Abstract
Seismograms from two borehole seismometers near the 2019 Ridgecrest, California, aftershock sequence do not return to pre‐mainshock noise levels for over ten days after the M 7.1 Ridgecrest mainshock. The observed distribution of root mean square amplitudes in these records can be explained with the Reasenberg and Jones (1989) aftershock occurrence model, which implies a continuous seismic “hum” of overlapping aftershocks of M > −2 occurring at an average rate of 10 events per second after ten days, which prevents observing the background aseismic noise level at times between the body‐wave arrivals from cataloged and other clearly observed events. Even after the borehole noise levels return at their quietest times to pre‐mainshock conditions, the presence of overlapping low‐magnitude earthquakes for 80 days is implied by waveform cross‐correlation results provided using the matrix profile method. These results suggest a hidden frontier of tiny earthquakes that potentially can be measured and characterized even in the absence of detection and location of individual events.
Introduction
Aftershock rates are very high immediately following large earthquakes and then generally decay roughly as 1/t following Omori’s law (e.g., Utsu et al., 1995). Rates are so extreme during the first few minutes to hours that it becomes difficult to detect, associate phases, and locate the many overlapping events, resulting in a noticeable deficit of smaller earthquakes in the catalog during these times (e.g., Kagan, 2004; Kagan and Houston, 2005; Lolli and Gasperini, 2006) and challenging attempts to estimate true earthquake rates and determining if and when the 1/t rate curve flattens as t → 0 (e.g., Peng et al., 2006, 2007; Enescu et al., 2007; Zhuang et al., 2017; Lippiello, Cirillo, et al., 2019; Lippiello, Petrillo, et al., 2019; Sawazaki, 2021). Even after the aftershock rate decays to more manageable levels for reliable earthquake detection and location, it remains significantly elevated for months to years after the mainshock, depending upon the aftershock productivity of the mainshock and the pre‐mainshock background rate. Studies of aftershock sequences often involve efforts to create more complete aftershock catalogs using advanced methods, such as template matching or machine learning (e.g., Ross et al., 2019; Zhu et al., 2019; Gulia et al., 2020; Liu et al., 2020; Mousavi et al., 2020; Shelly, 2020; Zhou et al., 2022; Fonzetti et al., 2024; Si et al., 2024). Aftershock modeling efforts often consider differences in aftershock decay rates, the role of secondary aftershocks, and variations in b‐value (e.g., Omi et al., 2013; Hainzl, 2016; Dascher‐Cousineau et al., 2020; van der Elst, 2021; Hainzl et al., 2024; Hardebeck et al., 2024).
Here, we adopt a new approach to analyze aftershocks of the 2019 Ridgecrest, California, M 6.4 and 7.1 mainshocks using continuous records from two nearby borehole seismometers, B918 and B921 (see Fig. 1). Because these sensors are located below 100 m depth, they suffer from less background noise (e.g., wind and anthropogenic signals) than surface seismometers. Rather than examining individual aftershocks, we focus on the time periods between aftershocks and search for the quietest such periods. We find that for over ten days after the Ridgecrest mainshocks, even the lowest amplitude signals remain noticeably stronger than the quietest pre‐mainshock noise levels. Although the seismograms during these times contain no visible earthquakes and visually resemble pre‐mainshock records, our results imply that continuous background “noise” during these ten days is actually composed of many tiny overlapping aftershocks. Modeling based on Omori’s law and the Gutenberg–Richter (G–R) law (e.g., Reasenberg and Jones, 1989) can explain the root mean square (rms) amplitude distribution of the borehole data and implies that earthquakes down to at least M −2 are required, which occur at rates of about 10 events per second even ten days after the mainshocks.
We also examine matrix profile (MP) results for the borehole records for the period from 10 days before the first mainshock to 80 days after the second mainshock. MP is a computationally efficient algorithm to obtain autocorrelation results for long time series (e.g., Yeh et al., 2016; Zhu et al., 2016; Zimmerman et al., 2019), that is, to cross‐correlate everything with everything (Shabikay Senobari et al., 2024). We find that MP minimum correlation values do not return to their pre‐mainshock condition even 80 days after the mainshocks, suggesting that nearly continuous aftershocks remain present and detectable in the borehole records even after their rms amplitudes have decreased to pre‐mainshock noise levels. These results have implications for earthquake monitoring efforts that focus on measuring and characterizing small earthquakes with magnitudes well below those currently studied.
The Ridgecrest Aftershock Sequence
The 4 July M 6.4 and 5 July 2019 M 7.1 Ridgecrest mainshocks produced an abundant and long‐lasting aftershock sequence that illuminated a complex network of both northwest and northeast‐trending faults (Fig. 1). The Southern California Seismic Network (SCSN) catalog recorded 25,918 aftershocks in the 50 days following the second mainshock. As shown in Figure 2, the SCSN rate of recorded events decreases during this time period, as expected from Omori’s law, and the catalog lacks small events in the first few hours to days, consistent with catalog incompleteness during a time of very high aftershock rates. However, the SCSN catalog also exhibits some abrupt changes that are likely artifacts caused by differences in network coverage and catalog operations. This can be seen most clearly in the temporal variations in the minimum magnitude and the sharp increases in the average interevent interval that occurred at about 7 and 25 days after the M 7.1 mainshock. A more complete catalog of 34,091 events is available for the initial 12 days of the Ridgecrest sequence (Shelly, 2020), obtained using template matching of the catalog events with continuous waveform data to identify additional events. This catalog behaves more smoothly with time than the SCSN catalog over the same time period, with the average interevent interval increasing gradually from about 25 to 50 s in the eleven days after the M 7.1 mainshock.
Minimum Amplitude Analysis
We examine continuous data from the vertical components of Plate Boundary Observatory stations B918 and B921 (see Fig. 1), sited within boreholes at depths of 196.7 and 139.9 m, respectively. Because these stations have lower non‐seismic noise levels than surface seismometers and are located only 10–50 km away from the bulk of the aftershock activity, they should be capable of recording very small earthquakes and are ideal for exploring the lower‐magnitude limits of aftershock detectability. Indeed, it is easy to see many uncatalogued events in these borehole records, even many days into the aftershock sequence (Fig. 3) and there are numerous low‐amplitude possible phase arrivals between the more obvious events.
Early in the aftershock sequence, overlapping P and S waves from multiple earthquakes create a continuous seismic “noise” level that is higher than the pre‐mainshock noise level caused by surface winds and other non‐seismic sources. Our study is motivated by the question of how long this lasts, that is, when can pre‐mainshock noise levels again be observed, and what does this imply for the rates of very small earthquakes (M < 0). Our starting point is measuring the minimum rms amplitude as a function of time, that is, determining the amplitude during the quietest periods.
Minimum rms amplitudes for station B918 are shown in Figure 4a up to 30 days after the second mainshock. Before the first mainshock, the minimum rms noise levels exhibit daily fluctuations indicative of surface noise sources (e.g., McNamara and Buland, 2004; Bonnefoy‐Claudet et al., 2006; Groos and Ritter, 2009; Boese et al., 2015; La Rocca and Chiappetta, 2022). The minimum rms levels spike sharply at the time of each mainshock and then decay rapidly. However, the daily fluctuations do not become visible again until about 10 days after the M 7.1 event and the minimum levels continue to decrease until about day 15. Interestingly, the noise level eventually flattens to a level that is substantially higher than the pre‐mainshock level. However, because the daily noise cycles are once again visible, this offset in level is most likely caused by a change in the instrument sensitivity related to strong shaking during the mainshocks. Interestingly, during the minimum rms intervals the original seismograms and their spectra do not appear noticeably different during the ∼15 days of elevated rms levels compared to the pre‐mainshock period or after the daily noise cycle resumes (Fig. 4c,e). If absolute amplitude information is discarded, it is hard to distinguish non‐seismic noise from seismic noise resulting from overlapping aftershocks, at least for 1–10 Hz filtered records.
Analogous minimum rms amplitudes for station B921 are shown in Figure 4b. In this case, the minimum amplitudes return to pre‐mainshock conditions after about 15 days, with the daily fluctuations again visible, but without the offset in amplitudes seen for B918. The B921 seismograms during the minimum rms periods do not show much variation, but the pre‐mainshock spectra have a spectral peak at about 3.5 Hz that is not seen in the post‐mainshock spectra (except near days 27–30).
These results indicate that minimum noise levels for these two borehole stations are continuously elevated for over 10 days after the M 7.1 earthquake, most likely due to overlapping arrivals from multiple tiny earthquakes. This result is not particularly surprising, given the high signal‐to‐noise for most of the catalog events visible in the borehole records (e.g., Fig. 3). The Shelly (2020) catalog has an average interevent interval of about 35 s at 7–10 days after the mainshock but few events below M 0.5. Assuming a b‐value of one and M > 0.5 for the catalog, the G–R relation predicts an average interevent interval of 3.5 s for M > −0.5 events and 0.35 s for M > −0.5 events.
Modeling the rms Amplitude Distribution for Station B921
The rms time series ( from equation 1) has information about the distribution of amplitudes as a function of time at 5‐s intervals. Figure 5 plots the number of rms amplitudes within bins of equal log amplitude for four different time periods (the 10 days before the M 6.4, the ∼34 hr between the M 6.4 and 7.1 mainshocks, the 10 days following the M 7.1 mainshock, and the 10–30 days period after the M 7.1 event). The counts are adjusted for the length of the time period, that is, by dividing by the number of days. Note that, as expected, there are many more small‐amplitude time periods than large‐amplitude time periods. The time between the M 6.4 and 7.1 earthquakes has the highest rate of the large rms amplitudes, followed by the 10 days after the M 7.1 mainshock. The pre‐M 6.4 time period has the lowest rates of high rms times. Assuming the log(amplitude) is proportional to the magnitude and an underlying seismicity b‐value of one, then the amplitudes should have a slope of −1 in this figure. However, this is observed only for the 10–30 days period after the M 7.1. The aftershock time periods closer to the mainshocks exhibit slopes of −1 only for the larger amplitude bins, with gradually shallowing slopes for the smaller amplitude bins. This behavior makes sense if we consider the effect of overlapping events, which will reduce the count of the smaller events because their low amplitudes are likely to be obscured by other events occurring at nearly the same time.
Modeling these results is complicated by the fact that the earthquakes occur at a wide range of distances from the station, so we do not expect a single magnitude versus log(amplitude) scaling relation to apply to the entire set of events. However, as discussed in Shearer et al. (2023), the sum of two distributions with different log(A) to magnitude scaling coefficients but identical b‐values will have a log N(log A) distribution with the same slope (b‐value) as log N(M).
To explore the relationship between these rms amplitude observations and earthquake magnitudes, we find the 5 s bin with the highest rms amplitude within 10 s before and 15 s after each event in the SCSN catalog. Figure 6 plots the resulting log rms amplitude versus magnitude for four different bins in epicentral distance from station B921 (0–12 km, 12–25 km, 25–50 km, and 50–100 km). Within each bin, a roughly unit slope is seen between log(A) and magnitude with lower amplitudes seen for the more distant events. Of course, some scatter should be expected even at a constant epicentral distance because any single station provides only a crude magnitude estimate, given differences in earthquake radiation pattern and directivity. At each of the four distance ranges, we compute the median scaling factor, C, of the differences between log(amplitude) and magnitude; the C‐values and numbers of events are shown in each figure panel.
Initialize to zero an array that will contain the 5 s rms amplitude bins for comparison to the real data.
Assume Reasenberg and Jones (1989) values of a, b, c, and p, using for the mainshock magnitude.
Assume a minimum and maximum aftershock magnitude, and , and magnitude sampling interval .
For every interval between and , compute the expected number of aftershocks by integrating equation (4) between t = 0 and days (864,000 s). Randomly assign a time to each aftershock based on the Omori law values of p and c.
For each synthetic earthquake magnitude, randomly pick one of the four epicentral distance bins, weighted by their proportions in the SCSN catalog (e.g., 7,040/15,571 = 45% of our earthquakes are assigned to the 25–50 km bin). Using the scaling factor, C, for the assigned bin, compute an rms amplitude.
From the synthetic earthquake time, find the appropriate 5 s amplitude bin and add the square of the rms amplitude. Note that we assume amplitudes will add incoherently, that is, energy is preserved but amplitude is not.
After the squared rms amplitude contributions has been added to the appropriate bins for all the synthetic earthquakes, take the square root of the summed values to convert back from energy to amplitude.
Process the synthetic rms 5‐s bins in the same way as the real data to generate a distribution curve for equal bins in log(amplitude) and a curve of the minimum rms bins at 30 min intervals.
Compare the results to the real data for Ridgecrest, and adjust the parameters until a reasonable fit is obtained.
Figure 7 shows the fit we obtain for a = −1.9, b = 0.9, p = 0.895, c = 20 s, , , and . For comparison, Hardebeck et al. (2019) examined 1933–2017 California M ≥ 5 mainshocks from 1933 to 2017 and found productivity a‐values mostly between −3 and −1.5. Joint inversion of southern California mainshocks from 1980 to 2017 obtained a = −2.04, p = 0.83, and c = 285 s. Modeling specifically of the Ridgecrest M 7.1 mainshock indicates an a‐value of −2.1 (J. Hardebeck, personal comm., 2024). Note that we do not attempt a formal inversion for our model parameters or to estimate uncertainties; our goal is only to see if a reasonable fit to the data is possible with this simple approximate model using parameter values within the range of prior studies. The novel aspect of our study is that we are able to extend modeling based on Reasenberg and Jones (1989) to very‐low‐magnitude earthquakes. Our model cannot reproduce the high minimum rms amplitudes observed 5–10 days after the M 7.1 mainshock if only M >−1 aftershocks are included. It is the overlapping aftershocks of −2 < M < −1 that create the continuous aftershock “hum” observed at these times. Between days 9 and 10, our model predicts that aftershocks are occurring at an average rate of 10 earthquakes per second, or an average interevent interval of about 0.1 s. Under these conditions, it is understandable that individual small earthquakes cannot be identified, but the existence of these tiny events is apparent in the continuously elevated rms amplitudes compared to pre‐mainshock conditions.
Here, we use the term aftershock “hum” to refer to seismic noise that is composed of the wavetrains of many overlapping tiny earthquakes, which occur at rates sufficiently high that individual events cannot be identified. This model is similar to hypotheses for the origin of tectonic tremors that involve swarms of low‐frequency earthquakes (e.g., Shelly et al., 2007; Beroza and Ide, 2011). However, aftershock hum is composed of “ordinary” earthquakes that are special only in that they occur at very high rates. Note that aftershock hum is not necessarily easily distinguished from non‐seismic noise sources, as shown by the example seismograms in Figure 4. One defining property of aftershock hum is that its amplitude is normally expected to decay following Omori’s law and to lack the daily fluctuations often seen in anthropogenic or wind‐generated noise. Although our observations cannot exclude the possibility of non‐aftershock contributions (e.g., tremor, instrument response changes) to the hum signal, the fact that the Reasenberg and Jones (1989) model predicts the hum amplitude so well suggests that such contributions are likely small if they exist at all.
MP Results
Measuring the average rms amplitude in 5‐s bins provides a way to detect the presence of seismic noise caused by overlapping aftershocks, simply by comparing the amplitudes at the quietest times to pre‐mainshock levels. However, potentially we can learn more by examining other features of the waveforms that are sensitive to aftershocks. Template matching has proven to be an effective method to enlarge earthquake catalogs by cross‐correlating waveforms of known events with continuous data to identify cross‐correlation peaks indicative of additional events that have similar locations and waveforms to the template events (e.g., Ross et al., 2019; Shelly, 2020) and in principle this approach can detect individual events near or even below the noise level. The template-matching method is limited to detecting events close to those already existing in the catalog and may miss events in different locations. In principle, we can obtain more complete results by considering every part of the continuous seismogram as a possible template, but this is computationally challenging for long time series. However, recent developments in software and high‐performance computing have led to the MP—a computationally efficient algorithm to obtain cross-correlation results for every pair of windows in a long time series (e.g., Yeh et al., 2016; Zhu et al., 2016; Zimmerman et al., 2019), that is, to cross‐correlate everything with everything (Shabikay Senobari et al., 2024).
MP effectively performs template matching without predefined templates. If we consider a time series as composed of a series of sub‐windows, the correlation coefficients of every pair of windows form what is termed the “Similarity Matrix.” The sub‐window length for local earthquake detection might be 5–10 s and the similarity matrix computed at full resolution would consider a set of overlapping sub‐windows along the time series with sequential pairs shifted by only one data point. In this case, the similarity matrix has about elements, in which n is the number of data points, and becomes impractical to compute or store for long time series. MP was developed in the computer science community to efficiently extract the maximum of each column in the similarity matrix and the location of the most similar sub‐window. In seismology, it can be used to find the sub‐window most highly correlated to every possible target sub‐window in a seismogram. In this way, pairs and clusters of similar events can be identified even if none of the events are contained in existing seismic catalogs (Shabikay Senobari et al., 2024).
We applied the MP algorithm to 100 days of continuous data from the vertical components of borehole stations B918 and B921, extending from 10 days before to 90 days after the M 6.4 event. As in the amplitude analysis discussed earlier, we applied a 1–10 Hz band‐pass filter and downsampled to 20 samples per second. We used MP to find the maximum correlation between every 5 s interval in the time series with the rest of the time series, shifting the 5 s window to just one point between each search (see Shabikay Senobari et al., 2024, for more details). The MP output has considerable information about tiny pairs of events that cross‐correlate, even if neither event is contained in an existing catalog. However, our focus here is to identify parts of the time series that do not correlate well with any other part of the time series, that is, that are least likely to include hidden earthquakes. Accordingly, we divide the MP output into 5‐s intervals, compute the average MP correlation over each 5‐s window, and search for the window with the smallest average MP correlation value within 30‐min intervals.
These results are shown in Figure 8. For station B918, the pre‐mainshock minimum MP correlations are between 0.53 and 0.54. The minimum correlations increase following both the M 6.4 and 7.1 mainshocks and then decay for the next ten days. However, even 80 days after the M 7.1 earthquake, the minimum correlations are between about 0.54 and 0.56, that is, distinctly higher than they were before the M 6.4 event. As in the case of the minimum amplitude analysis, the seismograms at the time of the correlation minima do not appear unusual and have no obvious differences between the time periods before and after the mainshocks. Yet the MP analysis shows that the pre‐ versus post‐records are indeed different, that is, there are no periods after the mainshocks when the traces are as uncorrelated as they were before the mainshocks. This indicates that continuously overlapping aftershocks are likely present for at least 80 days after the mainshocks, even though they are too small to be visible directly and their integrated energy does not clearly exceed pre‐mainshock conditions (see Fig. 4).
The MP results for station B921 are complicated by the monochromatic noise signal near 3.5 Hz that is seen in the low‐amplitude traces before the M 6.4 mainshock. This elevates the minimum MP correlation values during the pre‐mainshock period. The average minimum correlations drop following the mainshocks and then decay further during the next 20 days. The minimum correlations are noticeably higher from 26 to 30 days following the M 7.1 event, a period during which the 3.5‐Hz noise signal reappears in the spectra during the lowest amplitude periods (see Fig. 4f). This supports the connection between this non‐seismic noise signal and periods of higher minimum average MP values. Despite the greater scatter in the minimum MP correlations for B921, it does not appear that the results show a complete return to pre‐mainshock conditions even after 80 days. However, because of the 3.5‐Hz noise, the B921 MP results are less easily interpreted than the B918 results, which provides the best evidence that overlapping aftershocks are still present at 80 days.
Discussion
As improved detection methods have lowered the magnitude limit for local earthquake catalogs (e.g., Ross et al., 2019, 2020; Mousavi et al., 2020; Zhou et al., 2022; Wilding et al., 2023; Si et al., 2024), no evidence has emerged that we are reaching any lower magnitude limit in the Gutenberg–Richter power law, which predicts roughly a tenfold increase in earthquake numbers for every unit decrease in magnitude. Thus, our result showing that ten days after the M 7.1 Ridgecrest mainshock, M ≥−2 earthquakes are occurring at an average rate of about 10 events per second is not surprising, as it follows directly from the Reasenberg and Jones (1989) aftershock model, given reasonable parameter choices in equation (3). What is more surprising is that the presence of a very high rate of M <−1 aftershocks after ten days is detectable in the data for the Ridgecrest sequence. We have focused on the two borehole sensors, B918 and B921, because of their low background aseismic noise levels and their proximity to the Ridgecrest mainshocks; it would be interesting to test Ridgecrest surface stations and other aftershock sequences to see if similar results can be obtained.
So far we have only applied relatively crude modeling based on Reasenberg and Jones (1989) parameters, which nonetheless is sufficient to explain the main features that we observe in the rms amplitude distribution within 5 s windows as a function of time. More sophisticated modeling based on generating synthetic time series using actually observed event waveforms, summed with randomly generated times and magnitudes, might provide insights regarding aftershock behavior in the earliest minutes and hours after the mainshock when only the largest events are included in earthquake catalogs. Better characterization of small earthquake statistics during this early time period could help improve real‐time aftershock forecasts (e.g., Omi et al., 2013, 2019). The distribution of rms amplitudes shown in Figure 5 and our modeling results suggest that it should be possible to estimate G–R b‐values, aftershock p‐values, and other seismicity parameters directly from continuous waveforms without the need to detect and characterize individual earthquakes, which might have some advantages for real‐time aftershock forecasts.
As more small earthquakes are detected and located, fault structures become better defined (e.g., Ross et al., 2020; Wilding et al., 2023; Yoon and Shelly, 2024) and the statistics of earthquake swarms and foreshock sequences become clearer (e.g., Trugman and Ross, 2019; Ross et al., 2020; Peng et al., 2025). There are two challenges to detecting and locating individual small earthquakes. The first is that the P and S arrivals may be near or below the noise levels of the recording stations, a difficulty that can be addressed using power spectral density analysis (Vaezi and Van der Baan, 2015), multistation cross‐correlation, or beamforming of seismic array data (e.g., Meng and Fan, 2021; Shearer et al., 2023). The second difficulty is that small events can occur so frequently that their phase arrivals begin to overlap and eventually merge to form the seismic “hum” that we observe at Ridgecrest. This has long been obvious in the very early stages of aftershock sequences but seismicity rates remain elevated for many months following large earthquakes (e.g., Utsu et al., 1995), implying that the seismic hum only gradually decreases in amplitude as smaller and smaller events are required to provide the overlapping arrivals. However, continuous seismic hum should also occur as part of background seismicity not associated with a specific aftershock sequence, assuming no break in the G–R law. For example, the QTM catalog of Ross et al. (2019) contains 1.8 million earthquakes in southern California from 2008 to 2017 (averaging 20 events per hour) and is complete to about M 0.3. Assuming a b‐value of unity, this implies an average rate exceeding one event per second for M >−2 earthquakes and 10 events per second for M >−3 earthquakes.
Even if detecting and locating individual events is impractical at these very low magnitudes, careful analyses of the lowest observable noise levels and application of waveform cross‐correlation to continuous data (using MP or other methods) might allow seismic hum amplitudes to be measured as a function of time. In this way, it is possible that swarms and foreshock activity could be detected and characterized even without knowledge of the details of individual earthquakes.
Conclusions
Observations from two borehole seismometers at 140 and 197 m depth located close to the 2019 M 6.4 and 7.1 Ridgecrest, California, mainshocks are used to explore the lower magnitude limits of aftershock detection. We find that vertical‐component rms amplitudes were elevated compared to pre‐mainshock noise levels for at least 10 days following the M 7.1 mainshock even during the quietest periods when no clear seismic arrivals are visible. Our rms amplitude results are well explained using the Reasenberg and Jones (1989) aftershock model, which suggests the presence of continuous seismic noise (here termed aftershock hum), generated by the overlapping arrivals from many tiny earthquakes. Our modeling predicts that 10 days after the M 7.1 mainshock, M >−2 aftershocks occurred at an average rate of about 10 events per second. The presence of tiny aftershocks embedded within the continuous borehole records is further supported by MP calculations that examine how well correlated different 5‐s segments of the seismograms are with each other (a form of template matching without predefined template events). The MP results show that the borehole seismograms during the first 80 days of the aftershock sequence have different correlation properties than existed before the M 6.4 mainshock, even during the least correlated periods. Both our rms amplitude and MP analyses indicate that the signature of very small (−2 < M < −1) earthquakes can be detected even when they are too weak to be visible directly, which has implications for future studies of swarms and foreshock sequences.
Data and Resources
The waveform data used is this study are from the Plate Boundary Observatory Borehole Seismic Network (network code PB) and are available from the EarthScope Data Center (Incorporated Research Institutions for Seismology Data Management Center [IRIS‐DMC]), which is funded through the National Science Foundation’s Seismological Facility for the Advancement of Geoscience (SAGE) Award under Cooperative Agreement EAR‐1724509. The earthquake catalogs used in this study are available from the Southern California Earthquake Data Center (SCEDC, 2013). The SCEDC and Southern California Seismic Network (SCSN) are funded through U.S. Geological Survey Grant G20AP00037 and SCEC. Matrix profiles are generated by SCAMP software (https://github.com/zpzim/SCAMP, last accessed April 2025) (Zimmerman et al. 2019).
Declaration of Competing Interests
The authors acknowledge that there are no conflicts of interest recorded.
Acknowledgments
This research was supported by the Southern California Earthquake Center (SCEC) (Contribution Number 14157). SCEC is funded by National Science Foundation (NSF) Cooperative Agreement EAR‐1600087 and U.S. Geological Survey (USGS) Cooperative Agreement G17AC00047. P. M. S. acknowledges support from NSF (FAIN‐2104240). N. S. S. acknowledges support from NSF (FAIN‐2103976).