physics.geo-phOct 12, 2026
SaveEvent-Centric Earthquake Monitoring with a Multi-Station Transformer
Organizations: Department of Earth & Planetary Science, University of California, Berkeley · Berkeley Seismology Laboratory, University of California, Berkeley
Abstract
Earthquake catalogs underpin studies such as fault structure, earthquake processes and seismic hazard, and their value depends on being both complete and reliable. Machine learning has improved individual tasks of earthquake monitoring, but processing stations independently discards the consistency of arrivals across the network, and running tasks in sequence fragments the information describing each earthquake. We present EQNet2, a geometry-aware multi-station Transformer that jointly predicts phase arrivals, first-motion polarities, source location and magnitude, and the association of arrivals into events. On a Northern California test dataset and on continuous waveforms from Ridgecrest and the San Francisco Bay Area, EQNet2 recovers more weak arrivals, produces fewer false detections, and associates more picks into located events. Its learned attention links information across stations along time--distance patterns consistent with seismic moveout. As seismic networks grow denser, exploiting multi-station coherence and event-centric prediction offers a path to more complete and reliable catalogs.
Figures & tables
Figure 1: Event-centric formulation and geometry-aware architecture of EQNet2. a, Multi-station waveforms and relative station geometry are encoded into a shared representation from which phase arrivals, first-motion polarities, event-arrival proxies, source-related estimates, and network-wide association probability masks are jointly predicted for the same candidate events. b, Detailed structure of EQNet2. Multi-station waveforms are processed by a hierarchical Swin Transformer with spatiotemporal relative position bias and fused through a feature pyramid network before task-specific decoding. Predicted event-arrival proxies serve as prompts for event-specific association. PPM, pyramid pooling module ( Zhao et al., (2017) ) . Detailed architectures of the prediction heads are shown in Supplementary Fig. 1 .
Figure 2: Example of event-level outputs from EQNet2. a, Vertical-component waveforms sorted by source–station distance. Blue and red ticks mark predicted P- and S-phase arrivals, respectively; green ticks mark station-wise origin-time estimates obtained from the predicted event-arrival proxies and travel times. The orange diamond denotes the query event-arrival proxy, and orange shading shows the corresponding association probabilities. b,c, Vertical-component waveforms aligned on the predicted P arrivals and colored by predicted first-motion polarities. Polarities are categorized as Up (U), Unknown (N), or Down (D) using a score threshold of 0.3. d, Seismic stations, station-wise epicentral estimates, their centroid, and the reference epicenter. Dashed lines indicate predicted backazimuths, and station colors indicate predicted first-motion polarities.
Figure 3: Phase-picking performance across signal-to-noise-ratio groups. Precision (triangles), recall (squares), and F1 score (circles) of P-phase (top) and S-phase (bottom) picking on the NCEDC 2023 dataset. The vertical axis is compressed below 0.95 (dashed line). Gray bars show the number of manual picks in each group (right axis).
Figure 4: Comparison of earthquake catalogs for the 2019 Ridgecrest sequence. a, Temporal event frequency. b, Magnitude distribution. c, Depth distribution. d–g, Spatial distributions for SCSN, PhaseNO+GaMMA, PhaseNet+, and EQNet2. Event counts are shown in the legend, and labels in panels e–g report event recovery relative to the SCSN catalog. Colors in panels d–g indicate event depth.
Figure 5: Comparison of EQNet2 and PhaseNet+ for the 2025 Berkeley earthquake sequence. a–d, Predictions in a 2-hour window around the 2025-09-22 Berkeley earthquake (event ID: nc75240492). Each square in panels b and d denotes a predicted event origin time at the corresponding station. Gray squares are unassociated predictions, colored squares indicate predictions associated with the same event, and the black star marks the USGS catalog event. Panels a and c enlarge the shaded interval and additionally show predicted P- and S-phase arrivals. e, Stacked counts of associated and unassociated P- and S-phase picks. Percentages indicate the fraction of raw picks retained after association and location.
Figure 6: Relative position biases across the EQNet2 backbone. One attention head is shown for each of the four Swin Transformer stages. The horizontal axes denote relative time between waveform tokens, and the vertical axes denote the signed spatial offset of one station relative to the other, in degrees. Because the learned biases are nearly rotationally symmetric, the inter-station azimuth is fixed in the north–south direction and the relative elevation at zero. Darker colors indicate larger bias values. More visualizations are provided in Supplementary Figs. 14 and 15 .
Figure 7: Cross-station attention for a representative event. The attention map for event nc73934641 from the last Swin Transformer block is shown together with the corresponding station distribution. The query token containing a P-phase arrival is enclosed by a purple rectangle, the query station is marked by a purple triangle, and darker colors indicate higher attention weights.
Supplementary Fig. 1 : Structure of task heads. The Association Head is modified from the Segment Anything Model ( Kirillov et al., (2023) ; Ravi et al., (2024) ) ; the other heads mainly consist of 1D convolution modules.
Supplementary Fig. 2 : Precision–recall curves of phase picking and event detection on the NCEDC 2023 dataset. Yellow stars mark the highest F1 score for each model.
Supplementary Fig. 3 : Phase picking and event detection residual distributions of EQNet2 and its single-station variant on the NCEDC 2023 dataset.
Supplementary Fig. 4 : Confusion matrix for P-phase first-motion polarity determination on the NCEDC 2023 dataset. The three classes, D, N, and U, denote Down, Unknown, and Up polarities, respectively.
Supplementary Fig. 5 : Window-wise association ratio on the NCEDC 2023 dataset by spatiotemporal clustering and prompt-based method. The association ratios are expected to concentrate on curves whose equations follow , where is the total number of signals and is the number of unassociated ones.
| Precision | Recall | F1 | Mean (s) | Std (s) | ||
| EQNet2 | P-phase | |||||
| S-phase | ||||||
| EQNet2-Single | P-phase | |||||
| S-phase | ||||||
| PhaseNO | P-phase | |||||
| S-phase |
Supplementary Table 1 : Phase-picking metrics and time-residual statistics at a score threshold of 0.3.
| Detection Metrics | Origin Time (s) | Distance (km) | Backazimuth (deg) | ||||||
| Model | Precision | Recall | F1 | Mean | Std | Mean | Std | Mean | Std |
| EQNet2 | |||||||||
| EQNet2-Single | |||||||||
| PhaseNet+ | 0.9260 | 0.9987 | 0.9610 | 0.004 | – | – | – | – | |
Supplementary Table 2 : Event detection residual statistics at threshold 0.3.
Supplementary Fig. 6 : Phase picking F1, precision and recall scores for P- and S-phase arrival picking on the NCEDC 2023 dataset across magnitude groups. Symbols are the same as Figure 3 in the main text.
Supplementary Fig. 7 : Phase picking F1, precision and recall scores for P- and S-phase arrival picking on the NCEDC 2023 dataset across epicentral-distance groups. Symbols are the same as Figure 3 in the main text.
Supplementary Fig. 8 : Phase picking F1, precision and recall scores for P- and S-phase arrival picking on the NCEDC 2023 dataset across azimuth groups. Symbols are the same as Figure 3 in the main text.
Supplementary Fig. 9 : A dense-event window from the NCEDC 2023 test dataset. Predicted P- and S-phase arrivals from EQNet2, PhaseNet+, PhaseNet, and PhaseNO are shown for a 90-s window containing two cataloged events, nc73848531 and nc73848536. Solid lines denote predictions and dotted lines denote catalog annotations; blue and red indicate P- and S-phase arrivals, green indicates origin-time estimates, and orange indicates event-arrival proxies. Event-level predictions are available only for EQNet2 and PhaseNet+.
Supplementary Fig. 10 : Epicenter locations predicted by EQNet2 with different station window sizes before the absolute location procedure: (a) SCSN (reference), (b) 16, (c) 45 (all sensors).
Supplementary Fig. 11 : Spatial distribution of P-axis azimuths from focal mechanism solutions: (a) SCSN catalog, (b) catalog of Cheng et al. ((2023)) , (c) PhaseNet+ catalog, and (d) EQNet2 catalog.
Supplementary Fig. 12 : Comparison of EQNet2 and PhaseNet+ for the 2019 Ridgecrest earthquake sequence. (a–d), Predictions in a time window selected from the 2019 Ridgecrest sequence around the mainshock (event ID: ci38457511). Each square in panels b and d denotes a predicted event origin time at the corresponding station. Gray squares are unassociated predictions, colored squares indicate predictions associated with the same event, and the black star marks the SCSN catalog event. Panels a and c enlarge the shaded interval and additionally show predicted P- and S-phase arrivals. (e) Stacked counts of associated and unassociated P- and S-phase picks. Percentages indicate the fraction of raw picks retained after association and location.
Supplementary Fig. 13 : Pick-level evidence for the Berkeley one-day comparison. (a) Counts of located picks as a function of single-trace signal-to-noise ratio for EQNet2 and PhaseNet+. (b) Unassociated picks through the day, separated into picks adjacent to a located event, picks coincident across stations, and isolated picks.
Supplementary Fig. 14 : Relative position bias for all attention heads in the shallow and intermediate backbone stages. Panels a–c show blocks 1_1, 2_1, and 3_5, respectively. The horizontal axes denote relative time, the vertical axes denote relative spatial offset, and darker colors indicate larger bias within each head. Because the bias ranges differ among heads, with approximate ranges of 0–8 or 0–16, every panel is normalized independently.
Supplementary Fig. 15 : Relative position bias for all 16 attention heads in the deepest backbone stage (block 4_1). The horizontal axes denote relative time, the vertical axes denote relative spatial offset, and darker colors indicate larger bias within each head. The diversity of patterns suggests specialization among attention heads. Because the bias ranges differ among heads, with approximate ranges of 0–8 or 0–16, every panel is normalized independently.
Supplementary Fig. 16 : Additional attention maps for event nc73934641 from the deepest Swin Transformer block. The three query stations were selected to represent different positions within and near the edge of the station network. The token containing the query P arrival is enclosed by a purple rectangle, the query station is indicated by a purple triangle, and darker colors indicate higher attention weights.
| Category | Metric | PhaseNet+ | EQNet2-Tiny | EQNet2-Small | EQNet2-Base | PhaseNO |
| Model size | GFLOPs | 2.72 | 3.49 | 10.69 | 37.75 | 167.23 |
| Params (M) | 1.55 | 1.02 | 4.02 | 15.97 | 2.37 | |
| Runtime | Time (ms), 16 sta. | 3.7 | 24.4 | 24.9 | 31.9 | 167.2 |
| VRAM (GB), 16 sta. | 0.14 | 0.25 | 0.27 | 0.42 | 6.88 | |
| Time (ms), 32 sta. | 5.1 | 26.9 | 29.3 | 54.9 | 685.3 | |
| VRAM (GB), 32 sta. | 0.27 | 0.49 | 0.53 | 0.76 | 27.25 |
Supplementary Table 3 : Computational cost and performance scaling statistics. EQNet2-Base corresponds to the default EQNet2 model used in the main-text experiments.
| Station window size | EQNet2-Tiny | EQNet2-Small | EQNet2-Base |
| 18 | 6.06% 2,801 / 46,206 | 7.68% 2,967 / 38,625 | 9.33% 2,683 / 28,762 |
| 36 | 8.66% 2,976 / 34,361 | 11.24% 3,139 / 27,921 | 12.73% 3,020 / 23,727 |
| 69 | 13.06% 3,417 / 26,171 | 17.60% 3,838 / 21,803 | 22.06% 3,501 / 15,871 |
Supplementary Table 4 : Catalog-level association ratio (%) in the Berkeley earthquake case for different EQNet2 model sizes and station-window sizes for attention. Each cell reports the association ratio, followed by the number of located picks over the number of raw picks (located/raw). The 69-station configuration uses the full station set.
More in this topic
Seismology
Seismology studies earthquakes and seismic waves to understand Earth’s structure, monitor faults, and assess geological and infrastructure hazards.
See all 43 papers