A series of source inversions are performed for the 25 May 2009 (Memorial Day) North Korean seismic event using intermediate period (10–50 s) complete waveform modeling. An earthquake source is inconsistent with the data and the best-fit full seismic moment tensor is dominantly explosive (∼60%) with a moment magnitude (MW) of 4.5. A pure explosion solution yields a scalar seismic moment of 1.8 × 1022 dyne-cm (MW4.1) and fits the data almost as well as the full solution. The difference between the full and explosion solutions is the predicted fit to observed tangential displacement, which requires some type of non-isotropic (non-explosive) radiation. Possible causes of the tangential displacement are additional tectonic sources, tensile failure at depth, and anisotropic wave propagation. Similar displacements may be hidden in the noise of the 2006 event. Future analyses of this type could be used to identify and characterize non-earthquake events such as explosions and mine collapses.
 In previous work, Ford et al.  showed it was possible to identify explosions at the Nevada Test Site with a full moment tensor inversion of seismic waveforms from nearby stations in the Western US. In this paper, we extend this analysis to the 2009 Memorial Day event near Kimchaek, North Korea and show that the best-fit source is dominantly isotropic, which is consistent with an explosion. We also compare the 2009 event with the 9 October 2006 declared North Korean nuclear explosion, finding very similar results, with the 2006 source strength approximately five to six times smaller than the 2009 event.
2. Data and Method
 Three component waveform data from Global Seismograph Network (GSN) and China Digital Seismograph Network (CDSN) stations MDJ, INCN, BJT and HIA along with station TJN from the Ocean Hemisphere Project Seismological Network (Figure 1) are instrument corrected, integrated to displacement, and band-pass filtered using an 6th order acausal Butterworth filter with corners at 0.02 and 0.10 Hz. Green's functions were computed using the MDJ2 velocity model (Table 1) for stations MDJ, INCN, and TJN, and filtered similarly to the data. The velocity model is a modification of the surface-wave-derived MDJ model [Nguyen, 1994]. The Green's functions were aligned with the data based on a location and origin time of 00:54:43.38, 25 May 09, 41.2986°/129.0694° (D. Dodge, unpublished data, 2009). Green's functions from station TJN required a shift of 2 sec relative to the data to produce a good match, which may be due to error in the assumed velocity model along this path. Ford et al.  found that shifts of less than a half-cycle still produced accurate solutions.
 However, the MDJ2 velocity model was too fast at the farther stations, BJT and HIA (1099 and 1147 km, respectively) and required shifts greater than a half-cycle (∼5 s). Therefore, the velocity model was modified below the source layer to better model the waveforms at these distant stations. These changes are given in Table 1. A source depth of 600 m was initially assumed, though the solution sensitivity to this assumption will be tested. Explosion, earthquake (double-couple or DC), deviatoric, and full moment tensor solutions were evaluated and are compared in Figure 2. The explosion and DC solutions were obtained with a grid-search to find the best-fitting parameters, whereas the deviatoric and full moment tensor solutions were calculated with a least-squares linear inversion. Details of the inversion are given by Minson and Dreger .
 The pure explosion model is able to fit the waveforms with a variance reduction of 75% and yields an isotropic moment of 1.8 × 1022 dyne-cm (MW4.1; all seismic moment values are calculated with the method of Bowers and Hudson ). In contrast, the pure DC earthquake solution fits the data much worse at 52% with M0 = 3.8 × 1022 dyne-cm (MW4.4). The fact that the single degree of freedom explosion model fits so much better than the four degree of freedom DC model is highly significant and indicates that such a comparison can be a useful discriminant. The strike, rake, and dip of the best-fit DC is 50°, −85°, and 10°, respectively. Such a steep dip-slip mechanism is unusual for an earthquake. Of all sources calculated by the Global CMT Project (globalcmt.org) less than 1.6% have dips less than 10°. This type of information can be used as an additional indication of anomalous sources. The differences in the fits between explosion (Best-iso) and DC (Best-DC) sources can be viewed in Figure 2a, where the DC overpredicts the Love wave amplitude on the transverse components at almost all stations and underpredicts the Rayleigh wave amplitudes, especially at station INCN.
 Comparisons of waveforms and spectra for the 9 October 2006 nuclear test and 2009 event indicates that the 2009 event is approximately 5.7 times larger than the 2006 event. The isotropic moment for the 2006 event was found to be 0.3 × 1022 dyne-cm [Walter et al., 2007], which agrees with the Koper et al.  value. Therefore, from the waveform comparison the 2009 event should be approximately 1.5 × 1022 to 2.1 × 1022 dyne-cm (scale factor of 5 to 7); a range that includes the pure explosion result (1.8 × 1022 dyne-cm) obtained from waveform modeling.
 The full moment tensor inversion fits the data at 81% and yields an isotropic moment of 3.6 × 1022 dyne-cm, and a total moment of 6.3 × 1022 dyne-cm (MW4.5). The deviatoric moment tensor inversion fits the data at 80% and a total moment of 3.2 × 1022 dyne-cm (MW4.3). If the deviatoric source is decomposed to a compensated linear vector dipole (CLVD [Knopoff and Randall, 1970]) and DC sharing the same principal axes, then the source is 70% CLVD. The similarity in fits between the dominantly CLVD deviatoric source and dominantly isotropic full moment tensor shows that at shallow depths, a vertical CLVD mechanism can effectively mimic an explosion at the distances and periods analyzed here. This can be seen in the waveform comparison in Figure 2a. The full moment tensor isotropic moment is two times larger than the pure explosion indicating that the compound source of the full moment tensor solution (DC + CLVD + Isotropic) required to fit the Love waves also modifies the Rayleigh waves causing the isotropic component to increase in order to compensate.
 The source-type parameters, ɛ and k [Hudson et al., 1989], for each source inversion are calculated and plotted on the source-type plot in Figure 2b. ɛ measures the ratio of deviatoric eigenvalues and plots along the horizontal axis. Here k measures the relative amount of isotropic moment and plots along the vertical axis. The axes are then transformed so as to best represent error in these parameters as a normal distribution (ellipse on the source-type plot). Using this plot, one can discern the relative source components allowing for a potential seismic discriminant. For example, sources that deviate from the center of the plot have a large non-DC component. The best-fitting sources are far from the center, indicating that the source is anomalously non-DC. Along with the best-fitting sources (corresponding to those shown in Figure 2a), Figure 2b plots solutions and their 95% error ellipses for explosions at the Nevada Test Site (NTS) and earthquakes in the Western US from Ford et al. . The best-fit full moment tensor plots in the same region as the explosions and away from the earthquake population. The error ellipses for the best-fit sources from this study are all smaller than the symbol used to plot the solutions due to the very high signal-to-noise of the event records.
 Unlike earthquake inversions, full moment tensor inversions of an explosive source cannot constrain the source depth by comparing fits at different depths at the frequencies examined here. Event locations put the source at less than 1 km, so the results discussed above assume a source depth of 600 m. However, the isotropic component and total moment are dependent on the assumed depth. Ford et al.  showed that though the source-type is robust, the total moment trades-off with the assumed depth, and more moment goes into the vertical elements of the moment tensor as the depth decreases. Since the Green's functions for these elements vanishes in the limit as the depth of the source decreases (due to vanishing traction at the free-surface) these elements grow large and increase the total moment, thereby reducing the percentage of isotropic moment. We vary the depth between 300 and 900 m and calculate the isotropic component as well as the total moment and give the results in Figure 3. The variance reduction at each of these depths is approximately the same, so an analysis of all the solutions less than 1 km is necessary. The median isotropic component is 57% (range = 48–63%) with a median total moment of 6.3 × 1022 dyne-cm (range = 5.2–9.7 × 1022 dyne-cm) (MW4.5; range = 4.4–4.6).
 As mentioned previously, some non-isotropic radiation is required to fit the source due to the observed Love waves as can be seen on the tangential component waveforms in Figure 2a. The same amount of non-isotropic energy could have been present in the recordings of the 9 October 2006 nuclear test, but were obscured due to noise. Figure 4 shows the raw waveforms at station MDJ for both the 2006 and 2009 North Korea events filtered between 10 and 50 sec. When the waveforms of the 2006 test are magnified by a factor of six, the north-south and vertical components are very similar to the 2009 event. However, the tangential energy that is clear in the 2009 event due to the high SNR is still too small to peak above the noise in the 2006 event.
 There are several possible causes of the energy seen on the tangential component. The compressional (P) and tensional (T) axes of the full moment tensor (shown in Figure 2b) are consistent with principal strain axes calculated for South Korea by Jin and Park . The P and T axes are also similar to those from a study by Herrmann et al.  of earthquakes in the Yellow Sea/Korean Peninsula (http://www.eas.slu.edu/Earthquake_Center/MECH.KR). Tangential displacements could have been caused by tectonic release that accompanied the explosion in response to crustal stresses. Another possible source is shock-induced, deep-seated tensile failure as described by Patton and Taylor . Anisotropic radiation could be created if the CLVD in their model were tilted several degrees off vertical. However, both these models cannot explain the displacements seen at all stations, where for example the DC component of the full solution fails to match the amplitude at station INCN and the CLVD component of the deviatoric solution fails to match the phase at station HIA. A third cause of displacement recorded on the tangential component is multipathing and anisotropic wave propagation through a heterogeneous crust and mantle. Pasyanos et al.  imaged very thick sediment along the path to BJT in the vicinity of Bohai Bay (Figure 1), and Hong et al.  note effects caused by interaction with the continental margin along the paths to stations INCN and TJN. Rodgers et al.  and Kim et al.  have shown that, even at the periods examined in this study, three-dimensional effects can have a large impact on observed displacements.
 The MS:mb of the 2006 nuclear test is similar to some Eurasian earthquakes [Bonner et al., 2008, Figure 3] and dissimilar to NTS and STS explosions [Patton and Taylor, 2008, Figure 4]. Patton and Taylor  suggest the greater than expected MS of the 2006 explosion is due to the lack of a tensile failure component (modeled as a CLVD). Since at periods greater than at least 10 sec the 2009 event is a scaled version of the 2006 event (Figure 4), the results presented here could be used to comment on the Patton and Taylor  hypothesis regarding the MS:mb performance of the older test. However, due to the long periods analyzed here and the uncertainty in absolute depth, the relative contribution of tensile and monopole (CLVD and explosion) components cannot be constrained (Figures 2 and 3) and this study cannot support or refute the Patton and Taylor  hypothesis. The results presented here do show that there is a non-isotropic component to the radiation pattern of the 2009 event, and the similarity of the waveforms suggests the same for the 2006 event. Bonner et al.  comment that there is some scatter in the MS measurements of the 2006 nuclear test, which they speculate could be due to path-specific effects or a Rayleigh-wave radiation pattern. This study shows that indeed there should be a pattern associated with the amplitudes of the 2006 event.
 Modeling of intermediate period, regional distance waveforms identifies the Memorial Day event in Kimchaek, North Korea as decidedly non-tectonic with the best-fit model dominated by an explosion source. While the source type is well determined to be non-DC, the isotropic moment of the full moment tensor inversion has some uncertainty and the MW is between 4.4 and 4.6. Comparison of pure explosion and pure double-couple models indicate that the simpler explosion model fits the waveform data substantially better than the higher degree of freedom DC model, where the isotropic moment of the explosion model is 1.8 × 1022 dyne-cm (MW4.1). However, there are Love waves observed at several stations indicating that the source must have some non-isotropic component. Possible causes of the tangential displacement are additional tectonic sources, tensile failure at depth, and anisotropic propagation. The same non-isotropic component could have been present in the previous 2006 test, but was masked by the noise.
 We are grateful for very helpful internal reviews by Artie Rodgers and Mike Pasyanos. This is LLNL contribution LLNL-JRNL-414641 and BSL contribution 09-21. DD was supported by the National Nuclear Security Administration, contract DE-FC52-06NA27324.