Improved Resolution across the Global Seismographic Network: A New Era in Low‐Frequency Seismology Open Access

The Global Seismographic Network (GSN) is a network of approximately 150 globally distributed stations designed to provide high‐fidelity digital recordings of the largest earthquakes (⁠$Mw≈9.5$⁠) spanning the entire range of earthquake signals from free oscillations (0.1 mHz) to teleseismic body waves (≈15 Hz) (Lay et al., 2002). Although this goal has largely been met (Butler et al., 2004), free oscillation observations below 1 mHz remain difficult to record. For example, few stations (e.g., less than 20) can record low‐frequency toroidal spectra (analogous to Love waves) for even the largest earthquakes (Schneider and Deuss, 2021). Increased availability of high‐fidelity low‐frequency data should substantially improve normal‐mode measurements, and provide additional constraints on the velocity and density estimates of the Earth. Potential improvements enabled by additional high signal‐to‐noise ratio (SNR) low‐frequency data include more accurate estimates of splitting parameters and will enhance normal‐mode spectra inversions (e.g., Li et al., 1991; Jagt and Deuss, 2021).

Deviations of the Earth from being a spherical nonrotating elastic isotropic Earth cause the eigenfrequencies of the Earth to split (Dahlen and Tromp, 1998). The splitting of free oscillations contains information about rotation, ellipticity, as well as Earth structure. Estimation of the splitting parameters of the gravest free oscillations of the Earth requires various stacking and stripping methods (e.g., Buland et al., 1979). The estimation of free oscillation splitting parameters depends on various approximations of mode coupling (e.g., self‐coupling, group‐coupling, or full‐coupling; Dahlen and Tromp, 1998). Such approximations ultimately limit our ability to extract the 3D density structure of the Earth using normal modes (Al‐Attar et al., 2012; Akbarashrafi et al., 2018). To capture the splitting of free oscillations, it is necessary to use relatively long time series (≈1.1Q cycles; in which a Q cycle is the time for the signal amplitude to decay by $e−π$⁠; Dahlen 1982) as well as isolate the spectrum around the frequency band of interest. Free oscillations fall into two categories based on the nature of their motions: spheroidal modes, governed by vertical and horizontal surface motion, and toroidal modes, governed by only horizontal surface motion. Therefore, it is important not only to have high‐quality vertical seismic records but also horizontal component records to measure the splitting of toroidal modes or more complicated cross‐coupling of spheroidal modes (e.g., between fundamental spheroidal and toroidal modes).

Low‐frequency data quality is controlled by two major factors: (1) background “nonseismic” noise at the location and (2) the response and self‐noise of the sensor. Nonseismic noise at low frequencies is typically dominated by temperature and pressure variations (e.g., Ringler et al., 2020, and references therein), which are generally more pronounced for surface and near‐surface vault emplacements and have less effect for borehole emplacements due to better thermal stability and attenuation of pressure variations. Unfortunately, until recently, GSN borehole sensors (such as the Geotech KS‐54000) tended to demonstrate poorer performance at low frequencies (Ringler and Hutt, 2010; Laske and Widmer‐Schnidrig, 2015) than the GSN vault sensors (typically Streckeisen STS‐1), and therefore not all the potential benefits of a borehole emplacement were realized. Recently, very broadband sensors have become available in two different configurations—those that can be installed in a vault or those that are designed to be installed in a posthole or borehole. The ongoing installation of these sensors as replacements for the original GSN sensors promises to dramatically increase the amount of high‐quality low‐frequency data. The goal of this article is to evaluate the impact of these ongoing sensor replacements on data quality at the frequencies (<1 mHz) necessary for normal‐mode observations and to detail expectations for the future.

Until recently, the STS‐1, which can only be installed in vaults (with a flat velocity response to 360 s period), has been the only force‐balance seismometer with self‐noise levels low enough to allow the recording of data necessary for low‐frequency (≈0.1 mHz) spectra. Although many STS‐1 seismometers remain in operation as of 2021 and continue to provide exceptionally high‐quality data (e.g., Forbriger et al., 2021), the continued operation of these sensors has become more challenging as the components gradually age, the instruments are no longer in production, and support from the manufacturer has ended.

The new very broadband (VBB) sensors, namely the Nanometrics T‐360GSN and the Streckeisen STS‐6 (both with a flat to velocity response to 360 s period; Fig. 1), have been acquired, and over the last five years 45 of these new sensors have been installed at GSN stations. The T‐360GSN is available as vault, posthole, or borehole configurations, while the STS‐6 is designed for postholes and boreholes. This capability makes it possible to replace both the STS‐1 vault sensor and the KS‐54000 borehole seismometer, which are also no longer available or supported.

The effort to replace and develop a new generation of very broadband seismometers was laid out as a “high priority” in the Seismological Grand Challenges in Understanding Earth’s Dynamic Systems (Lay et al., 2009). The installation effort, which began in 2017, has been a combination of installing instruments at depth, drilling new boreholes, as well as repairing and reinstalling STS‐1 sensors at stations where the instruments have become compromised. Repairing the STS‐1 seismometers (e.g., Hutt and Ringler, 2011) has been possible from the surplus of STS‐1 seismometers removed from other stations where posthole and boreholes installations of STS‐6 and T‐360GSN sensors were feasible. At the end of 2021, the GSN was operating 45 STS‐1 seismometers, 31 Streckeisen STS‐6 seismometers, 14 T‐360GSN seismometers (13 borehole or posthole and 1 vault), and 7 KS‐54000 seismometers (many of these are performing poorly or not at all).

Given this large‐scale ongoing effort, it is essential to understand the potential impact on future observational capabilities of the GSN at low frequencies. Co‐located STS‐1 and STS‐6 sensors at the Black Forest Observatory (BFO) suggest that the STS‐1 may still retain superior performance at specific frequency bands (Forbriger et al., 2021), but it is unclear whether this extends across a wider frequency band or across the diverse range of deployments in the GSN.

In this study, we investigate the current ability of the GSN to make robust below 1 mHz normal‐mode observations by comparing data recorded after large earthquakes in 2021 with one in 2014. We use data following three large (⁠$Mw>8$⁠) earthquakes in 2021 and compare these to data collected following an $Mw$ 8.2 earthquake in 2014. These analyses allow us to systematically compare our ability to resolve normal modes between STS‐1s, KS‐54000, and newer replacement instruments. It also enables us to evaluate the current state of the remaining STS‐1s, many of which have been refurbished. In this work, we report on the positive effects that sensor upgrades have made to the observational abilities of the GSN to resolve normal modes.

To assess the current ability of the GSN to resolve normal modes from large earthquakes, we process seismic data from three large (⁠$Mw>8$⁠) earthquakes, which occurred in 2021 and one event which occurred in 2014 (Table 1). We note that the 2021 $Mw$ 8.1 Kermadec Islands event was preceded (<2 hr) by an $Mw$ 7.4 event at the same location and by an $Mw$ 7.3 (≈6 hr) about 900 km to the south. Similarly, the South Sandwich Islands event directly followed (<3 min) an $Mw$ 7.5 event at the same location. The $Mw$ 8.2 event in 2014 is therefore most comparable to the 2021 $Mw$ 8.2 Perryville, Alaska, earthquake (in terms of moment and depth; Table 1).

We use both qualitative and quantitative methods to assess how the sub‐1 mHz normal‐mode observations have changed between 2014 and 2021. For each event, we analyzed the excitation of modes below 1 mHz in two ways. First, we did a visual inspection of the spectra and discarded all spectra that did not contain high‐amplitude, high‐fidelity normal modes below 1 mHz (i.e., these discarded spectra were dominated by noise or spectral peaks that did not correspond to normal‐mode frequencies of the Earth). Specifically, we looked for excitation of the modes $S00$ (0.81 mHz), $S20$ (0.31 mHz), $S30$ (0.47 mHz), $S40$ (0.65 mHz), $S50$ (0.84 mHz), $S21$ (0.68 mHz), and $S31/3S1$ (0.939 mHz; Fig. 2). Very few stations were able to record $S20$ with high fidelity. Here, we simply compare the number of stations that we deemed capable of making normal‐mode observations for each event. For this we collected 126 hr of data (1.1Q cycles for $S50$⁠) starting at the event origin time for all stations in the GSN (Fig. 1; network codes: IC, II, and IU). We restrict our analysis to the vertical components and primary sensor (location code: 00, which is usually the highest quality sensor) at each station. We remove the linear trend and applied a Kaiser taper window with a side lobe tradeoff of $2π$⁠. We then zero padded the data to 20 days and calculated the discrete Fourier transform.

In Figure 3, we plot the SNR for each station and event in our study, and consider stations with an SNR > 2.5 as having “good” fidelity normal‐mode observations. This method provides an objective criterion but makes no claim of the usefulness (e.g., spectra with a high SNR but various other odd characteristics). We did not analyze $S00$⁠, as it requires a much longer time window for analysis (1.1Q cycles for $S00$ is approximately 83 days) or any of the other modes because they were not well excited.

For the 1 April 2014 $Mw$ 8.2 earthquake, we visually found 18 stations that were able to record sub‐1 mHz spectra (Fig. 2a). Notably, these 18 stations all had Streckeisen STS‐1s as their VBB sensor. Additionally, although these 18 stations clearly recorded $S00$ and $S50$⁠, these stations were unable to record $S20$⁠, and the SNR of $S30$ was relatively low at some stations (e.g., COR and CTAO). In total, we calculated spectra for 114 GSN stations for the 2014 event and found that only 21 stations (18%) had an SNR greater than 2.5 for the four modes analyzed (Fig. 3a).

In contrast, for the similarly sized 2021 $Mw$ 8.2 Perryville, Alaska, earthquake, we found 36 stations that visually recorded sub‐1 mHz spectra with high fidelity (Fig. 2b). Of these stations, 16 use STS‐6 seismometers, 17 use STS‐1s, 1 uses a T-360GSN borehole sensor (EFI), 1 uses a Nanometrics T-240 (SIMI), and 1 uses a Streckeisen STS-2.5 (TLY). We calculated spectra at 108 stations for this event and found that 37 (34%) recorded with an SNR greater than 2.5 for the normal modes examined (Fig. 3c). Similar trends are observed for the two other $Mw$ 8 events in 2021, with 32 (28%) of GSN stations recording the $Mw$ 8.1 Kermadec Islands event with an SNR of at least 2.5 (Fig. 3b) and 44 (41%) recording the $Mw$ 8.1 South Sandwich Islands earthquake. Overall, between 2014 and 2021, we see that the percentage of GSN stations that record the modes analyzed in our study after $Mw$ 8 earthquakes with “good” fidelity (SNR > 2.5) has jumped from 18% to over 30%. To verify that the improvements in our spectra are coming from lower noise levels and not just earthquakes that produces larger amplitude signals in this band, we estimate noise levels by sensor type for the frequency band of the four modes in our analysis (Fig. 4). From Figure 4 we see that the Streckeisen STS‐6 (orange) record the modes with the highest fidelity (the lowest nearby noise levels), suggesting an overall improvement in the network by installing these new instruments.

We attribute this increase in the improvement in normal‐mode spectra at GSN stations to recent sensor and infrastructure upgrades (Ringler et al., 2020). For example, on 1 April 2014, the GSN had 66 STS‐1 seismometers in operation and 31 KS‐54000 borehole seismometers. We did not find any KS‐54000 seismometers that were able to record sub-1 mHz spectra for the 2014 Iquique, Chile, earthquake, likely due to high self-noise levels of these instruments at low frequencies (Ringler and Hutt, 2010). Notably, Laske and Widmer‐Schnidrig (2015) had reported similar difficulties where the KS‐54000 at QSPA (South Pole, Antarctica) was unable to record $3S2$ at 1.10 mHz. We hope this situation will be rectified with the deployment of a very broadband seismometer installed at greater than 2 km at QSPA (Anthony et al., 2021). Although STS‐1 seismometers are about 15 dB quieter than KS‐54000s at 1 mHz (Ringler and Hutt, 2010), they must be installed in vaults and thus not useful to replace the KS‐54000 instruments.

In total, we found that only 17 of the 44 GSN stations operating STS‐1 seismometers were visually able to record $Mw$ 8.2 Perryville, Alaska event with high fidelity (Fig. 2b). We attribute this relatively low number (39%) of observations to many low-frequency STS‐1 records being compromised by nonseismic noise from the surface installations. In contrast, 53% of the stations operating STS‐6s visually recorded multiple sub-1 mHz normal modes from this event with high fidelity. The number of 1-mHz spectra is similar for the $Mw$ 8.1 Kermadec Islands and the $Mw$ 8.2 Chignik earthquakes. We directly attribute these improvements in spectral records to the lower noise levels of the STS‐6 (Fig. 4).

From Figure 3b–d we also see that the number of stations that were able to record the normal modes for events in 2021 with an SNR greater than 2.5 was far greater than for the 2014 event (Fig. 3a). Notably, roughly 60% (Table 1) of all GSN stations operating STS‐6s can record spectra of each of the three 2021 $Mw>8$ events. Consistent with visual observations, this “successful observation rate” is about 20% higher than stations operating STS‐1s for these same 2021 events. Therefore, both visual and quantitative analysis of normal-mode spectra suggests that we can record sub-1 mHz spectra at more stations in 2021, and that the fidelity of these spectra is higher.

We have shown that compared to the KS‐54000 sensors it replaces, across the GSN the STS‐6 can produce data with quality high enough to allow calculation of meaningful sub-1 mHz spectra for large earthquakes. Additionally, a higher percentage of STS‐6s in the GSN records normal-mode spectra than STS‐1s. We attribute these observations to the borehole and posthole installations where they are less susceptible to nonseismic noise sources compared to surface sensors. This suggests that at the network level the borehole sensors can produce superior low-frequency data as compared to the currently deployed STS‐1s.

Along with replacing the STS‐1s at many GSN stations, several stations have also been converted from vault to borehole installations. We expect, at these stations, that we will see improved (i.e., lower noise) vertical and horizontal low-frequency data. While our preliminary findings suggest that deep boreholes (e.g., greater than 100 m of overburden) with low-noise sensors are able to produce superior low-frequency data, such installations will not be possible everywhere (e.g., islands where drilling is logistically impractical). Our results can be viewed as an interim situation with continued network-wide improvements in the future, as more GSN stations are converted to boreholes. During these efforts, network operators rely on collocated instruments to verify the improvement from these new sensors. For example, Ringler et al. (2020) and Forbriger et al. (2021) carry out such collocated comparisons at a few select locations.

In comparison with the 2014 event, we see that we have increased the number of stations with high‐fidelity (SNR > 2.5) sub‐1 mHz spectra from approximately 18% in 2014 to over 30% in 2021. This suggests that we should record high‐quality low‐frequency spectra from smaller moderate‐to‐large earthquakes than was possible previously. In other words, observations of modes at many GSN stations that were previously only observable with $Mw$ 8.5 and greater events might now be observed with smaller $Mw$ 8 events. Between 2014 and 2021, the number of GSN stations capable of resolving sub‐1 mHz normal modes following $Mw$ 8 earthquakes has more than doubled. Although it is possible that certain GSN stations have recorded sub‐1 mHz normal modes for smaller earthquakes (e.g., deep events; Deuss et al., 2013), we think the events used in our analysis give a good estimate of the improvements in the network from 2014 to 2021. These changes will improve the global coverage over which we can observe low‐frequency spectra, which are critical to estimate the splitting parameters that help resolve 3D Earth structure.

With these improvements in low‐frequency observations, it could be possible to improve estimates of properties of the Earth’s interior. For example, by increasing the number of stations and frequency of observing low‐frequency spectra, it should be possible to increase the time resolution of identifying changes in any inner‐core rotation. This is because we will obtain more spectra for estimating splitting parameters over shorter time scales, as compared to the ±2 or more years from the previous studies (Laske and Masters, 1999). Uncertainty in splitting functions coming from self‐coupling, group‐coupling, and full‐coupling approximations (Akbarashrafi, et al., 2018) might also indirectly benefit from the increased resolution of free oscillation records at low frequencies. For example, the relatively low SNR that many spectra are recorded at could compromise the fidelity of the spectra and poorly fitting synthetics. Despite these improvements, seismic observations made on modern seismometers like the STS‐6 and T-360GSN are noisier than those made on superconducting gravimeters at frequencies below 1 mHz.

While, seismometers have substantially better global coverage than superconducting gravimeters (Häfner and Widmer‐Schnidrig, 2013). Therefore, normal‐mode measurements that require multiple measurements taken at different points on the Earth often rely on observations made on seismometers. Ultimately, improved fidelity of normal‐mode observations across the GSN should aid in estimating splitting parameters that are used for inversion of Earth structure.

Although this work has focused on vertical spectra, Ringler et al. (2020) showed that installing low-noise borehole sensors (e.g., STS‐6 or T-360GSN) at depth has the potential to increase the number of high-quality low-frequency horizontal spectra from large earthquakes. By isolating the sensors from surficial tilt noise, we can produce horizontal spectra well below many of the best-performing STS‐1 seismometers in the world (e.g., an STS‐6 in a shallow posthole can show low-frequency noise levels that are 10 dB lower than the STS‐1 at BFO, Rudolf Widmer‐Schnidrig, personal comm., BFO, 2021). We note that although horizontal noise levels remain well above the vertical noise levels at low frequencies, these reductions in tilt noise likely provide a means for recording infrequently measured toroidal modes. In turn, more common toroidal mode observations (e.g., Schneider and Deuss, 2021) could supplement vertical‐component normal‐mode observations and better constrain the known amplitude uncertainties in low‐frequency toroidal modes observed by Park et al. (2008) on underground laser extensometers. In combination with vertical‐component records, they could improve observations of cross‐coupling between fundamental spheroidal and toroidal modes, which might in turn provide constrains on Earth’s anisotropic structure (Park and Yu, 1992; Park, 1993).

We used data from the Incorporated Research Institutions for Seismology (IRIS)/U.S. Geological Survey (USGS) network (network code IU; Albuquerque Seismological Laboratory (ASL)/USGS, 1988), the New China Digital Seismograph Network (network code IC; ASL/USGS, 1992), as well as the International Deployment of Accelerometers (IDA) network (network code II; Scripps Institution of Oceanography, 1986). These data are all freely available at the IRIS Data Management Center (DMC). We made use of ObsPy in the analysis of this work (Megies et al., 2011), and the scripts to reproduce the figures in this article is available at https://code.usgs.gov/asl/papers/ringler/normal_mode_gsn (last accessed April 2022). The facilities of IRIS Data Services, and specifically the IRIS DMC, were used for access to waveforms, related metadata, and derived products used in this study. IRIS Data Services are funded through the Seismological Facilities for the Advancement of Geoscience and EarthScope (SAGE) Proposal of the National Science Foundation under Cooperative Agreement EAR‐1261681. The Global Seismographic Network (GSN) is a cooperative scientific facility operated jointly by the IRIS, the USGS, and the National Science Foundation (NSF), under Cooperative Agreement EAR‐1261681. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

The authors declare that there are no conflicts of interest recorded.

This work has benefited from discussions with colleagues at the Black Forest Observatory (Thomas Forbriger, Rudolf Widmer‐Schnidrig, and Walter Zürn) as well as Robert Freudenmann and Joseph Steim. The authors thank two anonymous reviewers for a very careful review that improved the article. The authors thank Jessica Irving for handling the manuscript and additional comments that lead to Figure 4. The authors also thank James Holland, Bob Hutt, Brian Shiro, and Janet Slate for helpful reviews of this manuscript.