The Changning region in Sichuan Province is a key industrial area for shale gas extraction and salt mining. Constructing a high-precision earthquake catalog for this region is critically important for investigating regional seismicity and the mechanisms of induced earthquakes. Historically, regional seismicity was relatively quiescent; however, it has increased markedly since 2006, culminating in the M_S6.0 Changning earthquake on 17 June 2019.

The study area is located in the southern Sichuan Basin, covering 28.22°N−28.52°N and 104.60°E−105.12°E. This study uses continuous three-component waveform data recorded by 28 broadband seismic stations deployed in the region from June to October 2021. The inter-station distance ranges from 3 to 5 km, and the sampling rate is 100 Hz. We adopt the DiTingSFM-0.1B seismic foundation model for simultaneous P- and S-wave phase picking, obtaining 1.68 million phase picks （839 847 P and 839 847 S）. For comparison, the PhaseNet method yields 2.35 million phase picks （1 231 511 P and 1 118 579 S）. Under identical conditions, phase association is performed using the REAL algorithm, which produces approximately 30 000 and 50 000 earthquake events, respectively. After association, P-wave travel-time residuals are mostly within 0.2 s and S-wave travel-time residuals within 0.3 s; the DiTingSFM-0.1B model shows a more concentrated distribution of S-wave residuals. Following absolute earthquake location and residual control with the HypoInverse algorithm, 29 249 and 49 277 events are retained. Most travel-time residuals fall in the range of 0.1−0.3 s, and more than 95% of events have residuals smaller than 0.2 s. The travel-time residuals of relocated events show a consistent distribution with hypocentral distance. We also note that isolated S-wave phases detected by PhaseNet are more frequently associated with low-quality events.

Subsequent precise relocation using the HypoDD algorithm demonstrates that S-wave travel-time residuals from DiTingSFM-0.1B remain comparatively smaller. The derived travel-time curves yield average shallow-crust P- and S-wave velocities of approximately 5.0 km/s and 2.86 km/s, respectively, with a v_P/v_S ratio of about 1.75. Depth estimates from precisely relocated hypocenters indicate that S-wave phases picked by DiTingSFM-0.1B produce much tighter depth distributions, concentrated at 2−4 km, enabling the construction of a more reliable earthquake catalog. In contrast, the PhaseNet catalog exhibits more dispersed depth distributions, with more events deeper than 6 km. The final catalog shows that seismicity within the study area mainly forms two distinct clusters trending NW−SE along the Changning anticline. In the western cluster, earthquakes occur over a wide depth range: shallow events at 2−4 km and deeper events extending down to 6−8 km, forming a spoon-shaped pattern consistent with the regional tectonic framework of Changning. In the southeastern cluster, seismicity concentrates at 2−4 km, matching well with local salt-mining activities at about 3 km depth （well depths: 2 600−3 346 m）.

Temporally, the region sustained continuous microseismicity from June to October 2021. Swarm-like activity occurred in mid−late June; July−August was characterized by high seismicity with quasi-periodic fluctuations, followed by steady background activity during September−October. The eastern cluster dominates the overall seismicity, while the western cluster the exhibits weak background seismicity. Microseismic events with M_L＜1.0 predominate throughout the catalog; M_L1.0−3.0 events mainly occur during swarm episodes, and M_L≥3.0 events are rare and nearly all located within the eastern cluster. The magnitude−frequency distribution obeys a clear Gutenberg-Richter relationship, with a completeness magnitude M_c of −0.4 and a b-value of 1.030.

The strong correlation between the spatial pattern and depth characteristics of seismicity, tectonic features and industrial activities illustrates that AI-driven workflows can effectively build high-precision earthquake catalogs and provide robust support for exploring regional earthquake generation mechanisms. Compared with PhaseNet, the paired P-S picking strategy of DiTingSFM-0.1B generates higher-quality phases and imposes tighter depth constraints. PhaseNet, by contrast, detects more events but includes more low-quality events constrained only by isolated single phases. This highlights the critical role of phase quality control in catalog construction. These results also provide valuable guidance for phase identification and quality control when applying AI models to build high-precision catalogs using dense seismic array data.