Selection and ranking of ground-motion models for the 2018 National Seismic Hazard Assessment of Australia

Transcription

1 Record 2018/29 ecat Selection and ranking of ground-motion models for the 2018 National Seismic Hazard Assessment of Australia Summary of ground-motion data, methodology and outcomes H. Ghasemi and T. Allen APPLYING GEOSCIENCE TO AUSTRALIA S MOST IMPORTANT CHALLENGES

2

3 Selection and ranking of ground-motion models for the 2018 National Seismic Hazard Assessment of Australia Summary of ground-motion data, methodology and outcomes GEOSCIENCE AUSTRALIA RECORD 2018/29 Hadi Ghasemi and Trevor Allen

4 Department of Industry, Innovation and Science Minister for Resources and Northern Australia: Senator the Hon Matthew Canavan Secretary: Dr Heather Smith PSM Geoscience Australia Chief Executive Officer: Dr James Johnson This paper is published with the permission of the CEO, Geoscience Australia Commonwealth of Australia (Geoscience Australia) 2018 With the exception of the Commonwealth Coat of Arms and where otherwise noted, this product is provided under a Creative Commons Attribution 4.0 International Licence. ( Geoscience Australia has tried to make the information in this product as accurate as possible. However, it does not guarantee that the information is totally accurate or complete. Therefore, you should not solely rely on this information when making a commercial decision. Geoscience Australia is committed to providing web accessible content wherever possible. If you are having difficulties with accessing this document please clientservices@ga.gov.au. ISSN X (PDF) ISBN (PDF) ecat Bibliographic reference: Ghasemi, H. and Allen, T Selection and ranking of groundmotion models for the 2018 National Seismic Hazard Assessment of Australia: summary of ground-motion data, methodology and outcomes. Record 2018/29. Geoscience Australia, Canberra.

5 Contents 1 Introduction Candidate Ground-Motion Models Stable, non-cratonic continental region (SCR-non-cratonic) Stable, cratonic continental region (SCR-cratonic) Active crustal region (ACR) Method Ground-Motion Data Ground-motion model ranking Discussion and Concluding Remarks References Appendix A List of selected events for ranking of ground-motion models Appendix B Comparison of the observed spectral accelerations with the selected groundmotion models Appendix C Residual analysis of the selected ground-motion models Insert document title here iii

6

7 1 Introduction One of the key challenges in assessing earthquake hazard in Australia is understanding the attenuation of ground-motion through the stable continental crust. There are now a handful of groundmotion models (GMMs) that have been developed specifically to estimate ground-motions from Australian earthquakes. These GMMs, in addition to models developed outside Australia, are considered in the 2018 National Seismic Hazard Assessment (NSHA18; Allen et al., 2018). In order to assess the suitability of candidate GMMs for use in the Australian context, ground-motion data from small-to-moderate Australian earthquakes have been gathered. Both qualitative and quantitative ranking techniques (e.g., Scherbaum et al., 2009) have been applied to determine the suitability of candidate GMMs for use in the NSHA18. This report provides a summary of these ranking techniques and provides a discussion on the utility of these methods for use in seismic hazard assessments in Australia; in particular for the NSHA18. The information supplied herein was provided to participants of the Ground-Motion Characterisation Expert Elicitation workshop, held at Geoscience Australia on 9 March 2017 (Griffin et al., 2018). Selection and ranking of GMMs for NSHA18 1

8 2 Candidate Ground-Motion Models In recent years, the number of ground-motion models (GMMs) has increased significantly due to the improvement and expansion of global and regional seismic networks. Douglas (2018) summarizes the characteristics of 440 GMMs for the prediction of peak ground acceleration (PGA) and 282 models for the prediction of elastic response ordinates. Given the number of available models, several studies have attempted to define selection criteria to shortlist GMMs that can capture the expected range of possible ground motions in the target region, such as the tectonic region type, magnitude type, spectral period range, magnitude and distance ranges and calibration of site effects (e.g., Cotton et al., 2006; Bommer et al., 2010; Stewart et al., 2011; Stewart et al., 2014). Such criteria are considered in the pre-selection of candidate GMMs for the NSHA18. We also considered GMMs developed for active tectonic regions, despite the selection criteria recommended by aforementioned studies. Researchers have previously suggested that the attenuation of ground-motions in Australia should be similar to active tectonic regions, such as California, based on the similarities in geologic setting (Gibson and Dimas, 2009) and also from observed acceleration records (Brown et al., 2001). Indeed, the findings of Allen et al. (2011) showed that the Chiou and Youngs (2008) model developed for active shallow crustal earthquakes appears to approximate moderate-magnitude Australian groundmotions well at short source-receiver distances. However, this model tended to overestimate ground motions for periods longer than approximately 0.2 s for small-magnitude earthquakes. Hoult et al. (2014) subsequently demonstrated the utility of modern active tectonic GMMs to estimate response spectral accelerations for recent, well-recorded earthquakes in eastern Australia. Table 1 lists the main characteristics of the pre-selected GMMs for the NSHA18 including the range of magnitude and distance, the spectral period band, as well as the target region. The GMMs were preselected based on experience of the authors and the models known performance in other hazard assessments worldwide, and also to provide a range of models for different tectonic environments (e.g. cratonic, stable shallow crust and active crust). Following is a short descriptions of the selected models grouped based on their respective tectonic regimes. 2.1 Stable, non-cratonic continental region (SCR-non-cratonic) Two of the selected SCR-non-cratonic ground-motion models (out of 11) are models developed specifically for Australian crustal conditions, i.e. A12 (Allen, 2012), and SEA09_non_craton (Somerville et al., 2009; cratonic and non-cratonic models). These models are developed for the stable continental region of south-eastern Australia. The selected Australian-specific models are derived from the regression analysis of synthetic ground-motion data where the records from small-to-moderate magnitude local earthquakes were used to determine the input parameters of numerical simulations. However, it should be noted that these models are based on different simulation algorithms. For example, the A12 model is based on stochastic finite-fault simulation technique while the SEA09_non_craton model is developed using a hybrid of stochastic simulations at high frequencies and physics-based modelling at low frequencies. We note that two other Australian GMMs have been developed: those of Gaull et al. (1990) that were used to underpin their national hazard model and Liang et al. (2008). Both of these GMMs are calibrated to local magnitude M L with Allen et al. (2011) noting these GMMs generally perform poorly for the available ground-motion records (after M L to M W conversions ere applied). Therefore they are not considered further in these analyses. Selection and ranking of GMMs for NSHA18 2

9 Four additional candidate SCR-non-cratonic models are developed for application in central and eastern U.S. (CEUS), a region often considered to be a tectonic analogue for Australia. These models are: AB06 (Atkinson and Boore, 2006), Sea02 (Silva et al., 2002), Pea11 (Pezeshk et al., 2011) and YA15 (Yenier and Atkinson, 2015) (Table 1). As with the Australian-specific models, all of the selected CEUS models are also derived using synthetic data generated by different simulation techniques. Notably, the Sea02 and YA15 models are based on approximation of the earthquake source as a point source with double-corner spectrum. All of the aforementioned models have physically-based and relatively simple functional forms with terms modelling the magnitude scaling, geometrical spreading and anelastic attenuation as functions of moment magnitude and rupture distance (the closest distance to rupture). The YA15 model, in addition to M W requires an event specific stress parameter (Δσ) to model the earthquake source. The stress parameter predominantly affects the short periods of the response spectrum and can be expressed as a function of M W (Rietbrock et al., 2013; Yenier and Atkinson, 2015) or focal depth (Allen et al., 2004; Allen, 2012; Yenier and Atkinson, 2015). Other selected models predict the level of ground motion assuming a region-specific average stress parameter that is used in the simulation process. In Table 1, the models of Eea16_Zea06, and Eea16_CY08 (Edwards et al., 2016) represent the adjusted versions of the Zhao et al. (2006; Japan), and Chiou and Youngs (2008; California) models for application in Switzerland, which is also classified as a stable continental region (Johnston, 1994). Edwards et al. (2016) derived model-specific, period-dependent factors to adjust the response spectra represented by the active tectonic ground-motion models by taking into account the differences between the host (i.e., Japan and/or California) and target (i.e. Switzerland) crustal velocity profiles, as well as near-surface attenuation parameters. In addition, using records from small-to-moderate earthquakes, they derived model-specific correction factors to make the GMMs suitable for predictions at small-to-moderate magnitudes (e.g. M W 5.5) in the Swiss region. 2.2 Stable, cratonic continental region (SCR-cratonic) The selected SCR-cratonic model, i.e. SEA09_craton (Somerville et al., 2009; cratonic and noncratonic models), is an Australian specific one developed for the Yilgarn craton, but is recommended to be applied for the Proterozoic and Achaean regions of central and western Australia. Similar to SEA09_non_craton model, this model is also developed using a hybrid of stochastic simulations at high frequencies and physics-based modelling at low frequencies. 2.3 Active crustal region (ACR) Among the selected ground-motion models, the models of Bea14 (Boore et al., 2014) and CY14 (Chiou and Youngs, 2014) are the only models developed mainly based on the recorded data. These models are developed as part of NGA-West2 project (Bozorgnia et al., 2014) where the majority of data is from California earthquakes, complemented with recordings from large earthquakes from other active tectonic regions (Ancheta et al., 2014). Both of these GMMs have more complex functional forms with some additional terms to model basin effects, hanging-wall effects, non-linear site response, etc. One of the key advantages of these models is that they are applicable over a wide range of earthquake magnitudes, source-to-site distances, and spectral periods (Table 1). In contrast, the complexity of their parametric form often requires user assumptions that may be difficult to characterise a priori. Table 2. defines the parameters that Selection and ranking of GMMs for NSHA18 3

10 may be required by the candidate models. Some of the listed parameters are not readily available for earthquakes and recording stations in Australia and need to be estimated either subjectively or using empirical relationships (e.g. Kaklamanos et al., 2011) that are not based on Australian data. Selection and ranking of GMMs for NSHA18 4

11 Table 1 Description of the selected ground-motion models. For definitions of magnitude, distance and site parameters see Table 2. Model ID Magnitude Type Magnitude Range Distance Type Distance Range Site Condition Horizontal Component Type 1 Period Range Region 2 Active shallow crust BEA14 MW RJB VS30 ( ) GMRotD50 PGA-10 California CY14 MW RRUP VS30 ( ) GMRotD50 PGA-10 California Stable continental crust SEA02 MW RJB VS30 (2800) GM PGA-10 CEUS AB06 MW RRUP VS30 (760) GM PGA-5 CEUS SEA09_craton MW RJB VS30 (865) GM PGA-10 Aus-cratonic SEA09_non_craton MW RJB VS30 (865) GM PGA-10 Aus-non-cratonic PEA11 MW RRUP VS30 (2000) GM PGA-10 CEUS A12 MW RRUP VS30 (820) GM PGA-4 SEA YA15 MW RRUP VS30 (760) GM PGA-10 CEUS EEA16_CY08 MW RRUP VS30 (1100) GMRotI50 PGA-10 Swiss EEA16_ZEA06 MW RRUP VS30 (1100) GMRotI50 PGA-5 Swiss 1 In this column: GMRotD50, and GMRotI50: orientation-independent measures of ground-motion (Boore et al., 2006); GM: geometric mean of two horizontal components. 2 In this column: CEUS: central and eastern U.S.; Aus-cratonic: Australian cratonic crust; Aus-non-craton: Australian non-cratonic crust; SEA: southeastern Australia. Selection and ranking of GMMs for NSHA18 5

12 Table 2: definition of the parameters that may be required by the candidate GMMS Parameter M W Z TOR λ Δσ δ R EPI R HYP R RUP R JB R X V S30 Z 1.0 Z 2.5 Definition Earthquake source parameters Moment magnitude Depth to the top of the rupture (km) Rake angle (degree) Stress parameter (bar) Fault dip (degree) Path parameters Distance from epicentre (km) Distance from hypocentre (km) Closest distance to the rupture plane (km) Horizontal distance to the surface projection of the rupture plane (km) Hor izontal distance to top edge of rupture measured perpendicular to the strike (km) Site parameters Time-averaged shear-wave velocity over the top 30 meters of the subsurface (m/s) The shear wave velocity horizon of 1 km/s (m) The shear wave velocity horizon of 2.5 km/s (km) Selection and ranking of GMMs for NSHA18 6

13 3 Method To study the characteristics of the selected models, we formed high dimensional model representative (HMR) matrices by sampling the magnitude range from M W , the distance range of km, and the period range of second. The selected ranges represent those of engineering interest. The ground-motion model predictions and corresponding standard deviation values were subsequently estimated for each sample. The magnitude and distance ranges are each discretised by 10 samples, equally spaced in linear space and logarithmic space, respectively. The period range is represented by 10 samples at 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 1.0, 2.0, 3.0, 4.0 sec. The calculations are for pure reverse faulting with dip angle of 45 (Figure 1). The stress parameter required by YA15 model is set to 300 bars, i.e. the average value for earthquakes in CEUS with M W values larger than 5.0, and focal depth values larger than 10.0 km (Yenier and Atkinson, 2015). The fault plane dimensions for each of the magnitude bins are determined based on the magnitude-area scaling law by Wells and Coppersmith (1994) for reverse faults. Figure 1: Illustration of the earthquake rupture geometries. The magnitude range of is represented by 10 samples, equally spaced in linear space. The rupture aspect ratio is set to 1.5, and Wells and Coppersmith (1994) scaling relationship is used to estimate the rupture area for the given magnitude The selected ground-motion models use different definitions of the source-to-site distance, i.e. rupture distance and/or Joyner-Boore distance parameters (Table 1). The CY14 model requires an additional Selection and ranking of GMMs for NSHA18 7

14 distance metric known as R X to model hanging-wall effects. For a given pair of source to site locations and magnitude, all of the other required distance metrics are calculated from the specific rupture scenario (Figure 2). The model predictions are calculated for hanging-wall, and hard-rock sites with V S30 of 760 m/s, which is similar to average site conditions obtained from geotechnical studies at several seismograph stations across Australia (Collins et al., 2006; Kayen et al., 2014). It should be noted that the meaning of hard rock varies among the selected models (Table 1), reflecting the unique site characteristics of their host region. Consequently, to compare ground-motion models, ideally the predictions of the selected models should be adjusted to a common site profile (e.g., Edwards et al., 2016). However, in this study we did not take into account the site effects adjustments and the definition of the reference hard-rock site condition is consistent with that of the selected ground-motion model. This is justified as the original, and not the adjusted version of the groundmotion models, will be used to calculate earthquake hazard in Australia. The other site parameters, i.e. Z 1.0 and Z 2.5 required by CY14 to model basin effects are estimated for V S30 of 760 m/s using empirical relationships between Z 1.0 -V S30 of Chiou and Youngs (2014) and Z 2.5 -Z 1.0 of Kaklamanos et al. (2011), respectively. Figure 2: Comparison of several source-to-site distance measures versus RRUP for an earthquake with magnitude of 7.0. The corresponding rupture geometry is illustrated in Figure 1. The dashed line indicates the 1:1 relationship. Point source distance metrics, i.e. REPI, and RHYP are also presented for comparison. Note that, in this case, using these distance measures rather than those reflecting the dimensions of the fault rupture, i.e. RJB, and RRUP will significantly bias the model predictions. The HMR-matrices for each of the models are visualized as three-dimensional voxel grids. Each volume in Figure 3 and 4 represent the mean values (i.e. the logarithm of the median ground-motion) and sigma values of each of the selected models at specific magnitude-distance-period bins, respectively. Figure 3 shows significant variations among selected models in terms of predicted ground motions. The models of SEA02 and SEA09_craton clearly stand out as outliers characterized Selection and ranking of GMMs for NSHA18 8

15 by large spectral accelerations at short distances (R RUP <~30 km). The southeastern Australian models of A12 and SEA09_non_craton show certain similarities in terms of the level of predicted spectral acceleration as well as the ground motion attenuation rate with distance. However, in comparison with SEA09_non_craton model, A12 model predicts slightly larger spectral accelerations at short distance and period ranges. Figure 3: Voxel grids of the logarithms of ground-motion median values for models shown in Table 1. Sigma values of the selected models also show significant variations with the largest values linked with A12 and SEA02 models (Figure 4). All of the models express sigma as a period dependent variable, and certain models (e.g. CY14) represent sigma as a magnitude-distance-period dependent Selection and ranking of GMMs for NSHA18 9

16 (Figure 4). It is also interesting to note that in comparison with A12 model, SEA09_craton and SEA09_non_craton models have significantly lower sigma values except at longer periods (period > 1.0 sec) where there is a sudden notable increase in the values. The author of A12 recognises that the sigma model was likely overestimated for rock ground-motions, and should be revised to lower values. Figure 4: Voxel grids of the ground-motion sigma values for models shown in Table 1. Selection and ranking of GMMs for NSHA18 10

17 4 Ground-Motion Data The ground-motion dataset is a subset of data compiled by Allen (2012) and Allen et al. (2006) for southeastern and southwestern Australian earthquakes, respectively. The original dataset contains 1562 records from 179 earthquakes that occurred between 1989 and 2014 with moment magnitudes in the range of 2.0 to 5.4. In this study we selected 140 records with hypocentral distances less than 300 km recorded from earthquakes with M W 3.8 (Appendix A). To compare the observations with the model of YA15, we assigned a stress parameter of 300 bar to the causative earthquake sources. The far-field records as well as records from smaller events are of low engineering significance and are not included in our model evaluation. We assumed the local site condition of the recording stations as engineering rock with V S30 of 760 m/s. Figure 5 shows the distribution of the selected records with respect to magnitude and distance. Figure 5 The magnitude versus hypocentral distance distribution of the selected records. The ground-motion records from southeast and southwest Australia are shown as circles and triangles, respectively. Selection and ranking of GMMs for NSHA18 11

18 There are two clear gaps in the data. Firstly, there is no record from moderate to large earthquakes with M W > 5.0. Such earthquakes have the largest potential for damage. Secondly, there are few records in the near-field region (e.g., hypocentral distance less than approximately 30 km). The majority of data are recorded at hypocentral distances between 30 and 300 km from earthquakes with magnitudes in the range of 3.8 to 4.8. Hence, the performance of the candidate ground-motion models can only be verified within aforementioned distance and magnitude ranges. Figure 6: (a) Unfiltered accelerogram recorded at hypocentral distance of 85 km from 2014-Feb-26 earthquake with M W=4.34 in southwestern Australia. Displacement time-series from filtered accelerogram using causal and acausal Butterworth filters with corner frequencies at 0.25 Hz and 0.1 Hz, are also presented. The blue shaded area shows the selected signal+noise window, and the red shaded area shows part of the selected noise window (b) Fourier acceleration spectra of the manually selected signal and noise windows. The shaded area represents the passband of the filter inferred from inspection of Fourier amplitude spectrum and the level of signal-to-noise ratio (c) Ratio of 5%-damped acceleration response spectra for the filtered accelerograms using causal filtering with corner frequency (fc) of 0.25 Hz as well as acausal filtering with fc of 0.1 Hz to the acceleration response spectrum of the unfiltered accelerogram. Selection and ranking of GMMs for NSHA18 12

19 The selected database represents a mixture of accelerograms and velocity seismograms with varying sampling rates. The baseline of the records were adjusted by subtracting the mean of the whole record to remove long-period biases. The first and last 5% of the time-series were cosine tapered and the instrument response was removed to recover the actual ground velocity and/or acceleration. The velocity seismograms were then differentiated to obtain ground acceleration. To account for low- and high-frequency noise, records were filtered using causal, fourth order Butterworth filters with passbands of sampling rate in Hz. Each of the velocity and displacement time-series obtained through integration of the filtered acceleration were visually inspected to check whether or not they appear to be reasonable. The acceleration time-histories that produced unphysical velocity and displacement records were not considered for further processing. It should be noted that the noise characteristics of each of the ground-motion records is unique and hence, ideally, each record should be filtered with record specific corner frequencies chosen based on the shape of Fourier amplitude spectrum and the level of signal-to-noise ratio. However this process is rather time consuming and requires significant effort to treat each record individually. In this study we analysed recorded ground-motions at spectral periods of 0.1, and 1.0 sec. The selected periods represent ground motions at short and mid period ranges, and are well within the filtering passband. However, in following the above processing procedure, some valuable long-period information that could be retrieved from records may be lost; hence preventing to test the performance of candidate ground-motion models at longer periods, i.e. spectral periods>3.0 sec. Figure 6 shows an example of such high quality ground-motion record at hypocentral distance of 85 km from an earthquake in southwestern Australia with M W For this particular record careful inspection of Fourier amplitude spectrum and the level of signal-to-noise ratio suggests that the corner frequency of the high-pass filter can be set as 0.1 Hz. Applying causal and acausal Butterworth filters with corner frequencies at 0.25 Hz and 0.1 Hz, respectively yield reasonable velocity and displacement timeseries, i.e. the corrected waveforms show no clear baseline drift; however it is clear that by applying the causal filter, long-period spectral accelerations (spectral periods > ~3.0 sec) are significantly underestimated. It should be also noted that the choice of acausal filtering is more appropriate than causal filtering as in the former case the spectral ordinates within the passband of the filter are not sensitive to the filter corner frequencies. Ongoing work will focus on developing a comprehensive Australian ground-motion database including all available Australian ground-motion records. The database will be implemented by first defining a ground-motion database schema followed by collecting and processing of the waveforms in a systematic and consistent way. The related metadata such as earthquake source parameters, local site conditions at the recording stations, and other relevant engineering parameters, such as peak ground-motion parameters, will be also gathered and stored in the database. Finally the database will become available to the public. Selection and ranking of GMMs for NSHA18 13

20 5 Ground-motion model ranking The performance of the selected GMMs are evaluated against southeastern and southwestern Australian data at hypocentral distance ranges of <150 km and km, and at spectral periods of 0.1 sec and 1.0 sec. Figure 7 (a), as an example, compares the observed spectral accelerations at 0.1 sec with the model of Allen (2012). The plotted data are from southeastern Australian earthquakes recorded at hypocentral distances less than 150 km. To help the visualisation of the performance of the candidate GMMs, the observed spectral accelerations, shown as filled circles, are normalised to a reference earthquake with M W 4.0 occurring at depth of 5.0 km, such that: logsa norm = logsa obs + μ(gmm(eve ref, Site rec )) μ(gmm( Eve obs, Site rec )) (1) where: SA norm : Normalised spectral acceleration at period of interest SA obs : Observed spectral acceleration at period of interest GMM : Ground-motion model of interest Eve ref : Set of input parameters required by GMM for the reference earthquake Eve obs : Set of input parameters required by GMM for the observed earthquake Site rec : Set of input parameters required by GMMfor the recording station μ : GMM s predicted mean value of the natural log of the spectral acceleration at the given period It should be noted that, as described above, the displayed observed data points (filled circles) are normalised spectral acceleration values and hence subject to change if GMM of interest changes. In Figure 7a, the mean and one standard deviation boundaries predicted by ground-motion model for reference event are shown as solid and dashed curves respectively. It can be seen that, overall the A12 model fits the data reasonably well with most of the observations lie within one standard deviation of the predicted mean curve. However, in general the model slightly underestimates the observations over the entire distance range. To further evaluate the fitness of the candidate ground-motion models, histograms of the normalised residuals are plotted for periods of interest (e.g. Figure 7b). For each observation, i.e. computed spectral acceleration at period of interest, the normalised residual is calculated as: R norm = (log(obs) log(pre)) σ (2) Selection and ranking of GMMs for NSHA18 14

21 where: SA obs : Observed spectral acceleration at period of interest SA pre : predicted spectral acceleration value at the given period σ : corresponding standard deviation as indicated by the candidate ground-motion model at period of interest Figure 7: a) Comparison of the observed spectral accelerations at period of 0.1 sec with the ground-motion model of A12. The observations are normalised to the reference earthquake with M W4.0 occurring at depth of 5.0 km. The mean and one standard deviation boundaries predicted by ground-motion model for the reference event are shown as solid and dashed curves respectively. b) Histogram of the normalised residuals. Comparison between fitted normal distribution (black curve) and standard normal distribution (red curve) are also presented. c-d) Distribution of the ground-motion model s residuals with respect to magnitude and distance. The fitted lines to the residuals are shown as dashed lines. Selection and ranking of GMMs for NSHA18 15

22 The normalised data residuals should follow the standard normal distribution, i.e. N(μ = 0, σ = 1), if data perfectly matches the model predictions. In this figure the standard normal distribution (red curve) is compared with the curve fitted to the normalised residuals (black curve). It can be seen that overall the A12 model slightly under-predicts the observations, and the scatter in the observations, measured by standard deviation of the fitted normal distribution, is comparable with that from the ground-motion model. The scatter plots of the residuals with respect to magnitude and distance are presented in Figure 7 (c-d), respectively. The dashed solid lines in these graphs are best fitting least-squares lines, calculated in order to detect any possible trends in the distribution of the residuals with respect to magnitude and distance. Based on this limited dataset, there is a significant negative trend in the distribution of the residuals with a positive intercept suggesting that the performance of the model improves as magnitude or distance increases. Appendix B includes similar plots to Figure 7 generated for each combination of ground-motion model, tectonic region, period of interest, and distance range parameters. To verify and compare the performance of the selected ground-motion models in a quantitative way, Appendix C summarizes, for each combination of the aforementioned parameters, the mean, median, and standard deviation values of the residuals as well as log-likelihood (LLH) measures. Following Scherbaum et al. (2009) the log-likelihood measure is computed as: LLH = 1 N log N i=1 2(g(x i )) (3) where: N : Number of observations x i : i th observation g( ) : Normal probability density function as computed by the selected ground-motion model The LLH measure reflects the relative information loss if the perfect ground-motion model (unknown) is replaced by the ground-motion model of interest. Hence, the lower the LLH measure means the closer distance between ground-motion model of interest and the perfect model that represents the data. Visual inspection of the results (Appendix B) indicates that while there is no model that perfectly matches the Australian dataset, the ground-motion models that appear to perform relatively well across all selected periods, tectonic regions, and distance ranges are A12, BEA14, and CY14 models. Note that the minimum magnitude of the compiled dataset (M W 3.8) is slightly outside the applicability range of the A12 model. Consistent with the results of qualitative comparison of the selected models against observations, relatively low LLH values are computed for the aforementioned models. The good performance of A12 model against data from southeastern Australia is expected as part of these data had been used in developing this model. Consequently, this can also bias the estimation of LLH value towards lower numbers (Scherbaum et al., 2009). Other Australian specific models, i.e. SEA09_craton and SEA09_non_craton do not perform as well against the observations from smallmagnitude earthquakes. It should be noted, however, that the entire magnitude range of the compiled database is not covered by the applicability range of these models (Table 1). Several studies have shown that ground-motion Selection and ranking of GMMs for NSHA18 16

23 models should not be extrapolated to magnitudes outside the range supported by the model due to magnitude-scaling problems (Bommer et al., 2007; Cotton et al., 2008). Note that the LLH measure reflects the overall performance of the model against the entire data, which may undermine some desirable features of the model. For example, SEA09_non_craton model predicts the near-field data (R JB <~30 km) from southeastern Australian earthquakes reasonably well, while over-predicting the data recorded at distance range of km. A large LLH value is computed for this model as the majority of the data is recorded at R JB >30 km. The performance of the other selected models, i.e. the Swiss and CEUS models, seems to vary among different tectonic regions, periods as well as distance ranges. One interesting example is the relatively good performance of the SEA02 model against southwestern Australian data at a period of 1.0 second and distance range of <150 km in comparison with the poor performance of this model in southeastern Australia (see Appendix B: Figure 21 & 65). This can be substantiated using both qualitative, i.e. visual inspection of the model fit to data, as well as quantitative, i.e. the computed LLH values, approaches. This comparison may imply regional variations of ground motions, but as follows the results should be interpreted with great caution given the limitations in the datasets used. Based on the results, as shown in Appendix B: Figure 17 & 61, it can be seen that for the same distance and period criteria, the model of CY14 performs reasonably well in both tectonic regions. Therefore, let us assume that CY14 represents the true model, and sample the model space two times following the distribution of the selected data with respect to magnitude and distance (Figure 5). Now for each data point within each of the two sample spaces, we can calculate the Kullback-Leibler distance (KL-distance) between CY14 and Sea02 models as: D KL (P Q) = p(x) log p(x) dx (4) q(x) where p and q denote the probability density functions of CY14 and SEA02 models, respectively. In this way, the KL-distance measures the differences between the true model, i.e. CY14 model, and the target model, i.e. SEA02 model. Figure 8 shows the box plots of the KL-distances computed for the two sample spaces. It is clear that the performance of the target model is sensitive to the number of samples. The target model represents the true model reasonably well for the first sample space, i.e. the one that represents data points for southwestern Australia, and the opposite holds for the second sample space, i.e. the one representing data points for southeastern Australia. This suggests that the performance of the models is quite sensitive to the number of observations and the results should be interpreted with great caution for datasets with limited observations. Selection and ranking of GMMs for NSHA18 17

24 Figure 8: Box plots of the KL-distances computed for the two sample spaces: Group1 and Group2 representing the model space sampled following the data distribution of the southwestern Australia, and southeastern Australia, respectively. For each sample the KL-distance is calculated between CY14 and SEA02 models. Selection and ranking of GMMs for NSHA18 18

25 6 Discussion and Concluding Remarks In this study we pre-selected and analysed 11 candidate ground-motion models to be considered for the Australian National Seismic Hazard Assessment (Table 1). The models were chosen subjectively, considering the findings of previous studies on selection of ground-motion models for Australia, and other stable continental regions. The selected models along with the key outcomes of this study were presented in the expert elicitation workshop to select and rank GMMs for NSHA18 (Griffin et al., 2018). The proximity of the selected ground-motion models was explored for magnitude, distance and period ranges of M W : , R RUP : km, T: sec, respectively. The results, for each groundmotion model, were presented as voxel grids of the ground-motion mean and sigma values (Figures 3 and 4). The results present the complete model space of the selected ground-motion models, and serve as a practical tool to assist the relative weighting, but not necessarily ranking, of the GMMs in the framework of a logic-tree. To verify, and potentially rank the performance of ground-motion models, we compiled a groundmotion database that included 140 records with hypocentral distances less than 300 km recorded from earthquakes with M W 3.8 (Figure 5). One of the advantages of the dataset is that earthquake magnitudes are expressed in terms of M W, ensuring consistency with the selected ground-motion models, which are calibrated to the moment magnitude scale. The magnitudes are taken from systematic studies on measuring M W for small-to-moderate earthquakes in Australia (Allen et al., 2006; Allen, 2012; Ghasemi et al., 2016). Furthermore, considerable effort was taken to collect reliable instrument response information for the recording stations. In contrast, local site class information for these sites are not well characterised. For all of the recording stations, we assumed a generic rock site with V S30 of 760 m/s, which is similar to average site conditions obtained from geotechnical studies at several seismograph stations across Australia (Collins et al., 2006; Kayen et al., 2014). Recently, McPherson (2017) published a revised seismic site conditions map for Australia based on surficial geology information. This map can provide a good first-order approximation to assess local site conditions of recording stations included in our database. For future work, we also recommend taking advantage of the available seismic waveforms to explore local site characteristics using empirical and theoretical approaches. For an example, Zhao et al. (2016) classified strong-motion stations in Japan based on the horizontal-to-vertical spectral ratio curves of ground-motion recordings. It should be also noted that in the compiled ground-motion database, the majority of data are recorded at hypocentral distances between 30 and 300 km from earthquakes with magnitudes in the range of 3.8 to 4.8. Consequently, the performance of the candidate ground-motion models can only be verified within the aforementioned distance and magnitude ranges. The authors intend to expand this database by adding additional existing ground-motion data from national and regional seismic networks, as well as new data as they become available. In addition, the current ground-motion processing practices can be revised by manual and record-specific processing rather than the automatic processing used herein. Furthermore, the use of acausal filtering methods rather than causal methods is recommended as the response spectra computed from causally-filtered accelerations are sensitive to the choice of filter corner periods even at short periods (Figure 6). In undertaking this task for developing a systematic and rigorous ground-motion database, a schema will be developed (or adopted) to systematically store earthquake time histories and associated metadata in a transparent and accessible way. Selection and ranking of GMMs for NSHA18 19

26 Qualitative and quantitative techniques have been applied to verify and compare the performance of the selected ground-motion models against Australian ground-motion data. The qualitative assessments are conducted through visual inspection of the model fit to the observations, and also through the analysis of the model-specific distribution of the residuals. Using the method of Scherbaum et al. (2009), quantitative log-likelihood values were calculated to further evaluate ground-motion models against recorded data. Beauval et al. (2012) showed the applicability and stability of this measure against small-magnitude, as well as small number of data recorded in France. This is in good agreement with our results that shows a good correlation between performance of the selected ground-motion models using qualitative measures with those inferred based on computed LLH values. However, it should be noted that as shown here, the LLH measure reflects the overall performance of the model against the entire dataset, and it may undermine some desirable features of a ground-motion model such as model performance against near-field data, for example. The results suggest that the performance of the selected models varies among the studied tectonic regions (cratonic or non-cratonic regions), distance ranges (<150 km and km), and periods (0.1 and 1.0 sec). However, as demonstrated in this study (Figure 8), the performance of the models is quite sensitive to the number of observations and the results should be interpreted with great caution for datasets with limited observations. This clearly demonstrates the ongoing need for expert judgment in the process of selecting and ranking of ground motion models for 2018 Australian National Seismic Hazard Assessment. However, the results of this study provide a framework to assist with the decisionmaking process. Selection and ranking of GMMs for NSHA18 20

27 7 References Allen, T., J. Griffin, M. Leonard, D. Clark, and H. Ghasemi (2018). The 2018 National Seismic Hazard Assessment for Australia: model overview, Geoscience Australia Record 2018/27, Canberra, doi: /Record Allen, T., M. Leonard, and C. Collins (2011). The 2012 Australian Seismic Hazard Map catalogue and ground motion prediction equations, Australian Earthquake Engineering Society 2011 Conference, Barossa Valley, South Australia. Allen, T. I. (2012). Stochastic ground-motion prediction equations for southeastern Australian earthquakes using updated source and attenuation parameters, Geoscience Australia Record 2012/69, Canberra, pp 55. Allen, T. I., T. Dhu, P. R. Cummins, and J. F. Schneider (2006). Empirical attenuation of groundmotion spectral amplitudes in southwestern Western Australia, Bull. Seism. Soc. Am. 96, , doi: / Allen, T. I., G. Gibson, A. Brown, and J. P. Cull (2004). Depth variation of seismic source scaling relations: implications for earthquake hazard in southeastern Australia, Tectonophys. 390, 5 24, doi: /j.tecto Ancheta, T. D., R. B. Darragh, J. P. Stewart, E. Seyhan, W. J. Silva, B. S.-J. Chiou, K. E. Wooddell, R. W. Graves, A. R. Kottke, D. M. Boore, T. Kishida, and J. L. Donahue (2014). NGA-West2 database, Earthq. Spectra 30, , doi: /070913EQS197M. Atkinson, G. M., and D. M. Boore (2006). Earthquake ground-motion predictions for eastern North America, Bull. Seism. Soc. Am. 96, , doi: / Beauval, C., H. Tasan, A. Laurendeau, E. Delavaud, F. Cotton, P. Guéguen, and N. Kuehn (2012). On the testing of ground-motion prediction equations against small-magnitude data, Bull. Seism. Soc. Am. 102, , doi: / Bommer, J. J., J. Douglas, F. Scherbaum, F. Cotton, H. Bungum, and D. Fäh (2010). On the selection of ground-motion prediction equations for seismic hazard analysis, Seism. Res. Lett. 81, , doi: /gssrl Bommer, J. J., P. J. Stafford, J. E. Alarcón, and S. Akkar (2007). The influence of magnitude range on empirical ground-motion prediction, Bull. Seism. Soc. Am. 97, Boore, D. M., J. P. Stewart, E. Seyhan, and G. M. Atkinson (2014). NGA-West 2 equations for predicting PGA, PGV, and 5%-damped PSA for shallow crustal earthquakes, Earthq. Spectra 30, , doi: /070113EQS184M. Boore, D. M., J. Watson-Lamprey, and N. A. Abrahamson (2006). Orientation-independent measures of ground motion, Bull. Seism. Soc. Am. 96, Bozorgnia, Y., N. A. Abrahamson, L. Al Atik, T. D. Ancheta, G. M. Atkinson, J. W. Baker, A. Baltay, D. M. Boore, K. W. Campbell, B. S.-J. Chiou, R. Darragh, S. Day, J. Donahue, R. W. Graves, N. Gregor, T. Hanks, I. M. Idriss, R. Kamai, T. Kishida, A. Kottke, S. A. Mahin, S. Rezaeian, B. Rowshandel, E. Seyhan, S. Shahi, T. Shantz, W. Silva, P. Spudich, J. P. Stewart, J. Watson-Lamprey, K. Wooddell, and R. Youngs (2014). NGA-West2 research project, Earthq. Spectra 30, , doi: /072113EQS209M. Brown, A., G. Gibson, C. Sinadinovski, and K. McCue (2001). Measurements of PGA and attenuation in southeastern Australia, New Zealand Society of Earthquake Engineering 2001 Conference, Wairakei, New Zealand, Paper No Chiou, B. S.-J., and R. R. Youngs (2008). An NGA model for the average horizontal component of peak ground motion and response spectra, Earthq. Spectra 24, , doi: / Chiou, B. S.-J., and R. R. Youngs (2014). Update of the Chiou and Youngs NGA model for the average horizontal component of peak ground motion and response spectra, Earthq. Spectra 30, , doi: /072813EQS219M. Selection and ranking of GMMs for NSHA18 21

28 Collins, C., R. Kayen, B. Carkin, T. Allen, P. Cummins, and A. McPherson (2006). Shear wave velocity measurement at Australian ground motion seismometer sites by the spectral analysis of surface waves (SASW) method, Australian Earthquake Engineering Society 2006 Conference, Canberra, ACT, Cotton, F., G. Pousse, F. Bonilla, and F. Scherbaum (2008). On the discrepancy of recent European ground-motion observations and predictions from empirical models: analysis of KiKnet accelerometric data and point-sources stochastic simulations, Bull. Seism. Soc. Am. 98, Cotton, F., F. Scherbaum, J. J. Bommer, and H. Bungum (2006). Criteria for selecting and adjusting ground-motion models for specific target regions: Application to Central Europe and rock sites, J. Seismol. 10, , doi: /s Douglas, J. (2018). Ground motion prediction equations , Department of Civil and Environmental Engineering, University of Strathclyde July 2018, pp 624. Edwards, B., C. Cauzzi, L. Danciu, and D. Fäh (2016). Region-specific assessment, adjustment, and weighting of ground-motion prediction models: Application to the 2015 Swiss seismichazard maps, Bull. Seism. Soc. Am. 106, , doi: / Gaull, B. A., M. O. Michael-Leiba, and J. M. W. Rynn (1990). Probabilistic earthquake risk maps of Australia, Aust. J. Earth. Sci. 37, , doi: / Ghasemi, H., J. Griffin, S. Heimann, M. Leonard, and T. Allen (2016). Towards a homogeneous earthquake catalogue for Australia, Australian Earthquake Engineering Society 2016 Conference, Melbourne, Victoria. Gibson, G., and V.-A. Dimas (2009). Earthquake hazard at Newcastle, Australian Earthquake Engineering Society 2009 Conference, Newcastle, New South Wales. Griffin, J., M. Gerstenberger, T. Allen, D. Clark, R. Cuthbertson, V.-A. Dimas, G. Gibson, H. Ghasemi, R. Hoult, N. Lam, M. Leonard, T. Mote, M. Quigley, P. Somerville, C. Sinadinovski, M. Stirling, and S. Venkatesan (2018). Expert elicitation of model parameters for the 2018 National Seismic Hazard Assessment: Summary of workshop, methodology and outcomes, Geoscience Australia Record 2018/28, Canberra, doi: /Record Hoult, R. D., A. Amirsardari, D. Sandiford, E. Lumantarna, H. M. Goldsworthy, G. Gibson, and M. Asten (2014). The 2012 Moe earthquake and earthquake attenuation in south eastern Australia, 2014 Australian Earthquake Engineering Society Conference, Lorne, Victoria. Johnston, A. C. (1994). Seismotectonic interpretations and conclusions from the stable continental region seismicity database; In Johnston, A. C., Coppersmith, K. J., Kanter, L. R., and Cornell, C. A., Eds., The earthquakes of stable continental regions, Volume 1-Assessment of large earthquake potential, Electric Power Research Institute, Palo Alto, California, TR V1. Kaklamanos, J., L. G. Baise, and D. M. Boore (2011). Estimating unknown input parameters when implementing the NGA ground-motion prediction equations in engineering practice, Earthq. Spectra 27, Kayen, R. E., B. A. Carkin, T. Allen, C. Collins, A. McPherson, and D. Minasian (2014). Shearwave velocity and site-amplification factors for 50 Australian sites determined by the spectral analysis of surface waves method, U.S. Geological Survey Open-File Report Liang, J. Z., H. Hao, B. A. Gaull, and C. Sinadinovski (2008). Estimation of strong ground motions in southwest Western Australia with a combined Green's function and stochastic approach, J. Earthq. Eng. 12, McPherson, A. A. (2017). A revised seismic site conditions map for Australia, Geoscience Australia, Pezeshk, S., A. Zandieh, and B. Tavakoli (2011). Hybrid empirical ground-motion prediction equations for eastern North America using NGA models and updated seismological parameters, Bull. Seism. Soc. Am. 101, , doi: / Rietbrock, A., F. Strasser, and B. Edwards (2013). A stochastic earthquake ground-motion prediction model for the United Kingdom, Bull. Seism. Soc. Am. 103, 57 77, doi: / Selection and ranking of GMMs for NSHA18 22

29 Scherbaum, F., E. Delavaud, and C. Riggelsen (2009). Model selection in seismic hazard analysis: an information-theoretic perspective, Bull. Seism. Soc. Am. 99, , doi: / Silva, W., N. Gregor, and R. Darragh (2002). Development of regional hard rock attenuation relations for cenetral and eastern North America, Pacific Engineering and Analysis, El Cerrito, CA. Somerville, P., R. Graves, N. Collins, S.-G. Song, S. Ni, and P. Cummins (2009). Source and ground motion models for Australian earthquakes, Australian Earthquake Engineering Society 2009 Conference, Newcastle, New South Wales. Stewart, J., D. M. Boore, K. Campbell, M. Erdik, and W. Silva (2011). PEER GEM - Global GMPEs Task 1.b: Define a consistent strategy to model ground motion - site effects in parametric ground motion models. Stewart, J. P., J. Douglas, M. Javanbarg, Y. Bozorgnia, N. A. Abrahamson, D. M. Boore, K. W. Campbell, E. Delavaud, M. Erdik, and P. J. Stafford (2014). Selection of ground motion prediction equations for the Global Earthquake Model, Earthq. Spectra, Wells, D. L., and K. J. Coppersmith (1994). New empirical relationships among magnitude, rupture length, rupture width, rupture area, and surface displacement, Bull. Seism. Soc. Am. 84, Yenier, E., and G. M. Atkinson (2015). Regionally adjustable generic ground-motion prediction equation based on equivalent point-source simulations: Application to central and eastern North America, Bull. Seism. Soc. Am. 105, , doi: / Zhao, J. X., J. Zhang, A. Asano, Y. Ohno, T. Oouchi, T. Takahashi, H. Ogawa, K. Irikura, H. K. Thio, P. G. Somerville, Y. Fukushima, and Y. Fukushima (2006). Attenuation relations of strong ground motion in Japan using site classification based on predominant period, Bull. Seism. Soc. Am. 96, , doi: / Zhao, J. X., S. Zhou, J. Zhou, C. Zhao, H. Zhang, Y. Zhang, P. Gao, X. Lan, D. Rhoades, Y. Fukushima, P. G. Somerville, and K. Irikura (2016). Ground motion prediction equations for shallow crustal and upper mantle earthquakes in Japan using site class and simple geometric attenuation functions, Bull. Seism. Soc. Am. 106, , doi: / Selection and ranking of GMMs for NSHA18 23