyne cm) *SM, strong-motion data; Tele, teleseismic data. Maximum slip at northeast fault segment. Variance 2.2 10 27 4.7 10 27 0.045 9.58 2.3 10 27 4.6 10 27 0.043 5.23 Maximum Slip (m) 15 14 12 (8) 14 (20) Our preferred model using a two-segment fault shows that the region of large slip extends through the area where the fault bends to the northeast. A large slip of 20 m was obtained in the northeast fault segment to explain the observed GPS data. There is a local converging slip at the location where the fault bends toward northeast due to the requirement of a right-lateral slip in that region to explain the observed large western movement in the observed waveform of TCU068. Although the simple 2D fault geometry used in this study might not be sufficient to well explain all of the observed data, and there are differences among the models, which may be attributed to practical factors, such as data

1086 K.-F. Ma, J. Mori, S.-J. Lee, and S. B. Yu TCU068 TCU052 458 874 415 732 Figure 16. The comparison the synthetics (dashed line) and observed (solid lines) displacement waveforms of stations TCU068 and TCU052. The numbers on the right of the waveforms indicate the maximum displacement amplitudes in centimeters. Figure 17. The three-month aftershock distribution (solid circles) and the spatial-slip distribution from strong-motion and teleseismic data for the twosegment fault model. The slip distribution had been projected to the surface for comparison. The amount of slip is indicated by the color bar. The bold solid lines indicate the surface breaks of the Chelungpu fault. The asterisk indicates the hypocenter. 720 296 quality, data coverage, and model parameterization, our fault models exhibit the common feature of the spatial and temporal slip distribution during the fault rupture. Acknowledgments The authors would like to send our appreciation to the Central Weather Bureau for providing quick strong-motion data used in this study. We also thank the comments from two reviewers: David Wald and Steven Day of making the manuscript more complete. The discussions with H. Kanamori, T.-R. A. Song, and C. Ji at Caltech, and Y. Zeng at UN at Reno, was very helpful. This work was supported by National Central University and National Science Council, R.O.C., under Grant NSC-89-2921-M-008-012-EAF. References Bouchon, M. (1981). A simple method to calculate Green s functions for elastic layered media, Bull. Seism. Soc. Am. 71, 959 971. Brodsky, E. E., and H. Kanamori (2000). The elastrohydrodynamic lubrication of faults, J. Geophys. Res. (submitted). Chung, J. K., and T. C. Shin (1999). Implications of the rupture process from the displacement distribution of strong ground motions recorded during the 21 September 1999 Chi-Chi, Taiwan earthquake, TAO 10, 777 786. Chen, K. C., B. S. Huang, J. H. Wang, W. G. Huang, T. M. Chang, R. D. Hwang, H. C. Chiu, and C. C. Tsai (2000). An observation of rupture pulses of the September 20, 1999, Chi-Chi, Taiwan earthquake from the near-field seismograms, Bull. Seism. Soc. Am. (submitted). Cohee, B. P., and G. C. Beroza (1994). Slip distribution of the 1992 Landers earthquake and its implications for earthquake source mechanics, Bull. Seism. Soc. Am. 84, 692 712. Futterman, W. I. (1962). Dispersive body waves, J. Geophys. Res. 67, 5279 5291. Hartzell, S. H., and T. H. Heaton (1983). Inversion of strong ground motion and teleseismic waveform data for the fault rupture history of the 1979 Imperial Valley, California earthquake, Bull. Seism. Soc. Am. 73, 1553 1583. Hsu, Y. J., K. Johnson, P. Segall, S. B. Yu, and L. C. Kuo (2000). Fault geometry and slip distribution of the 1999 Chi-Chi, Taiwan earthquake imaged from inversion of GPS data in The 8th Taiwan Geophysical Annual Meeting, 74 77. Huang, B. S., K. C. Chen, W. G. Huang, J. H. Wang, T. M. Chang, R. D. Hwang, H. C. Chiu, and C. C. Tsai (2000). Characteristics of strong ground motion across a thrust fault tip from the September 21, 1999, Chi-Chi, Taiwan earthquake, Geophys. Res. Lett. 27, 2729 2732. Ji, C., D. V. Helmberger, T.-R. A. Song, K.-F. Ma, and D. J. Wald (2001). The 1999 Chi-Chi, Taiwan, earthquake: a massive wedge-shaped heave, Geophys. Res. Lett. (submitted). Kanamori, H., and T. H. Heaton (2000). Microscopic and macroscopic physics of earthquakes, in Geocomplexity and the Physics of Earthquakes, Geophy. Monograph 120, 147 163. Kao, H., and W. P. Chen (2000). The Chi-Chi earthquake sequence of September 20, 1999 in Taiwan: seismotectonics of an active, out-ofsequence thrust, Science 288, 2346 2349. Kuo, K. W. (1994). Site conditions of the TSMIP (III): Chianan region, Central Weather Bureau Report CW83-1A-12 (in Chinese). Langston, C. A., and D. V. Helmberger (1975). A procedure for modeling shallow dislocation source, Geophys. J. 42, 117 130. Lee, S. J., and K. F. Ma (2000) Slip distribution of the 921 Chi-Chi earthquake inferred from coseismic GPS data, in The 8th Taiwan Geophysical Annual Meeting, 29 36 (in Chinese). Lee, C. T., K. H. Kang, C. T. Cheng, and C. W. Liao (2000). Surface rupture and ground deformation associated with the Chi-Chi, Taiwan earthquake, Sino-Geotechnics 81, 5 18. Liu, K. S., T. C. Shin, and Y. B. Tsai (1999). A free-field strong motion network in Taiwan, TSMIP, TAO 10, 377 396. Ma, K. F., E. E. Brodsky, J. Mori, and T.-R. A. Song (2001). Dynamic rupture behavior of the 1999 Chi-Chi (M w 7.6) Taiwan, earthquake, Science (submitted). Ma, K. F., C. T. Lee, Y. B. Tsai, T. C. Shin, and J. Mori (1999). The Chi- Chi Taiwan earthquake: large surface displacements on inland thrust fault, EOS 80, 605 611. Ma, K. F., T. R. Song, S. J. Lee, and S. I. Wu (2000). Spatial slip distribution of the September 20, 1999, Chi-Chi Taiwan earthquake: inverted from teleseismic data, Geophy. Res. Lett. 27, 3417 3420. Ma, K. F., J. H. Wang, and D. Zhao (1996) Three-dimensional seismic velocity structure of the crust and uppermost mantle beneath Taiwan, J. Phys. Earth 44, 85 105. Okada, Y. (1985). Surface deformation due to shear and tensile faults in a half-space, Bull. Seism. Soc. Am. 75, 1135 1154. Oglesby, D. D., and S. M. Day (2001). Fault geometry and dynamics of