Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes

Transcription

1 Bulletin of the Seismological Society of America, Vol. 16, No., pp , October 216, doi: 1.178/12134 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes by Asako Iwaki, Takahiro Maeda, Nobuyuki Morikawa, Hiroe Miyake, * and Hiroyuki Fujiwara Abstract The robustness of broadband ground-motion simulation can promote seismic-hazard assessment. A broadband ground-motion simulation technique called the recipe is used in the scenario earthquake shaking maps of the National Seismic Hazard Maps for Japan. The recipe represents a fault rupture based on a multiple asperity model referred to as the characterized source model. Broadband groundmotion time histories on the engineering bedrock are computed by a hybrid approach of the 3D finite-difference method and the stochastic Green s function method for the long- (>1 s) and short-period (<1 s) ranges, respectively, using a 3D velocity structure model. The ground motion on the ground surface is computed using the 1D site response of the surface soil layers. Because the need for ground-motion simulations of scenario earthquakes is increasing, it is important to validate the method from seismological and engineering perspectives. This study presents a validation of the recipe using velocity waveforms, peak ground velocity (PGV), seismic intensity, and pseudoacceleration response spectra. The validation scheme follows the framework of the Southern California Earthquake Center Broadband Platform. We selected two M w 6.6 crustal earthquakes that occurred in Japan as the targets of this study: the 2 Tottori and the 24 Chuetsu (mid-niigata) earthquakes. The validation results are satisfactory except for those in the shortest-period range (.1.1 s) at large hypocentral distances (>7 km); such conditions are outside of the target range of the recipe. Simulations using a 1D velocity structure model were also examined. The simulation results for the 1D and 3D velocity structure models indicated that the 3D velocity structure models are important in reproducing PGV and the later phases with long duration, especially on deep sediment sites. Introduction Prediction of broadband ground motion for scenario earthquakes requires numerous processes including modeling the fault rupture, wave propagation, and site response within the near-surface soil structure. Such ground-motion prediction will be important for precise seismic-hazard assessments in modern society. Reproducibility of broadband ground motion by third-party researchers has become an important requirement. The Southern California Earthquake Center (SCEC) worked to solve this issue and established the Broadband Platform (BBP) (e.g., Maechling et al., 21). This is an ideal open-source tool that allows the integration of several rupture-generator models and ground-motion simulation techniques (e.g., Motazedian and Atkinson, 2; Graves and Pitarka, 21; Olsen and Takedatsu, 21). In the current SCEC BBP version, most fault rupture models *Also at Earthquake Research Institute, The University of Tokyo, Yayoi, Bunkyo-ku, Tokyo , Japan. are based on random slip realization or the k-squared slip distribution in the wavenumber domain. Their main objective is to perform a huge amount of ground-motion simulations with random source parameters. The Green s functions are based on 1D velocity structure models specified by the user. The ground-motion simulation methods are validated by comparing the simulated ground motion with the observed ground motion for past earthquakes or ground-motion prediction equations (GMPEs) for scenario earthquakes, using criteria for pseudoacceleration response spectra (PSA) (Goulet et al., 21). Lately, many works focus on validation of ground-motion prediction methods; a number of papers refer to the criteria used in the SCEC BBP validation (e.g., Boore et al., 214; Sun et al., 21). The Earthquake Research Committee (ERC) of the Headquarters for Earthquake Research Promotion (HERP) of Japan has developed a broadband ground-motion simulation method that integrates the characterized source model and the 3D velocity structure model into a standard 2214

2 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 221 procedure called the recipe (ERC, 29), which aims to predict broadband ground motion due to possible earthquake rupture scenarios for active faults. The ERC has been releasing and updating the National Seismic Hazard Maps for Japan, which consist of two types of maps: probabilistic seismic-hazard maps and scenario earthquake shaking maps (ERC, 2; Fujiwara,Kawai,Aoi,Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al., 29). The scenario earthquake shaking maps consider more than 9 active faults evaluated by the ERC (the long-term evaluation of active faults) and display the predicted broadband ground-motion intensity for scenario earthquakes, which are systematically computed by the recipe. The maps can be accessed at an open web platform called the Japan Seismic Hazard Information Station (J-SHIS), which was developed by the National Research Institute for Earth Science and Disaster Prevention (NIED). The recipe is composed of the source modeling part and the broadband ground-motion simulation part. Lessons from the 199 Kobe earthquake revealed that establishing a source model that reproduces rupture directivity pulses is a key issue in seismology and earthquake engineering. Based on the source characterization by Somerville et al. (1999), the asperity inside the fault area was found to follow the scaling as a function of seismic moment. Also, the asperity was found to efficiently generate short-period as well as long-period ground motions (e.g., Sekiguchi et al., 1996; Kamae and Irikura, 1998). This is why the recipe adopts a characterized source model with clear distinctions between the asperities and the background slip area. The recipe employs a hybrid simulation technique in which the long- and short-period components of the ground motion are independently computed, as is the case with many other broadband groundmotion simulation methods (e.g., Graves and Pitarka, 21; Mai et al., 21). The recipe describes the fault rupture by the characterized source model proposed by Irikura and Miyake (21, 211). In the scenario earthquake shaking maps, the recipe is used to simulate broadband ground motions using a 3D velocity structure model. The performance of the recipe is validated by comparing observed and simulated velocity waveforms, % damped response spectra, and seismic intensities. To obtain a better fit, the recipe has been revised by validating its performance for past crustal earthquakes (e.g., Morikawa et al., 211). In this study, we evaluate the applicability of the broadband ground-motion simulation using % damped PSA, following the framework of the SCEC BBP validation exercise. We choose two M w 6.6 crustal earthquakes that occurred in Japan, the 2 Tottori and 24 Chuetsu earthquakes, as the target events of the study. These two events are included in the SCEC BBP validation exercise. The Chuetsu earthquake is also known as the mid-niigata prefecture earthquake and is referred to as the Niigata earthquake in SCEC documents (e.g., Goulet et al., 21). The goals of the ground-motion modeling for the two earthquakes are summarized in Table 1. In addition to the 3D velocity structure Table 1 Target Earthquakes and Broadband Ground-Motion Simulations Earthquake 3D Velocity Structure Model 1D Velocity Structure Model* Tottori Chuetsu Validation of the recipe for a strike-slip fault on stiff rock Validation of the recipe for a reverse fault in thick sedimentary layers model, the 1D velocity structure model used in the SCEC BBP validation exercise is analyzed. In addition to PSA, we compare the observed and simulated ground motion in terms of velocity time histories, peak ground velocity (PGV), and seismic intensity by the Japan Meteorological Agency (JMA) I JMA. The importance of investigating the velocity waveforms of predicted ground motion is supported by recorded seismograms that have similar response spectra but significantly different waveforms (e.g., Koketsu and Miyake, 28). The characteristics of the ground motion in the time domain, including the duration and strong pulses near the fault, largely influence the dynamic response of structures that are not adequately predicted by response spectra (e.g., Bertero et al., 1978; Hall et al., 199; Baker, 27). Method Validation of the source models/evaluation of 3D effects on specific soil conditions *In addition to the 3D velocity structure model used in the recipe as a default, the performance of the 1D velocity structure model for the Southern California Earthquake Center (SCEC) Broadband Platform (BBP) validation model is analyzed. The framework of the ground-motion simulation by the recipe used in the HERP (e.g., Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al., 29; Morikawa et al., 211) is shown in Figure 1. In the recipe, the ground motion on the engineering bedrock and that on the ground surface is simulated using different approaches. On the engineering bedrock, where the S-wave velocity is approximately 3 7 m=s and the effect of nonlinear responses is small, the time series of ground motion is computed by a hybrid approach (Irikura and Kamae, 1999) that combines a 3D finite-difference method (FDM) and the stochastic Green s function method (SGFM; Kamae et al., 1998) for long- (>1 s) and short-period (<1 s) ranges, respectively. The characterized source model and the 3D velocity structure model that covers the crust and deep subsurface structure are implemented in the simulation. Once the waveforms on the engineering bedrock are computed, the ground motion on the ground surface is computed by 1D siteresponse analysis by an equivalent-linear method (Schnabel et al., 1972) or a linear method (Haskell, 196) (right side of Fig. 1a). On the other hand, PGV and I JMA are derived from the empirical relationships between V S3, the time-averaged

3 2216 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara Figure 1. (a) Flow diagram for the broadband ground-motion simulations following the recipe (modified from Morikawa et al., 211; a box for pseudoacceleration response spectra [PSA] RotD was added in this study). (b) Schematic illustration showing the velocity structure (from Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al., 29). The 3D velocity structure model covers the crust and deep velocity structure, and the surface soil layers correspond to the velocity structure between the engineering bedrock and the ground surface. S-wave velocity to a depth of 3 m, and the amplification factors of PGVand I JMA (left side of Fig. 1a; see the Appendix for detail). Characterized Source Models The recipe employs the characterized source model (Irikura and Miyake, 21, 211) to model the fault rupture of crustal earthquakes. The characterized source model is composed of multiple asperities and the surrounding background area, where the asperities are defined as regions, or patches, that have a larger slip than the average slip of the entire rupture area. Following Das and Kostrov (1986), asperities with large slips generate long-period as well as shortperiod seismic-wave radiations due to large stress drop. The background area with a stress-free field also generates long-period seismic-wave radiation. The scaling relationships of the rupture areas of the entire fault and asperities with

4 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 2217 (a) Case1 dip 9º 2 km S1 E Case2 dip 9 S1 E asp.2 asp.1 14 km asp.2 asp.1 26 km (b) Case1 dip 4 km N21 E Case2 dip N21 E asp.2 asp.1 16 km asp.3 asp.1 26 km Case3 N21 E Case4 N21 E dip dip asp.1 asp.2 asp.1 asp.2 Figure 2. Configuration of the characterized source models for the (a) Tottori and (b) Chuetsu earthquakes with 2 km grid lines. The horizontal and vertical directions correspond to the directions along strike and dip, respectively. More detailed parameters are listed in Tables 2 and 3. respect to the total seismic moment are used to derive three kinds of parameters: the outer, inner, and extra fault parameters. The outer fault parameters define the size, configuration, and seismic moment of the entire rupture area. The inner fault parameters describe the heterogeneity of the slip within the fault: the size, seismic moment, and the stress drop of the asperities. The extra fault parameters include the rupture starting point and the rupture velocity. The procedures for deriving the fault parameters used in the HERP are described in Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al. (29) and Morikawa et al. (211). The characterized source model well reproduces the rupture directivity pulse (e.g., Somerville et al., 1997), which enhances the response spectra, as in the 199 Kobe earthquake. In the recipe, information from the long-term evaluation of active faults by the HERP is used to set the location of the rupture fault as well as those of the asperities whenever possible. If such information is not available, the recipe recommends considering several patterns of asperity locations and rupture nucleation points. In this study, we assume that the approximate fault configuration, asperity locations, and the rupture starting point are roughly known, allowing the source characterization and ground-motion simulation processes to be validated. Although the recipe provides the basic procedure to derive the parameters following the scaling relationships, it also recommends incorporating other geological, seismological, or engineering information that is considered better suited to a specific situation. Therefore, we present two and four cases for the Tottori and Chuetsu earthquakes, respectively, using different fault parameters. The fault configurations are shown in Figure 2, and the fault parameters are given in Tables 2 and 3, which are described in more detail in later paragraphs. The Kostrov-like slip time function is given by the formulation of Nakamura and Miyatake (2). The rupture front propagates from the hypocenter with a constant rupture velocity V R. The fault is spatially

5 2218 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara Table 2 Source Parameters for the Tottori Earthquake Unit Length L* km 27(26) Width W* km 1(14) Latitude, longitude at top degrees ( ) 3.27, center Strike degrees ( ) N1E Dip degrees ( ) 9 Rake degrees ( ) 18 Depth to top km 2. Total seismic moment M N m 9: M w 6.6 Short-period level A N m=s=s 1: Average slip D m.68 Rupture velocity m=s 2 Hypocenter depth along km 12 8 dip Asperity area S a km Asperity 1 Area S a1 * km Seismic moment M a1 N m 1: : Effective stress σ a1 MPa Average slip D a1 m Rise time t r s Asperity 2 Area S a2 * km Seismic moment M a2 N m 6: : Effective stress σ a2 MPa Average slip D a2 m Rise time t r s Background area Area S b km Seismic moment M b N m 6: : Effective stress σ b MPa Average slip D b m.6.13 Rise time t r s Locations of the asperities are shown in Figure 2a. *Lengths and widths of the fault and the asperities that are adjusted to the 2 km 2 km subfaults are denoted in the parentheses. Acceleration source spectral amplitude at short periods inferred from the empirical relation with seismic moment by Dan et al. (21). Hypocenter depth from the top of the fault along dip by Goulet et al. (21). Estimated from Iwata and Sekiguchi (22). discretized into : : and 2: 2: km 2 subfaults for FDM and SGFM, respectively. 3D Velocity Structure Models The recipe employs three types of velocity structure models: the crustal structure of the seismic bedrock and deeper, the deep subsurface structure, and the shallow subsurface structure as shown in Figure 1b. The crustal structure model and the deep subsurface structure model are used to compute the waveforms on the engineering bedrock, whereas the shallow subsurface structure model is used to compute those on the ground surface. The velocity structure model is described in detail by Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Hayakawa, et al. (29), and also briefly by Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al. (29). The 3D crustal structure model that extends from the upper crust to the mantle is attached to the deep subsurface structure and is based on the tomography model of Matsubara et al. (28). The 3D deep subsurface structure is based on v.2 of the J-SHIS velocity structure model (J-SHIS V2) (Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Hayakawa, et al., 29; see Data and Resources for reference). It spans the structure between the seismic bedrock (V S 3 m=s) and the engineering bedrock (V S 1 m=s). The physical properties of the bedrock vary depending on the geological conditions. The first three layers (V S < m=s) are excluded in this study to reduce the computational cost because they are modeled only in limited small areas of Japan. The shallow subsurface structure is constructed based on the 2-m-grid Japan engineering geomorphologic classification map of Wakamatsu and Matsuoka (213). The distribution of V S3 is estimated from the map, considering the geomorphologic classification, elevation, slope gradient, and distance from mountains and hills. Broadband Ground Motion on the Engineering Bedrock by a Hybrid Approach As mentioned previously, the time series of the broadband ground motion on the engineering bedrock is computed by a hybrid approach of 3D FDM and SGFM for long- (>1 s) and short-period (< 1 s) ranges, respectively. The procedure of the hybrid approach is described in detail by Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al. (29) and Morikawa et al. (211). The long-period component is computed by the 3D FDM using discontinuous grids (Aoi and Fujiwara, 1999), which runs on an open-source code called Ground Motion Simulator (Aoi et al., 24). The rupture and wave propagation effects are incorporated into the FDM computation using the characterized source model and the J-SHIS V2 3D velocity structure model. Figure 3a,b shows the depth distribution of the seismic bedrock (V S 3 m=s) in the study area for the Tottori and Chuetsu earthquakes, respectively. The source area of the Tottori earthquake is on a stiff rock site, whereas that of the Chuetsu earthquake is on significantly deep sedimentary layers. The depth of the sedimentary layers varies spatially and extends to a maximum of nearly 1 km. The minimum V S is set to m=s, and the grid spacing is set to 8 m at depths smaller than 2 and 8 km for the Tottori and Chuetsu earthquakes, respectively, to ensure accuracy up to the crossover period of 1 s. The grid spacing is set to 24 m in the deeper region. In the FDM computation, Q-value is given by Q f Q S f=f in which f 1 Hz and Q S is the reference Q-value for S wave. The short-period component is computed by the SGFM code of Dan and Sato (1998). In the SGFM, the stochastic ground motion following the ω 2 source model (Boore, 1983) of a small earthquake, or the SGF, is synthesized on

6 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 2219 Table 3 Source Parameters for the Chuetsu Earthquake Unit Length L* km 26(26) Width W* km 16(16) Latitude, longitude at top center degrees ( ) 37.28, Strike degrees ( ) N21E Dip degrees ( ) Rake degrees ( ) 9 Depth to top km 3.6 Total seismic moment M N m 9: M w 6.6 Short-period level A N m=s=s 1: Average slip D m Rupture velocity m=s Hypocenter depth along dip km Asperity area S a km Asperity 1 Area S a1 * km Seismic moment M a1 N m 3: : : : Effective stress σ a1 MPa Average slip D a1 m Rise time t r s Asperity 2 Area S a2 * km Seismic moment M a2 N m 9: : : Effective stress σ a2 MPa Average slip D a2 m Rise time t r s Asperity 3 Area S a3 * km Seismic moment M a3 N m 4: Effective stress σ a3 MPa 1.1 Average slip D a3 m.96 Rise time t r s.83 Background area Area S b km Seismic moment M b N m : : : : Effective stress σ b MPa Average slip D b m Rise time t r s Locations of the asperities are shown in Figure 2b. *Adjusted lengths and widths are denoted in the parentheses (adjusted to the 2 km 2 km subfaults). Estimated from Hikima and Koketsu (2). Estimated from Asano and Iwata (2, 29). the engineering bedrock. Then, the SGF is summed over the fault of a large earthquake based on the self-similar scaling relations and the ω 2 source spectral model in a similar manner as in the empirical Green s function method (Irikura, 1986). The horizontal and vertical components are computed by considering SH and SV waves, respectively, with vertical incident, using a 1D velocity structure extracted from the J-SHIS V2 3D velocity structure model at each site. The empirical vertical-to-horizontal spectral ratio proposed by Nishimura et al. (21) is used to adjust the amplitude of the vertical component. The long- and short-period components are superpositioned in the time domain to create a time series of a broadband ground motion at the crossover period (1 s) after applying a pair of high- and low-cut filters. The sampling frequency is 2 Hz for both the long- and short-period components. We applied high-cut filter at 3 Hz to the broadband waveforms. Broadband Ground Motion on the Ground Surface As shown in Figure 1a, the recipe recommends two methods to simulate the ground motion on the ground surface. In this study, we take the method on the right side of Figure 1 to compute the velocity time series by the 1D siteresponse analysis (Haskell, 196) using the surface soil layers estimated from the logging data of K-NET and KiKnet. On the other hand, we compute PGV and I JMA based on the empirical relationship between V S3, PGV, and I JMA (left side of Fig. 1a). First, I JMA and PGV are calculated from the waveforms on the engineering bedrock. Then, the amplifica-

7 222 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara (a) 4º 3º (b) 4º 3º 13º 3º 13º 14º SMNH7 YMGH9 SMN14 14º NGN18 NIGH16 SMN12 SMN1 SMNH1 SMNH11 SMN1 SMNH1 OKY1 TTR7 OKYH1 SMNH2 TTRH2 SMNH3 SMNH12 SMN3 TTR9 OKYH9 OKYH12 OKYH7 HYGH2 OKY6 OKYH8 SMNH8 HYG11 HRS21 OKYH14 HRS2 OKYH EHM 1 132º 133º 1 13º NGNH31 NIG24 NIGH13 NGN4 NGNH29 tion factors, which are obtained from the empirical relationships between V S3, PGV, and I JMA (see the Appendix), are used to derive the I JMA and PGV on the ground surface. Broadband Ground-Motion Simulation 2 Tottori Earthquake When modeling the Tottori earthquake, we consider two cases, referred to as cases 1 and 2, with different source models. The source parameters and fault configuration for these cases are shown in Table 2 and Figure 2a, respectively. In case 1, the basic procedure of the recipe is followed to derive the outer and inner fault parameters. Meanwhile, the inner fault parameters for case 2 are modified such that the characteristics of the source model are similar to those of the kinematic slip inversion model by Iwata and Sekiguchi (22). HRSH OKY7 EHM1 NIG1 NIG16 NIGH6 NIGH7 NIGH9 NIGH1 FKS3 FKS28 NIG19 NIG2 FKSH21 FKSH NIGH11 FKS26 NIGH12 NIG21 FKS29 NIGH1 TCGH17 NIG23 NGNH7 GNM GNMH1 GNMH9 YMN6 SITH NIGH2 TCG3 TKS11 Seismic bedrock depth [km] FKSH8 YMT7 FKS1 IBRH16 HYG4 FKSH12 HYGH1 Seismic bedrock depth [km] º 14º 141º HYG9 km Figure 3. Study area for the (a) Tottori and (b) Chuetsu earthquakes. The epicenter, stations, and the surface trace of the fault are indicated by a star, triangles, and a thick line, respectively. The scale shows the depth of the seismic bedrock (V S 3 m=s) of the Japan Seismic Hazard Information Station (J-SHIS) velocity structure model. The color version of this figure is available only in the electronic edition. km OSK1 OSKH1 Specifically, the asperity area S a, the average slip D a, and the effective stress σ a in the asperities are modified. The location of the fault, length L, width W, dip, strike, and the locations of the asperities are set based on Iwata and Sekiguchi (22) for both two cases. The observed and simulated broadband velocity waveforms on the ground surface at selected stations are compared in Figure 4. Thex axis represents the lapse time from the occurrence time by JMA, which corresponds to the starting time of the simulations. The waveforms are rotated into the faultnormal and fault-parallel components only at the near-source stations (TTRH2, SMNH1, TTR9, and SMN1). The amplitudes of the fault-normal components are noticeably underestimated by both cases at these stations. Iwata and Sekiguchi (22) simulated near-source ground motion using their optimum source model, and discussed that the large ground velocity on the fault-normal component appears in the region above and northwest of the hypocenter, which could be explained by the upward and northwestward rupture propagation with large rupture velocity. In this study, the rupture propagates at a constant rupture velocity starting from the hypocenter, which differs from the complicated rupture propagation pattern of the optimum model. Nevertheless, the amplitudes of the fault-parallel component at SMNH1 and TTR9 are well represented by the simulations; this is likely to be caused by the large stress drop at the southeast asperity (labeled as asperity 1 in Fig. 2a). Strong pulses due to the rupture directivity observed at OKYH are also well represented by the simulations. However, the simulated waveforms do not explain the observed pulses at the near-source station TTRH2 where the hypocentral distance is smaller than 1 km. The observed waveforms at SMN1 and SMNH1, which are located near the coast in the north of the source area, show relatively long duration with later arrivals because of the velocity structure. Such long-period dominant later phases are underestimated by the simulations, which may reflect the insufficiency of the 3D velocity structure model in this area. 24 Chuetsu Earthquake For the Chuetsu earthquake, we consider four cases: cases 1, 2, 3, and 4. The source parameters and fault configurations of the cases are shown in Table 3 and Figure 2b, respectively. The fault planes of the four cases are determined by reference to the fault models by previous studies (Asano and Iwata, 2, 29; Hikima and Koketsu, 2). The outer and inner fault parameters of cases 3 and 4 basically follow the recipe, whereas some of the inner fault parameters of cases 1 and 2 follow the fault models by the previous studies. The ratio of the area of asperity 1 (S a1 ) to that of asperity 2(S a2 ) is 2:1 for case 3 and 8:3 for case 4, and their locations follow the basic rule of the recipe. The locations of the asperities, the average slip D a, and the asperity area S a of cases 1 and 2 are obtained by modifying the recipe based on the

8 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 2221 (a) TTRH2 Fault normal Fault parallel Fault normal Fault parallel SMNH TTR SMN (b) SMN NS EW NS EW 16.4 SMNH OKYH9 9.3 OKYH SMNH OKYH OKY6 3.3 SMNH (c) HYG NS EW NS EW 1.7 HYGH EHM 1.9 SMN Figure 4. Comparison of observed (top) and simulated (bottom two) broadband velocity waveforms on the ground surface for the Tottori earthquake at selected stations. The maximum amplitude of the two horizontal components is indicated on the right. Hypocentral distances are (a) at near-fault stations (<1 km) (waveforms are rotated into the fault-normal and fault-parallel components.); (b) at regional-distance stations (3 7 km); and (c) at far-source stations (> 7 km). The color version of this figure is available only in the electronic edition. kinematic slip inversion models by Hikima and Koketsu (2) and Asano and Iwata (2, 29), respectively. The observed and simulated velocity waveforms at selected stations are compared in Figure. Again, the waveforms are rotated into the fault-normal and fault-parallel components only at the near-source stations (NIG2, NIGH11, NIG19, and NIGH12). The discrepancy between the simulated and observed waveforms is relatively large at near-source stations (NIG2, NIG19, and NIGH12), as it is for the Tottori earthquake, presumably because the

9 2222 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara (a) Fault normal Fault parallel Fault normal Fault parallel NIG NIGH NIG NIGH (b) NS EW NS EW NIG23 FKS NIGH13.4 NIGH FKS3.9 NGNH (c) NS EW NS EW GNM 2.6 NGN FKSH8 1. YMN Figure. Comparison of observed (top) and simulated (bottom two) broadband velocity waveforms on the ground surface for the Chuetsu earthquake at selected stations. The maximum amplitude of the two horizontal components is indicated on the right. Hypocentral distances are (a) at near-fault stations (<1 km) (waveforms are rotated into the fault-normal and fault-parallel components.); (b) at regional-distance stations (3 7 km); and (c) at far-source stations (> 7 km). The color version of this figure is available only in the electronic edition.

10 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes º PGV º 133º 1 13º º 3º º 133º 1 13º 132º 133º 1 13º PGV Sim 1 1 PGV Sim PGV PGV Figure 6. erved and simulated (cases 1 and 2) peak ground velocities (PGVs) on the ground surface plotted on the map for the Tottori earthquakes. The stars represent the epicenters. The bottom panels compare the observed and simulated PGVs. simulation results at near-source stations are directly affected by the source parameters. As the source region of the Chuetsu earthquake is located on deep sedimentary layers, the waveforms are particularly complicated because of the combination of the fault rupture and the wave propagation effects, but they are generally well reproduced by the simulations at regional and far-source stations. For example, the observed waveforms at NIGH13 show not only large pulses from the source with short duration but also later arrivals from the surface wave, which is reproduced by the simulations. The simulated waveforms overestimate the surface waves at NGNH7, suggesting the need for improvement of the velocity structure model in this area. Broadband Ground-Motion Validation Peak Ground Velocity and Seismic Intensity PGV and I JMA, as well as the velocity waveforms, are the ground-motion indexes that have been conventionally used to validate the recipe (e.g., Fujiwara, Kawai, Aoi, Morikawa, Senna, Kudo, Ooi, Hao, Wakamatsu, et al., 29). The observed and simulated PGVs for the Tottori and Chuetsu earthquakes are compared in Figures 6 and 7, respectively. The overall spatial distribution of PGV is explained by the simulations, except for case 2 of the Tottori earthquake in which the simulated PGV at some regional- and far-source stations overestimates the observation. It should be noted that PGVs in Figures 6 and 7 may differ from the maximum amplitudes of the velocity waveforms in Figures 4 and, reflecting the difference between the two methods (left and right sides of Fig. 1a) for modeling the near-surface site effects. The observed and simulated I JMA values are compared in Figure 8. The observed I JMA is computed from the time series of the observed records. The simulated I JMA generally shows a good agreement with the observed values for both earthquakes. However, the largest observed I JMA of 7 at NIG19 for the Chuetsu earthquake was not reproduced by the simulations. Because I JMA is calculated from acceleration waveforms, short-period components of ground motion may have played a large role in such underestimation of I JMA.

11 2224 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara º 14º 141º 1 139º 14º 141º PGV º 14º 141º 1 139º 14º 141º 1 139º 14º 141º PGV Sim 1 1 PGV Sim PGV PGV PGV Sim 1 1 PGV Sim PGV PGV Figure 7. erved and simulated (cases 1 4) PGVs on the ground surface plotted on the map for the Chuetsu earthquake. The stars represent the epicenters. The bottom panels compare the observed and simulated PGVs. Pseudoacceleration Response Spectra In this section, we evaluate the performance of our broadband ground-motion simulations using PSA, following the framework of the SCEC BBP validation exercise (Dreger et al., 21; Goulet et al., 21). In this framework, % damped PSA of the average median horizontal component (RotD; Boore, 21) inthe.1 1 s period range is compared with the observed data. The validation result in the.1.1 s period range is not justified in this study, as it is out of the target period range of the recipe (.1 1 s). In addition, it should

12 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 222 (a) I JMA Sim (b) I JMA Sim I JMA Sim I JMA I JMA I JMA be noted that a high-cut filter at 3 Hz ( :3 s) is applied to the observed K-NET and KiK-net accelerograms to avoid high-frequency noise (Aoi et al., 211). PSA RotD from the simulated waveforms on the engineering bedrock is compared with the observed data at 4 K-NET and KiK-net stations in Figure 9. The observed PSA data are corrected to the site condition on the engineering bedrock of the J-SHIS V2 3D velocity structure model, using the empirical correction terms for site effects used in the GMPE by Kanno et al. (26). The correction terms are represented by V S3 and V b S3 at each station. Vb S3 represents the site condition on the engineering bedrock of the J-SHIS V2 model, defined as the time-averaged S-wave velocity from the top of the engineering bedrock to a depth of 3 m. V b S3 varies from approximately to 1 m=s, depending on the station. PSA is evaluated at 4 stations and 63 discrete periods from.1 to 1 s, following Goulet et al. (21). I JMA Sim I JMA Sim I JMA Sim I JMA I JMA I JMA Figure 8. Comparison of the observed and simulated Japan Meteorological Agency (JMA) seismic intensity I JMA on the ground surface for the (a) Tottori and (b) Chuetsu earthquakes. The agreement between the observed and simulated PSA at aperiodt i is evaluated using the logarithmic residual given by EQ-TARGET;temp:intralink-;df1;313;79r j;k T i ln O j T i =S j;k T i ; 1 in which O j T i and S j;k T i are the observed and simulated values, respectively, with j and k denoting the stations and simulation cases. Figure 1 shows a plot of the residuals at the 63 evaluated periods, taking the mean and standard deviation over all the stations and cases. The simulations for both earthquakes are underestimating the observed PSA in the short-period range (.1 1 s), whereas they are slightly overestimating it in the long-period range (1 1 s). The residuals at each station, taking the mean over the cases, for periods of.1,.2,., 1., 2., and. s are plotted on the maps in Figure 11. The map indicates that the residuals do not show systematic trends with distance and direction. The residuals are aggregated into four period bins (.1.1,.1 1, 1 3, and 3 1 s) and four distance bins (<, 2, 2 7, and >7 km), using the combined goodness-of-fit (CGOF) parameter: EQ-TARGET;temp:intralink-;df2;313;481 CGOF l; m 1 2 j X NC N C n S;m n T;l X NC 2 N C n S;m X n S;m X n ;l k 1 j 1 i 1 X n S;m X n T;l k 1 j 1 i 1 r j;k T i j jr j;k T i j 2 (Dreger et al., 21), in which n T;l and n S;m denote the numbers of the periods and stations in the lth period bin and mth distance bin, respectively. N C is the total number of the simulation cases, and the operation jjindicates the absolute value. Table 4 shows the CGOF values in the four period and distance bins. Each cell of Table 4 has two numbers separated by a slash, but we only discuss the number on the left in this section; the number on the right is the CGOF for the simulations using 1D velocity structure model, which will be discussed later in the Effects of 1D and 3D Velocity Structure Models section. Most of the cells have CGOF values of less than.7, and thus they are considered good (<:3) orfair (< :7) according to the threshold of the SCEC BBP evaluation. Some cells in the short-period ranges (.1.1,.1 1 s) do not meet the criteria, as the simulations underestimate the observation. The systematic underestimation at short periods is discussed in the Discussions and Conclusions section. CGOF values are also computed for the GMPE by Kanno et al. (26), by replacing S j;k T i with the PSA predicted by GMPE in equation (1). % damped acceleration spectra predicted by the GMPE with correction terms for site effects using V b S3 are compared with the observed PSA. The CGOF values for the simulations (numbers in Table 4) are divided by the corresponding values for the GMPE. In this approach, the performance of the simulation method is compared to that of the GMPE (Dreger et al., 21). The results shown in Table indicate that the performance of the simulations is superior to that of the GMPE in the long-period range

13 2226 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara (a) TTRH2 SMNH1 TTR9 SMN1 (b) NIG2 NIGH11 NIG19 NIGH SMN1 SMNH1 OKYH9 OKYH NIG23 FKS28 NIGH1 NIGH SMNH11 OKYH OKY6 SMNH NIG1 NIGH7 FKS3 NGNH HYG11 HYGH1 EHM SMN GNM NGN4 FKSH8 YMN Period Period Period Period Period Period Period Period Figure 9. % damped PSA RotD for the (a) Tottori and (b) Chuetsu earthquakes. The thick lines correspond to the observed data corrected to the site condition of the engineering bedrock. Simulated PSA are plotted by different lines given in the legend. The color version of this figure is available only in the electronic edition. (a) ln(obs/sim) (b) ln(obs/sim) Tottori 3D Period Chuetsu 3D Period Figure 1. Plot of PSA residuals (see equation 1) with a period.1 1 s for the (a) Tottori and (b) Chuetsu earthquakes. The mean and standard deviation for all stations and cases are shown by solid and dashed lines, respectively. (3 1 s), whereas many cells in the shorter-period ranges are indicated as being comparable to or inferior to the GMPE. We recognize that the values in Tables 4 and should not be directly compared with those of the SCEC BBP validation results shown in Dreger et al. (21) for two major reasons: the first is that we assumed that the approximate locations of the asperities are known prior to modeling the source, and the second is that we used a well-calibrated 3D velocity structure model. The latter will be discussed in a later section, by performing the broadband ground-motion simulation using the same 1D velocity structure as in the SCEC BBP validation. However, the current results demonstrate the performance of the recipe including the characterized source model and the 3D velocity structure model. Effects of 1D and 3D Velocity Structure Models As mentioned in the previous section, the velocity structure model used in this study is different from that used in the SCEC BBP validation. In the SCEC BBP validation, 1D layered velocity structure models for western and central Japan by Goulet et al. (21) are used for the Tottori and Chuetsu earthquakes, respectively (hereafter referred as 1D velocity structure model). We perform broadband ground-motion simulations using the 1D velocity structure models (1D simulations) with the same source models in the simulations discussed earlier (3D simulations) to validate the performance of PSA RotD. We use the corrected PSA data by Goulet et al. (21), which are adjusted to a site condition with V S3 of 863 m=s, which corresponds to V b S3 of the 1D velocity structure models. The numbers on the right of the slashes in Table 4 are the CGOF values obtained from the 1D simulations, and Figure 12 shows a plot of the residuals. 94% and 8% of the cells in Table 4 are found to meet the criterion (CGOF < :7) for the Tottori and Chuetsu earthquakes, respectively, which

14 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 2227 (a) Period =.1 s Period =.2 s Period =. s 3º 3º 3º º 133º 1 13º 132º 133º 1 13º 132º 133º 1 13º.9.6 Period = 1. s Period = 2. s Period =. s.3. 3º 3º 3º º 133º 1 13º 132º 133º 1 13º 132º 133º 1 13º (b) Period =.1 s Period =.2 s Period =. s 1 139º 14º 141º 1 139º 14º 141º 1 139º 14º 141º 1. Period = 1. s Period = 2. s Period =. s º 14º 141º 1 139º 14º 141º 1 139º 14º 141º 1. Figure 11. PSA residuals on a map at periods.1,.2,., 1., 2., and. s for the (a) Tottori and (b) Chuetsu earthquakes. Positive and negative values are plotted by circles and squares, respectively. The stars represent the epicenters. is considered acceptable because the cells that fail the criterion are in the.1.1 s period range. In the shorter-period ranges (<1 s), the CGOF values for the 1D simulations are slightly better than the 3D simulations (numbers on the left in Table 4). On the other hand, the 3D simulations performed better in the longer-period ranges (> 1 s), especially for the Chuetsu earthquake. Figures 13 and 14 show PSA RotD and broadband velocity waveforms, respectively, obtained from the 1D and 3D simulations, at stations located on deep sedimentary layers. Only case 3 of the Chuetsu earthquake is shown as an example. The difference between PSA of 1D and 3D simulations is especially large at long periods (>1 s). Waveforms from the 3D simulation have later phases with larger ampli-

15 2228 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara (a) ln(obs/sim) Tottori 1D Period (b) ln(obs/sim) Chuetsu 1D Period Figure 12. Plot of the PSA residuals for the 1D simulations for the (a) Tottori and (b) Chuetsu earthquakes. The mean and standard deviation of all stations and cases are shown by solid and dashed lines, respectively. tudes and longer durations and show better agreement with the observation. The differences between the 1D and 3D velocity structure models beneath the stations are shown in Figure 1. The observed later phases with long-period components at these stations are underestimated by the 1D simulation; this can be attributed to the lack of deep sediment with spatially dependent depth that varies from 1 to 8 km. The results suggest the importance of using 3D velocity structure models to obtain realistic groundmotion predictions. Table 4 The Combined Goodness-of-fit (CGOF) for the Broadband Ground-Motion Simulations with the 3D and 1D Velocity Structure Models PSA Period/ Distance (km).1.1 s.1 1s 1 3s 3 1 s Number of Stations* Tottori <.29/.32.63/.31.17/.38.1/ /.2.29/.26.4/.41.4/ /.7.33/.34.46/.6.39/.7 16 >7 1.9/.88.71/.6.9/.8.34/.8 16 Chuetsu < 2.82/.6.4/.41.4/.3.17/ /.72./.41.32/.41.2/.3 18 >7 1.33/1..68/.3.3/.41.3/ In each cell, the numbers on the left and right sides of the slash denote the CGOFs for 3D and 1D velocity structure models, respectively. The bold, regular, and italic fonts indicate good (<:3), fair (<:7), and poor (>:7) thresholds, respectively, following the Southern California Earthquake Center (SCEC) Broadband Platform (BBP) validation (Dreger et al., 21). PSA, pseudoacceleration response spectra. *The number of stations that are included in the corresponding distance bins. There is no station within km from the hypocenter for Chuetsu earthquake. NIGH GNM 1 1 Period Discussion and Conclusions We performed a validation of the broadband ground-motion method called the recipe for the Tottori and Chuetsu earthquakes (M w 6.6) using velocity waveforms, PGV, I JMA, and % damped PSA of the RotD component. The simulated velocity waveforms well reproduced the characteristics of the observed waveforms, including the near-fault pulses and later arrival phases with relatively long-period components. However, the simulated waveforms did not match well with the observation particularly at near-source stations. At near-source stations, the simulation results are mostly affected by the source models. It should be noted that our goal is not to perfectly reproduce the observed waveforms Table Ratio of CGOF for the Broadband Ground-Motion Simulations and the Ground-Motion Prediction Equation (GMPE) PSA Period/ Distance (km) NIG NGN4 1 1 Period NIG s.1 1 s 1 3 s 3 1 s Number of Stations Tottori < > Chuetsu < > D 1D The CGOF numbers for 3D simulations in Table 4 are divided by the corresponding numbers for the GMPE. Bold, regular, and italic fonts indicate better than GMPE (<1:), comparable to GMPE (<1:), and inferior to GMPE (>1:) thresholds, respectively, following Dreger et al. (21). 1 1 YMN6 1 1 Period Figure 13. Comparison of % damped PSA RotD from the case 3 simulation of the Chuetsu earthquake using the 3D (thin lines) and 1D (dashed lines) velocity structure models. The thick traces correspond to the observed data corrected to a site condition with V S3 of 863 m=s.

16 Validation of the Recipe for Broadband Ground-Motion Simulations of Japanese Crustal Earthquakes 2229 NS EW NS EW NIGH11.9 NIG D D D D 12.8 NIG GNM 2.6 3D D 1.6 1D 4.6 1D.8 NGN4 1.9 YMN6. 3D 1.2 3D.3 1D.9 1D Figure 14. Comparison of the broadband velocity waveforms on the ground surface for the simulations using the 3D (case 3) and 1D velocity structure models for the Chuetsu earthquake. The top lines represent the observed data. V S [km/s] V S [km/s] V S [km/s] D 1D Depth [km] Depth [km] Depth [km] 1 NIGH11 1 NIG21 1 NIG23 V S [km/s] V S [km/s] V S [km/s] Depth [km] Depth [km] Depth [km] 1 GNM 1 NGN4 1 YMN6 Figure 1. V S structures extracted from the 3D (solid lines) and 1D (dashed lines) velocity structure models at selected sites for the Chuetsu earthquake.

17 223 A. Iwaki, T. Maeda, N. Morikawa, H. Miyake, and H. Fujiwara because the goal of this study is to validate the recipe as a tool for predicting ground motion under conditions that the source parameters are poorly constrained. The simulations also underestimated the later phases at some stations, suggesting the insufficiency of the 3D velocity structure model. When the velocity structure model is further improved in the future, the simulation results are also expected to be improved. The simulated PGV and I JMA generally agreed well with the observation, except for I JMA for the Chuetsu earthquake at the nearest-fault station NIG19, where the simulations underestimated the observed I JMA of 7. We followed the procedure of the SCEC BBP validation presented by Dreger et al. (21) for the validation using PSA RotD and demonstrated that the performance of our simulations is generally acceptable. We found that the performance could be deemed poor in the short-period ranges (.1.1 and.1 1 s)at large hypocentral distance (> 7 km). It should be noted that the recipe has not been validated for the shortest-period range of.1.1 s so far; therefore, the results in this period range are not justified in this study. There are at least two factors that may have caused the systematic underestimation in the short-period ranges. The first is the choice of f max, the frequency at which an acceleration spectrum begins to fall off at high frequencies (Hanks, 1982). Currently, f max is set at 6 Hz for all crustal earthquakes in Japan in the scenario earthquake shaking maps. Because f max controls the amplitude of high-frequency ground motion, further investigation may be needed to determine appropriate regional values of f max. The second factor is the deep subsurface velocity structure model. As shown in Figure 9a, the simulations largely underestimated the observed PSA for periods in the.1 1 srangeatsomestationssuchasoky6, OKYH14, and HYG11, most of which are in a mountainous area where V b S3 is approximately 12 1 m=s. It is possible that the short-period (<1 s) components of the ground motion are not appropriately generated on the engineering bedrock. Improvement on the deep subsurface structure model is in progress (e.g., Senna et al., 213), considering the weathered layer on the bedrock and the interaction between the shallow and deep subsurface structure. We also conducted ground-motion simulations using the 1D velocity structure models used in SCEC BBP validation and compared the results with those obtained from the simulations using the 3D velocity structure model. We found that CGOF values for the 3D simulations were better than those for the 1D simulations at periods longer than 1 s. In addition, the time series of velocity waveforms from the two types of simulations significantly differed from each other. The velocity waveforms from the 3D simulation show better agreement with the observation, especially in reproducing the large later phases and long durations observed at deep sediment sites, which demonstrated the importance of using 3D velocity structure models in ground-motion prediction. Moreover, it is suggested that the characteristics of the ground motion should be evaluated not only by response spectra but also by some features of the time series waveforms, as shown by Paolucci et al. (21). Therefore, future work should concentrate on quantitative evaluation of waveforms, which is necessary for more comprehensive validation of broadband ground-motion prediction methods. Data and Resources The ground-motion data and the logging data were obtained from the National Research Institute for Earth Science and Disaster Prevention(NIED) strong-motion seismograph networks K-NET and KiK-net ( last accessed May 216; Aoi et al., 211). The National Seismic Hazard Maps for Japan, the long-term evaluation of the active faults, and the recipe were documented by the Earthquake Research Committee (ERC) of the Headquarters of Earthquake Research Promotion (HERP) of Japan ( last accessed May 216). The National Seismic Hazard Maps for Japan, the long-term evaluation of the active faults, and the Japan Seismic Hazard Information Station v.2 (J-SHIS V2) 3D velocity structure model can also be accessed online ( last accessed May 216). Most figures were drawn using the Generic Mapping Tools v.4..8 ( hawaii.edu/gmt, last accessed May 216; Wessel and Smith, 1998). Acknowledgments We thank the Earthquake Research Committee (ERC) of the Headquarters of Earthquake Research Promotion (HERP) of Japan, as well as Kazuki Koketsu, James Mori, Kojiro Irikura, Hiroshi Kawase, Paul Somerville, Arben Pitarka, Christine Goulet, Phil Maechling, Fabio Silva, and Norm Abrahamson for their helpful suggestions and discussions. Insightful comments from two anonymous reviewers and Associate Editor Luis A. Dalguer are greatly appreciated. References Aoi, S., and H. Fujiwara (1999). 3D finite-difference method using discontinuous grids, Bull. Seismol. Soc. Am. 89, Aoi, S., T. Hayakawa, and H. Fujiwara (24). Ground Motion Simulator: GMS, BUTSURI-TANSA 7, (in Japanese with English abstract). Aoi, S., T. Kunugi, H. Nakamura, and H. Fujiwara (211). Deployment of new strong motion seismographs of K-NET and KiK-net, in Earthquake Data in Engineering Seismology: Predictive Models, Data Management and Networks, Geotechnical, Geological, and Earthquake Engineering, S. Akkar, P. Gülkan, and T. van Eck (Editors), Vol. 14, Springer, Dordrecht, The Netherlands, Asano, K., and T. Iwata (2). Source rupture models of recent disastrous earthquakes in Japan, APRU/AEARU Research Symposium 2, Kyoto, Japan, P8. Asano, K., and T. Iwata (29). Source rupture process of the 24 Chuetsu, mid-niigata prefecture, Japan, earthquake inferred from waveform inversion with dense strong-motion data, Bull. Seismol. Soc. Am. 99, Baker, J. W. (27). Quantitative classification of near-fault ground motions using wavelet analysis, Bull. Seismol. Soc. Am. 97, Bertero, V. V., S. A. Mahin, and R. A. Herrera (1978). Aseismic design implications of near-fault San Fernando earthquake records, Earthq. Eng. Struct. Dynam. 6,