Refined Earthquake Focal Mechanism Catalog for Southern California Derived With Deep Learning Algorithms

Earthquake focal mechanisms, determined with P‐wave polarities and S/P amplitude ratios, are primary data for analyzing fault zone geometry, sense of slip, and the crustal stress field. Solving for the focal mechanisms of small earthquakes is often challenging because phase arrivals and first‐motion polarities are hard to be separated from noise. To overcome this challenge, we implement convolutional‐neural‐network algorithms (Ross, Meier, & Hauksson, 2018, Ross, Meier, Hauksson, & Heaton, 2018, https://doi.org/10.1029/2017jb015251, https://doi.org/10.1785/0120180080) to detect additional phases and polarities. Using both existing and these new data, we build a high‐quality focal mechanism catalog of 297,478 events that occurred from 1981 to 2021 in southern California with the HASH method of Hardebeck and Shearer (2002), https://doi.org/10.1785/0120010200, Hardebeck and Shearer (2003), https://doi.org/10.1785/0120020236. The new focal mechanism catalog is overall consistent with the standard catalog (Yang et al., 2012, https://doi.org/10.1785/0120110311) but includes 40% more focal mechanisms, and is more consistent with moment tensor solutions derived using waveform‐fitting methods. We apply the new catalog to identify changes in focal mechanism properties caused by the occurrences of large mainshocks such as the 2010 Mw7.2 El Mayor‐Cucapah and 2019 Mw7.1 Ridgecrest earthquakes. Such changes may be associated with co‐seismic stress drops, post‐seismic deformation processes, and static stress changes on a regional scale. The new high‐resolution catalog will contribute to improved understanding of the crustal stress field, earthquake triggering mechanisms, fault zone geometry, and sense of slip on the faults in southern California.

focal mechanisms of relatively small earthquakes (usually M L < 3) are derived by fitting the spatial distributions of amplitudes and first motion polarities of the radiated waves (e.g., Hardebeck & Shearer, 2002, 2003Jost & Herrmann, 1989). Focal mechanisms of population of earthquakes can be used to characterize the stress field, dominant faulting style, fault zone structure, and the variations of these properties in the spatial volumes and time intervals containing the events (e.g., Bailey et al., 2010;Hardebeck & Michael, 2006;Vavryčuk et al., 2015). The spatiotemporal resolution of these results can contribute significantly to a better understanding of earthquake and fault physics in the crust and depends on the smallest earthquake size and density of events for which reliable focal mechanisms are available.
In southern California, a set of studies progressively derived focal mechanisms for smaller and larger populations of events. To calculate focal mechanisms of small earthquakes, Shearer (2002, 2003) calculate focal mechanisms with the HASH algorithm, which approximates the events as double-couple point sources and determines the mechanisms by fitting the observed P-wave first motion polarities and the S/P amplitude ratios. Yang et al. (2012, updated to later years, referred to below as the YHS catalog) applied the HASH algorithm to all events in the earthquake catalog of Hauksson et al. (2012, updated to later years) (Figure 1). At the end of 2021, the YHS catalog had about 211,000 mechanisms that have been widely used to understand failure processes of earthquake sequences Meier et al., 2014), earthquake triggering mechanisms (Meng & Peng, 2014;Ross, Rollins, et al., 2017), detailed fault zone structures (Cheng et al., 2018;, and spatiotemporal variations of stress fields (Abolfathian et al., 2020;Yang & Hauksson, 2013). However, over 70% of the events in the Hauksson et al. (2012) catalog still do not have focal mechanism solutions, and 60% of the focal mechanisms in the YHS catalog have large uncertainties (>45°). Most of the events without focal mechanisms or with low-quality focal mechanisms are small-magnitude earthquakes, as it is difficult to pick the input first-motion polarities and P-/S-wave amplitude ratios from weak signals.
To overcome these challenges, Ross, Meier, and Hauksson (2018), Ross, Meier, Hauksson, and Heaton (2018) trained convolutional-neural-networks (CNN) using millions of manually labeled seismograms recorded in Southern California. The trained CNN pickers can successfully pick the catalog events' polarities and phases with higher precision compared with traditional methods and can potentially detect extra polarities and phases. Uchide (2020) trained a CNN polarity picker using seismic data in Japan and implemented it to obtain earthquake focal mechanism catalog and stress map of Japan (Uchide, 2020;Uchide et al., 2022). These results suggest that incorporating CNN polarity and phase pickers into focal mechanism calculations has the potential of providing mechanisms for smaller events and leveraging the accuracy and robustness of data in southern California. Here we first apply the CNN pickers of Ross, Meier, and Hauksson (2018), Ross, Meier, Hauksson, and Heaton (2018) to detect extra polarities and arrivals from earthquake waveforms. These additional inputs are then combined with catalog polarities and arrivals for calculating focal mechanisms with the HASH algorithm. In the following, we first introduce the data used in the study (Section 2.1) and the improved focal mechanism calculation workflow (Section 2.2). The new workflow is applied to all events from 1981 to 2021 in the earthquake catalog of Hauksson et al. (2012) to generate a new focal mechanism catalog (Section 3). Our new catalog, which we designate CNN_SoCal, has 40% more focal mechanisms than the YHS catalog (Section 3.2) with enhanced accuracy (Section 3.3). The improved catalog provides more comprehensive observations of earthquake fault geometry and sense of slip of the analyzed earthquakes (Section 3.4), giving improved constraints on the physics of earthquakes and faults. The results are discussed in Section 4.

Input Data
We use earthquake waveforms for all 716,513 events listed in the regional Southern California Seismic Network (SCSN) catalog for the period 1981(Southern California Earthquake Data Center, 2013. These events are shown in Figure 1 and the used stations are shown in Figure S1 in Supporting Information S1. The analysis employs only the earthquake waveforms recorded by broadband (HH, BH) and short-period (EH) channels. We also use S-/P-wave amplitudes as well as phase picks and polarities that were manually reviewed by the seismic analysts at the SCSN. All waveforms were resampled at 100 Hz. The different stages of the analysis benefit from each having a different band-pass filter; we use a 3-20 Hz filter for phase picking and polarity recognition for high-resolution picks of impulsive arrivals, and a 1-10 Hz frequency range for estimating S-/P-wave amplitude ratios for robust estimation of amplitude ratios in dominant frequency ranges of earthquakes.

Method
We develop a workflow for building a high-resolution focal mechanism catalog and apply it to the entire earthquake waveform archive of the SCSN. This workflow consists of the following main steps (Figure 2a): (Section 2.2.1) applying a deep-learning-based phase picking algorithm to earthquake waveforms recorded by all broadband and short-period channels to increase the number of available picks; (Section 2.2.2) using these phase picks to calculate S-/P-wave amplitude ratios ( Figure 2b); (Section 2.2.3) applying a deep-learning-based first-motion polarity picking algorithm to M < 2.0 events to increase the available polarities (Figure 2c); and (Section 2.2.4) using the first-motion polarity and S-/P-wave amplitude ratio data to calculate focal mechanisms (Figure 2d). The details of each step are discussed in the following subsections.

Phase Arrival Identification
While a substantial quantity of the manually reviewed phase picks is available for the entire earthquake catalog, not all stations and earthquakes have phase picks. To augment and improve upon the existing information, we apply the generalized phase detection (GPD) algorithm (Ross, Meier, Hauksson, & Heaton, 2018) to the earthquake waveform archive whenever picks are unavailable.
As the GPD algorithm was designed for use with 3-component data only, we focus here on these sensor types. For each station-event pair, we attempt to detect and pick P-and S-waves using the GPD algorithm. We operate on 4.0 s sliding windows of three-component waveforms with 0.1 s time step and analyze each time window with a previously trained deep neural network (Ross, Meier, Hauksson, & Heaton, 2018). This outputs the probability of the dominant signal class being a P-wave, S-wave or noise and assigns them to the time point at the center of the window. We designate the time points with probability values greater than 0.95 as the onset times of phase arrivals. When there are multiple consecutive values above 0.95, the time point with the maximum value is chosen as the detection. Then we compare the obtained CNN phase arrivals with the corresponding theoretical arrival times estimated using the event's origin time and hypocenter along with the 1D velocity model in Hutton et al. (2010). We discard any picks that are more than 1s from the theoretical arrivals, which is appropriate for these local distances. If multiple picks are within this window, the closest one is taken. Applying the GPD algorithm to all events' waveforms yields 4,036,357 P-wave arrivals and 3,092,738 S-wave arrivals. More than 90% of the P-and S-wave arrivals are within 0.2 and 0.3 s from their theoretical arrivals, respectively ( Figure S2 in Supporting Information S1).

S-/P-wave Amplitude Ratio Calculation
We calculate P-and S-wave amplitudes from pre-processed waveforms recorded at three-component stations. Pand S-wave arrivals in the SCSN catalog or those picked by the GPD algorithm are used to measure amplitudes. We choose 0.5 s before to 1.5 s after P-and S-wave arrivals as the P-and S-wave signal windows, respectively, and 2.5-0.5 s before P-wave arrivals as the noise windows. We obtain the vector summation over three-component waveforms and take the difference between the maximum and minimum amplitude values in each time window to be the estimation of the signal or noise amplitudes. If the time difference between P-and S-wave arrivals is larger than 2 s and the SNR of P-wave amplitude is larger than 3, we calculate the S-/P-wave amplitude ratio and incorporate it into the focal mechanism calculation.

Polarity Recognition
The other ingredient for calculating focal mechanisms is the polarity of P-wave first-motion. To augment the manually reviewed polarities already available from the SCSN, we use a deep learning algorithm (CNN polarity picker) for picking first motion polarities (Ross, Meier, & Hauksson, 2018). For each P-wave arrival time available, we select a 4 s waveform segment of the vertical component waveform centered at the phase arrival time as the input to determine the P-wave first-motion polarity using the CNN polarity picker. For each event, we estimate the CNN polarity picker precision as the number of true positives (the number of CNN polarities that are the same as catalog polarities) divided by the total number of CNN polarities with corresponding catalog records. We apply the polarity picker to all catalog events and estimate event-based polarity picker accuracy for all events. The fraction of events with polarity picker precision >80% decreases with increasing magnitude from over 90% for M L < 1 events to less than 70% for M L > 1.9 events ( Figure S3 in Supporting Information S1). The reason for this is that most events in the training data set for the CNN polarity picker are small. To control the input polarity quality, we only apply the CNN polarity picker to M L < 2 events.

Focal Mechanism Calculation
The obtained S-/P-wave amplitude ratios, catalog polarities and the additional CNN polarities are used to calculate focal mechanisms with the HASH program ( Figure 2a). We use the same input parameters as those in Yang et al. (2012). The minimum number of required P wave polarities (npolmin) is 8 and only stations within 120 km of the epicenter (delmax) are used. The maximum allowed azimuthal gap (max_agap) and takeoff angle gap (max_pgap) of the stations are 90° and 60°, respectively. Each focal mechanism is calculated for 30 trials (nmc) using a 5° grid angle for searching focal mechanism solutions. We set the fraction of the bad polarities (badfrac) to be 5% and the acceptable variation of the S/P amplitude ratios (qbadfrac) to be 0.3 on a base 10 logarithmic scale. For the obtained focal mechanisms, we use a 45° angle for computing the mechanism probability (cangle) and set the probability threshold for multiples (prob_max) as 0.1. The take-off angles used in the calculation are computed from nine 1D velocity models ( Figure S1 in Yang et al., 2012) for Southern California to address velocity model uncertainties in the focal mechanism uncertainty estimation.

Results
For all 716,513 events from 1981 to 2021 in Hauksson et al. (2012), we obtain 297,478 focal mechanisms along with their associated uncertainties and other parameters. We analyze this catalog to determine the parameters for quality classification (Section 3.1, Figure S4, Table S1 in Supporting Information S1) and use the quality definition to classify the focal mechanisms both in the CNN_SoCal catalog and the YHS catalog. The quality A-D focal mechanisms are used for comparisons. Here, we present and discuss results of quantity improvements (Section 3.2), quality improvements (Section 3.3), and new observations using the CNN_SoCal catalog (Section 3.4) compared with the YHS catalog.

Quality Classification
To derive a refined catalog with high quality focal mechanisms that can provide refined information on fault structures, stress inversions and other topics, we use the following parameters to characterize focal mechanism quality: nodal plane uncertainties (var_avg; root-mean-square angular difference of the acceptable nodal planes from the preferred planes), mechanism probability (prob; fraction of acceptable mechanisms within 45° from the preferred solutions), fraction of polarity misfit (mfrac), and station distribution ratio (STDR; distribution of observations relative to the focal sphere). Since most M L > 3 events are well recorded and have large SNR at most stations within 120 km epicentral distances, we use the distribution of their quality parameters in the new catalog to define A-D focal mechanism qualities ( Figure S4 in Supporting Information S1). We define focal mechanisms with var_avg <25°, prob >0.8, mfrac ≤0.2, and STDR >0.4 as quality A. For quality B to C, we gradually relax the requirements. Quality D focal mechanisms are those that do not satisfy the requirements of quality A to C but have maximum azimuthal gaps <90° and maximum takeoff angle gaps <60° (Table S1 in Supporting Information S1). We use the quality definitions to classify the focal mechanisms both in our new catalog and in Yang et al. (2012). The quality A-D focal mechanisms are used in the following analysis. All the parameters used for quality classification are also provided in the catalog (Cheng et al., 2023) and users can customize the quality definition based on their needs.

Quantity Improvements
Since different times and locations have different station settings, we first check the temporal dependency of the available focal mechanisms. For results from 1981 to 2007, the CNN_SoCal catalog has about 20% more focal mechanisms than the YHS catalog due to the limited number of three-component stations ( Figure 3a). After 2007, our new catalog has at least 40% more solutions for most years (Figure 3b). For all catalog events from 1981 to 2021, more than 40% of the events have at least 1 additional S-/P-wave amplitude ratios and 1 or more additional first-motion polarities (Figure 4b), leading to 40% more focal mechanisms than the YHS catalog. Near the edge of the network (like the Northridge area, Ridgecrest area, and the Brawley Seismic Zone), the numbers of the obtained solutions are more than doubled (Figure 4a). There are 15% more M L ≥ 2.0 and 50% more M L < 2.0 focal mechanisms compared with the YHS catalog ( Figure 4c). The numbers of quality A, B, C, and D focal mechanisms increase by 36%, 52%, 40%, and 22%, respectively (Figure 4d).

Quality Improvements
By including the CNN pickers in the workflow for determining focal mechanisms, the CNN_SoCal catalog has about 40% more focal mechanisms in all quality categories. However, the focal mechanism quality (defined in Section 3.1) only represents the solution uncertainty instead of the accuracy. As mentioned in Hardebeck and Shearer (2002), about 80%-90% of the manually picked polarities are correct. If we have many measured polarities for a certain event but over 20% of them are incorrect, the determined focal mechanism may still have low uncertainty and high quality, but these solutions may be highly biased due to the errors in the input data. To 10.1029/2022JB025975 6 of 18 overcome this problem, we use multiple independent observations to evaluate the accuracy of the derived focal mechanisms.

Comparison With the YHS Catalog
To check the consistency with the YHS catalog, we extract common events and compare the consistency of their focal mechanisms in both catalogs. We measure the similarity of two focal mechanisms using the Kagan angle  defined as the minimum rotation angle needed to make two focal mechanisms identical (Kagan, 1991), and the matrix angle defined as the dot product angle of moment tensor matrices of the two focal mechanisms (Tape and Tape, 2012). For all 200,278 common quality A-D focal mechanisms, the 80 percentiles of both the Kagan and matrix angles are less than 45° (Figures 5a and 5b), which is the uncertainty value used to define the quality C focal mechanisms. Over 80% of 56,681 common quality A and B focal mechanisms have both the Kagan and matrix angles less than 25° (Figures 5c and 5d), which is the uncertainty value used to define the quality A focal mechanisms. Our results are thus consistent with the YHS catalog with the relative differences generally within the uncertainty range.

Comparison With Moment Tensor Catalogs
To evaluate the focal mechanism catalog further, we compare the focal mechanisms with the corresponding moment tensors obtained by waveform-fitting methods. If the results obtained from these two different methods are consistent, we assume they are accurate. We compared both the CNN_SoCal catalog and the YHS catalog to see which one is more consistent with the moment tensors. We use three different moment tensor catalogs for comparison: 610 M w > 3.5 moment tensors in SCSN catalog (Figures 6a and 6b), 211 M w > 3.5 events in Tape et al. (2010) (Figures 6c and 6d), and 145 M w > 3.0 events in Wang and Zhan (2020) (Figures 6e and 6f). Our results are more consistent with these three moment tensor catalogs with overall smaller Kagan and matrix angle differences ( Figure 6). However, some events have highly different solutions in the CNN_SoCal catalog compared with those in the SCSN moment tensor catalog (Figures 6a and 6b). Figure S5 in Supporting Information S1 shows the spatial distribution of the Kagan angles and matrix angles between SCSN moment tensor solutions and focal mechanisms in the CNN_SoCal catalog. Most of the events with large Kagan and matrix angles are located at the edge of the station networks ( Figure S5 in Supporting Information S1), which suggests the inaccuracies might be due to incomplete azimuthal coverage. Note that focal mechanism solutions are solved using high-frequency waveforms and reflect the beginning of the rupture processes, while moment tensor solutions are solved using relatively low frequency waveforms and reflect the whole rupture processes. Therefore, the inconsistency between focal mechanisms and moment tensors are not necessarily due to inaccurate solutions.

Comparison of Focal Mechanisms of Similar Earthquakes
For M L < 3.0 earthquakes, it is challenging to obtain highly accurate focal mechanism for each individual earthquake and quantify the accuracy of the obtained solutions. To overcome this limitation, we next use relative focal mechanism similarity of similar earthquakes to evaluate their accuracy. We first obtain a set of similar earthquakes in the Cahuilla Valley Pluton area Ross et al., 2020) by extracting closely located events with highly similar waveforms ( Figure S6, Text S1, and Table S2 in Supporting Information S1). We compare their focal mechanisms in the YHS catalog with those in the CNN_SoCal catalog and assume that the mechanisms of these similar events are more accurate if they are more consistent.
To compare the obtained focal mechanisms in a controlled way, we first separate them into three categories, quality A & B (Figures 7a and 7b) and quality C & D (Figures 7c and 7d) mechanisms with solutions in the YHS catalog, as well as additional solutions that are not in the YHS catalog (Figure 7e). For each group, we plot the nodal plane uncertainties and the principal axes of focal mechanisms in the same panel, including P-axis (the middle of the dilatational quadrant), T-axis (the middle of the compressional quadrant), and B-axis (the intersections of nodal planes). The CNN_SoCal catalog has comparable number of quality A and B focal mechanisms (Figures 7a and 7b) with similar principal axes and similar uncertainties to the YHS catalog. For quality C and D mechanisms, the solutions in the CNN_SoCal catalog have more consistent principal axes and smaller nodal plane uncertainties compared with those in YHS catalog (Figures 7c and 7d). Moreover, the CNN_SoCal catalog has 10 additional focal mechanism solutions (Figure 7e).

New Observations From the CNN_SoCal Catalog
The CNN_SoCal catalog has higher focal mechanism density, delineates more detailed spatiotemporal variations of focal mechanisms and brings new insights into various loading processes. Therefore, we utilize quality A-C Figure 6. The comparisons of common focal mechanisms from the CNN_SoCal catalog (green) and the YHS catalog (purple) with multiple moment tensor catalogs that are determined by waveform-fitting. Histograms of (a, c, e) Kagan angles and (b, d, f) matrix angles among the common focal mechanisms between the focal mechanism catalogs and the moment tensor catalog from (a), (b) Clinton et al. (2006), (c), (d) Tape et al. (2010), and (e), (f) Wang and Zhan, (2020). Vertical lines denote the 80th percentile Kagan angles and matrix angles. focal mechanisms to investigate the spatiotemporal variations of focal mechanism properties. To better understand the variations of a group of focal mechanisms, we calculate the composite potency tensor by summing up all source mechanism tensors: where ̂ is the normalized potency tensor with √̂̂= 1 (Bailey et al., 2010). The maximum, intermediate and minimum eigenvectors of the composite potency tensor are assumed to represent maximum pressure axis (P-axis), maximum tensile axis (T-axis) and null axis (N-axis), respectively (Dziewonski & Woodhouse, 1983). The principal axes are used to calculate the strike, dip, and rake angle of the best-double-couple mechanism. To compare diverse focal mechanisms in a uniform way, we further express the mechanisms' faulting styles on a continuous scale from −1 to 1, with normal faulting having a value of −1, strike-slip denoted by 0 and thrust faulting denoted by 1 (Shearer et al., 2006). The above-mentioned P-axis orientation and faulting style of composite potency tensors can be used to investigate detailed changes of focal mechanism properties and related driving factors, like stress perturbations caused by major ruptures. In this section, we investigate the pre-and post-seismic focal mechanism variations of the most recent two large earthquakes recorded in the catalog, the 2010 M w 7.2 El Mayor-Cucapah earthquake (Figures 8 and 10) and the 2019 M w 7.1 Ridgecrest earthquake (Figures 9 and 11). The implications of the results are discussed in Section 4.2.

Temporal Variation of P-Axis Azimuth Before and After Major Ruptures
To understand the style of deformation, we plot the earthquake P axes of focal mechanisms near the 2010 M w 7.2 El Mayor Cucapah earthquake (Figure 8a) and the 2019 M w 7.1 Ridgecrest earthquake (Figure 9a). The P-axes lie roughly in the NW-SE direction for the 2010 M w 7.2 El Mayor Cucapah earthquake and N-S direction for the 2019 M w 7.1 Ridgecrest earthquake. To identify potential spatial and temporal variations of stress conditions, we divide the study areas into spatial bins from the NW to the SE with 10 km distance interval and analyze the temporal variations of the focal mechanisms. In each spatial bin, we calculate the P-axis azimuth and plunge of the composite potency tensor of 100 focal mechanisms with 99% overlapping time window. For each time window, we apply a bootstrap method to estimate the uncertainties of P-axis azimuth and plunge based on 100 resampled data sets. Since P-axis plunges are near-vertical in most spatiotemporal bins ( Figures S8 and S9 in Supporting Information S1), we focus on the variation of P-axis azimuth in the following analysis.
After the 2010 M w 7.2 El Mayor-Cucapah earthquake, the composite P-axis shows clockwise rotation after the mainshock in the NW and SE part of the AA′ cross-section (Figures 8b, 8c, 8g, and 8h) and counterclockwise rotation in the central part of AA′ (Figures 8d, 8e, and 8f). After several months, the P-axis azimuth becomes stable and does not show significant changes. Because there are more focal mechanism solutions in the CNN_ SoCal catalog (green curves), the corresponding results show a more rapid P-axis rotation after the major rupture compared with those solved using the YHS catalog (blue curves). Similarly, the 2019 M7.1 Ridgecrest sequence shows a clockwise rotation in the NW end of the aftershock zone (Figure 9b), a counterclockwise rotation near the mainshock hypocenter and major slip area (Figures 9c, 9d, 9f, and 9g), and no significant rotation in the intersection of conjugate fault area (Figure 9e). The results using the YHS and CNN_SoCal catalogs display similar patterns. However, because of more sampling points in time, the CNN_SoCal catalog illuminates a more gradual post-seismic clockwise P-axis rotation near the Garlock fault (Figure 9g), which are not captured by the YHS catalog.

Spatial Correlation Between Faulting Style Changes and Shear Stress Changes After Major Ruptures
We estimate and compare the pre-and post-seismic stress state by calculating the faulting style of the composite potency tensor of all quality A-C focal mechanisms on a 0.1 • × 0.1 • grid in southern California. Figures 10  and 11 present the distribution of faulting type within 3 years before (Figures 10a, 10d, 11a, and 11d), 1 year after (Figures 10b, 10e, 11b, and 11e) the 2010 M w 7.2 El Mayor-Cucapah earthquake and the 2019 M w 7.1 Ridgecrest earthquake, as well as the differences of faulting style between these two time periods. Compared with the faulting style variation illuminated by the YHS catalog (Figures 10c and 11c), the results from the CNN_SoCal catalog show more consistent patterns with the shear stress changes estimated from the permanent displacement of the major ruptures (Figures 10f and 11f). For the 2010 M w 7.2 El Mayor Cucapah earthquake, the results calculated from CNN_SoCal catalog (Figure 10f) show more transpressional focal mechanisms near the mainshock epicenter, along the Elsinore fault and San Jacinto Fault (SJF), and more transtensional focal mechanisms around the NW of the mainshock epicenter. For the 2019 Ridgecrest earthquake (Figures 11c and 11f), the faulting style changes solved using the CNN_SoCal catalog illuminates a broader area with more transtensional focal mechanisms around the major rupture and more transpressional to the NW and SW of the major rupture. The CNN_SoCal catalog shows a stronger spatial correlation between the change of faulting style with the shear stress changes estimated from the permanent displacement of the major ruptures in a broader area (Meng & Peng, 2014;.

Improvements of the CNN_SoCal Focal Mechanism Catalog
We apply the CNN phase picker and polarity picker for focal mechanism calculations and develop a new procedure to determine focal mechanisms for small-to-moderate-magnitude earthquakes with the HASH algorithm. We apply our approach to all events in the southern California catalog of Hauksson et al. (2012) and built a new focal mechanism catalog (CNN_SoCal catalog). The additional inputs obtained using the CNN pickers result in overall 42% more focal mechanisms compared with YHS catalog. In addition to additional focal mechanisms and reduced uncertainties, The CNN_SoCal catalog also has a higher consistency with multiple moment tensor catalogs and similar earthquakes. The improved accuracy of the focal mechanism solutions indicates that CNN pickers perform very well in accurately detecting additional phases and polarities, which help to improve the quality of the obtained focal mechanisms. Although the CNN_SoCal catalog exhibits high consistency with moment tensor catalogs, there are still some M w > 3.5 events with significantly different SCSN moment tensor solutions and focal mechanisms. Most of these events are located near the edge of the network and lack good azimuthal coverage ( Figure S5 in Supporting Information S1). This implies that further improvements of focal mechanisms by the advanced calculation workflow requires better azimuthal coverage. Take-off angle coverage is also very important for the quality of focal mechanisms. The take-off angle coverage is strongly related to the station spacing, station aperture, and the event depth ( Figure S7 in Supporting Information S1). For events shallower than 6 km, take-off angle varies greatly within 0-0.2° (about 22 km epicentral distance) but approaches zero when the epicentral distance is larger. With increasing event depth, take-off angle changes more gradually with the epicentral distance. To achieve a good take-off angle coverage, shallow events require small station spacing with small epicentral distances while deep events require large aperture of the station distribution. Besides station and events locations, small magnitude events radiate smaller amplitude signals, which increase the difficulty in picking accurate phase arrivals and further limits the azimuthal and take-off angle coverage of the data. Therefore, densely spaced stations that provide adequate azimuthal coverage with high SNR waveforms, are very important for obtaining high-quality focal mechanisms.
The derived CNN_SoCal catalog has around 30%-50% additional focal mechanisms in the areas with densely distributed stations (e.g., SJF zone, Landers and Hector Mine area) and over 80% additional focal mechanisms near the edge of the network (Figure 4). The significant improvements near the edge of the network suggests the importance of additional observations to the events with limited original inputs. Increasing input observations can be achieved by adding more stations and picking more arrivals in the areas with limited stations. In the areas with dense station coverage, additional observations are not as helpful as they are near the edge of the network. It is necessary to develop advanced algorithms to further improve the quality and quantity of focal mechanisms in these regions. For instance, we can introduce other measurements and constraints into focal mechanism determination, like relative amplitudes and relative polarities among event pairs (Shelly et al., 2016). The methods can also be applied to the Quake Template Matching catalog , which has about 10 times more events than the standard earthquake catalog , to obtain many more earthquake focal mechanisms. By adding more stations, using better phase and polarity pickers, involving additional radiation pattern measurements into calculation, and applying the method to more detected earthquakes, it would be possible to obtain many more high-quality focal mechanisms for even smaller events and lead to the next generation of focal mechanism data set. Note that the ability to analyze signals from earthquakes and solve source properties is limited by the noise levels and attenuation of seismic waves with propagation distance, which is strongly related to earthquake magnitudes, stress drops, and rupture velocities in crustal regions .

Local P-Axis Rotation After Major Ruptures
Rotations of principal stress axes after major earthquakes have been inferred in various regions and mainly attributed to weak faults and large coseismic stress drops (Hardebeck, 2012;Hardebeck & Hauksson, 2001;Hauksson, 1994). The orientation and magnitude of stress rotation could be related to the ratio between stress and stress drop, as well as the angle between the pre-mainshock stress field and the fault orientation (Hardebeck & Hauksson, 2001). Based on the model of Hardebeck and Hauksson (2001), the postseismic stress is expected to rotate toward the fault-normal orientation when the maximum principal stress is oriented at high angle to the fault and toward fault-parallel direction when the maximum principal stress is at low angle to the fault near the mainshock rupture zone.
To examine such relations, we obtain the major co-seismic slip area and the fault strike angle of the El Mayor-Cucapah and the Ridgecrest mainshocks from source inversion results Wei et al., 2011). The major rupture area for the 2010 El Mayor-Cucapah earthquake is between 20 and 70 km along AA′ (Figures 8d-8h) with N312 • E striking fault plane between 20 and 50 km (Figures 8d-8f) and N355 • E striking fault plane between 50 and 70 km along AA′ (Figures 8g and 8h) (Wei et al., 2011). The post-seismic P-axis orientations in these spatial bins all rotate from low-angle to the fault to more fault-parallel orientations. For the 2019 Ridgecrest earthquake, the major slip area is between 20 and 50 km with ∼N322 • E striking fault plane (Figures 9d-9f). The P-axis orientation in these spatial bins all show counterclockwise rotation from ∼N5 • E to ∼N10 • W. The rotations of P-axis azimuth in the major co-seismic slip areas are all consistent with the model of Hardebeck and Hauksson (2001). Because there are more focal mechanisms in the CNN_SoCal catalog, the results show much clearer stress rotation in a much shorter time (<2 months) for the 2010 El Mayor-Cucapah earthquake compared with the results from the YHS catalog (>6 months) (Figures 8d-8h). These results suggest that the stress rotation is mainly related to the stress perturbations in the co-seismic and early post-seismic time periods. Note that the observed stress rotations may not imply temporal variations of the stress field but may reflect noise due to inherent uncertainties in the focal mechanisms and stress inversions (Hardebeck, 2012;Townend and Zoback, 2001;Townend et al., 2012), and/or local stress heterogeneities prior to the mainshock (Smith and Dieterich, 2010;Yukutake et al., 2007Yukutake et al., , 2010 which contradict the assumed uniform stress field in the chosen spatial bin. Local stress heterogeneities are likely to be common and can be caused by geometrical changes of faults, heterogeneous non-tectonic local loadings and other mechanisms (Ben-Zion, 2008;Kim et al., 2021).
After the mainshock ruptures, similar to the observations from many other mainshocks (e.g., Hardebeck & Okada, 2018), many areas show little P-axis back rotations or even forward rotations (Figures 8d, 8e, 8f, 8g, 8h, 9d, and 9f) while some areas have significant P-axis rotations followed by a gradual return within several months (Figures 8b, 8c, 9b, 9c, and 9g). The observed return of P-axis azimuth may not imply the reloading of coseismic stress drop (Hardebeck, 2012) because the major co-seismic slip areas have negligible stress backrotations, suggesting limited stress reloading in the area with large co-seismic slip. The areas with significant stress backrotations are mainly located near the edge of major rupture zones and have more significant postseismic deformations observed from geodetic data compared with the other aftershock zones (Cheng & Ben-Zion, 2020;Ross, Rollins, et al., 2017;. This has also been observed near some other strike-slip faults, such as the 1992 M7.2 Landers earthquake in California (Hardebeck & Okada, 2018;Hauksson, 1994) and1999 M7.4 Izmit earthquake in Turkey (Ickrath et al., 2014). Therefore, the significant P-axis backrotations observed near the edge of major rupture zones may be a combined effect of large co-seismic shear stress perturbations (stress rotation) and post-seismic afterslip and aftershocks (stress backrotation).

Regional Fault Type Changes After Major Ruptures
Major earthquakes typically can cause strong stress perturbations even at large distances and trigger remote aftershocks. Generally, aftershocks are triggered by either static stress changes due to permanent fault displacement or oscillatory dynamic stresses carried by seismic waves (Hill et al., 1993;Kilb et al., 2000). The relative importance of these two mechanisms in earthquake triggering of aftershocks has been debated over decades but is still controversial (e.g., Freed, 2005;Harris, 1998;Richards-Dinger et al., 2010), because both static and dynamic triggering can increase the seismic rate within overlapping time window. One of the significant differences between static and dynamic stresses is the direction of stress perturbations. Static stress changes are described by a single stress tensor while dynamic stress changes do not produce permanent loading but produce forcing in opposite directions when the seismic waves pass by. Therefore, static stress changes can modify focal mechanisms in a consistent way in a large area while dynamic stress triggered earthquakes can exhibit more diverse mechanism patterns.
After the 2010 M w 7.2 El Mayor Cucapah and the 2019 M w 7.1 Ridgecrest earthquakes, the observed systematic spatial variation of the changes of composite faulting style, and the spatial correlation between faulting style with the estimated shear stress perturbation of major ruptures (Meng & Peng, 2014; suggest clear static stress changes even in areas located 100 km away from the mainshock hypocenter. Areas with transtensional focal mechanisms also highlight stress shadows with the opposite direction of static stress tensor and background stress tensor (Harris, 1998). This is difficult to explain from other stressing mechanisms: subsequent earthquake triggering (Meier et al., 2014), dynamic triggering (Felzer & Brodsky, 2005;Meng & Peng, 2014), pressure changes due to fluid motions (Ross, Rollins, et al., 2017), and complex fault geometries (Marsan, 2006). These mechanisms can only affect local areas in short time periods with heterogeneous stress perturbations and cannot explain the observed large areas with transtensional events within 1 year after the mainshock. Moreover, some areas distant from major ruptures with negligible estimated shear stress changes (e.g., the Cajon pass area) also show consistent changes of faulting type, which may be driven by other mechanisms or other sources of stress perturbations instead of oscillatory dynamic stresses. These observations illustrate that the changes of faulting type solved from the CNN_SoCal catalog can be used to infer the changes of horizontal stress, help to differentiate earthquake triggering mechanisms, and identify stress shadows after major earthquake ruptures.

Potential Applications of the CNN_Socal Catalog
In this study we analyze variations within groups of focal mechanisms to examine stress changes, which indicate strong spatiotemporal correlation with major ruptures' co-seismic stress drops and static stress perturbations, and postseismic stress back rotations. In future studies, the high-resolution CNN_SoCal catalog can be used as the input for stress inversions to improve the stress model in southern California (Abolfathian et al., 2020;Yang and Hauksson, 2013), map faults geometry in depth (Plesch et al., 2020), and facilitate studies of major rupture monitoring, modeling, and aftershock forecasting (Hardebeck, 2020). Many other scientific questions can be revisited using the CNN_SoCal catalog. For example, the catalog can be used to analyze various stress loading processes at different time scales, like tidal loading (Cochran et al., 2004), seasonal water loading (Johnson et al., 2017), and inter-seismic stress accumulation (Cheng & Ben-Zion, 2020). Spatially, the fault geometry at depth inferred from focal mechanisms can be used to understand the on/off-fault orientation changes (Cheng et al., 2018), the relationship between surface deformation and seismic deformation at depth (Zhai et al., 2021), as well as the effect of brittle-ductile transition on the changes of fault geometry and stress field near the bottom of seismogenic zone (Schulte-Pelkum et al., 2020). The insights from these topics will further contribute to understanding crustal deformation and potential seismic hazard.

Data Availability Statement
We analyzed waveforms and parametric data from the Caltech/USGS Southern California Seismic Network (SCSN); https://doi.org/10.7914/SN/CI; stored at the Southern California Earthquake Data Center. https://doi. org/10.7909/C3WD3xH1. The seismicity hypocenter parameters from 1981 to the end of 2021 are from the waveform-relocated catalog as described by Hauksson et al. (2012), which uses GrowClust for relocating the most recent version of this catalog (Trugman & Shearer, 2017). The derived focal mechanism catalog is available through the Mendeley Data (https://data.mendeley.com/datasets/9s54cy253d/5) (Cheng et al., 2023). We appreciate the support provided by more than 20 SCSN and SCEDC staff members who maintain stations and communications systems, as well as data flow, processing, and archiving.