Earthquake early warning and rapid earthquake reporting require early estimates of event size and location, particularly for offshore earthquakes and sparsely instrumented regions, where conventional multi-station association may incur delays. We develop a prototype for rapid single-station event characterization and teleseismic warning using the DiTing seismic waveform foundation model. This study demonstrates an integrated workflow on representative cases, without claiming universal warning performance. The workflow integrates waveform reception, P-wave identification, catalog magnitude and epicentral-distance estimation, back-azimuth estimation, progressive location updates, and message dissemination.

DiTing provides a pretrained waveform encoder and independently fine-tuned scalar regression heads for catalog magnitude and epicentral distance. Three-component records are ordered as Z, N, E and resampled to 100 Hz. Each component is standardized using its own mean and standard deviation. Waveform patches are converted into tokens, processed by the full pretrained encoder, and fused before being fed into task-specific multilayer perceptron heads. The regression target is the source catalog magnitude, denoted as M_cat. Because the contributing magnitude scales are not physically interchangeable, magnitude errors quantify agreement with catalog labels rather than the absolute accuracy on a uniform physical scale. The reproducibility description covers prototype preprocessing and inference only, excluding undocumented training parameters of the foundation model.

Continuous processing adopts 10 000-sample windows with a 5 000-sample step, corresponding to 100 s and 50 s, with zero padding for incomplete segments. Magnitude estimation uses a time window spanning 10 s before to 90 s after the P-wave arrival. Online distance estimation uses the window from 0.03 s before to 3.32 s after arrival, with neighboring offsets of −0.03, 0, and ＋0.03 s. These online settings differ from the observation windows used in the standalone distance evaluation. Predictions from adjacent windows are aggregated and assessed for consistency. We monitor medians, dispersion, magnitude variations and temporal persistence, instead of treating a transient peak value as a stable estimate. Component orientation and polarity corrections rely on station metadata.

Back azimuth is estimated via an independent network trained and validated using terrestrial STEAD data. The polarization-estimation method proposed by Li et al serves as a methodological reference and does not supply Martian training data. The input consists of three-component waveforms from 0.5 s before to 3.5 s after the P-wave arrival, supplemented by nearby windows offset by −0.02, 0, and ＋0.02 s. Maximum-absolute-amplitude normalization is applied, and polarization information is extracted from the eigenvalues and eigenvectors of the three-component covariance matrix. Convolutional waveform and polarization branches are combined prior to the fully connected layers, which predict the sine and cosine components of back azimuth. Circular statistics are used to aggregate predictions from neighboring windows. At the single-station stage, epicentral distance and back azimuth are used to compute an initial epicenter released in the prototype’s first report. This constitutes an early estimate, and its issuance does not guarantee operational location accuracy.

As additional stations detect P waves, their predicted epicentral distances are incorporated into the location constraints, and the previous estimate provides an initial solution for the subsequent update. The prototype framework evaluates station residuals, weighted root-mean-square residuals （WRMS）, location shifts, and the contribution of individual observations. Waveform quality, component completeness, prediction stability, and uncertainty govern the weighting of each record. Inconsistent observations may be downweighted and examined via leave-one-out recalculation. Back azimuth remains quality-controlled auxiliary information rather than an unconstrained directional constraint. The quality-control framework accounts for waveform anomalies, model-output stability, and consistency across location constraints. Thresholds and weighting rules require further calibration beyond the test cases presented here; the limited evaluation cannot yield values directly transferable to all seismic networks and instruments. Initial and updated results are distributed via MQTT messages.

Evaluations adopt separate datasets for distinct purposes. Catalog-magnitude regression is assessed using 2 009 near-field P-wave records from 589 stations with epicentral distances less than 150 km. The resulting bias, mean absolute error （MAE）, and root-mean-square error （RMSE） are ＋0.236, 0.349, and 0.406 magnitude units, respectively. Epicentral-distance estimation is tested using 1 194 records from 429 permanent stations, recorded between 28 February and 12 July 2023, at epicentral distances less than 500 km. The same dataset is used for P-wave observation windows of 3, 5, 10, 20, 60, and 90 s. The corresponding distance biases are −0.70, −0.41, −0.47, −0.31, −0.08, and −0.07 km; the respective MAEs are 5.78, 4.92, 4.04, 3.28, 2.72, and 2.32 km, and RMSEs are 8.04, 6.82, 5.58, 4.54, 3.76, and 3.23 km. Longer windows improve the overall consistency of distance estimation for this dataset. Residual distributions still exhibit large outliers, and no assumption of normality is made.

Magnitude evolution at larger distances is illustrated using six station records from three large earthquakes: the 2024 Hualien event （Taiwan）, the 2024 Kyushu event （Japan）, and the 2025 central Myanmar event, with catalog magnitudes of 7.3, 7.1, and 7.9, respectively. Station epicentral distances range roughly from 300 to 1 400 km. Some records show rapid magnitude growth within the first few seconds after P-wave arrival, yet subsequent temporal trajectories vary across stations. The two Myanmar records, both beyond 1 000 km, show magnitude underestimation within the first 10 s, with estimates converging toward catalog values over the full window. Wave propagation distance, limited information within early short windows, waveform evolution, and domain mismatch between training and test data are plausible explanations, though none are conclusively verified. These records do not demonstrate systematic underestimation for all great earthquakes, nor reliable generalization to all distant source scenarios.

System-level case testing uses 15 single-station records from four off-network events in Xinjiang, offshore Taiwan, Kashmir, and Mongolia. Across all 15 records, the catalog-magnitude MAE is 0.234 and the median absolute error is 0.233. The median absolute epicentral-distance error is 8.86 km. In contrast, the median absolute back-azimuth error reaches 163.26°, and the median single-station epicentral offset is 148.32 km. Near-180° directional errors in some records highlight a major limitation of single-station initial location. Low magnitude and distance errors alone therefore cannot guarantee an accurate single-station epicenter. Potential contributing factors include component orientation and polarity, waveform noise, polarization stability, wave propagation effects, and domain differences between training and application datasets. The available cases do not quantify their individual contributions.

Progressive location updates refine some event solutions but do not yield monotonic error reduction. For the Xinjiang event, epicentral error decreases from 74.12 km in the initial stage to 13.81 km and 8.23 km, followed by a rise to 9.92 km at the fourth update stage. For the Kashmir event, the error first increases from 39.64 km to 119.99 km before dropping to 10.81 km and 5.64 km. For the Mongolia event, the third-stage error of 7.49 km increases to 30.24 km after incorporating an additional observation, while WRMS rises from 1.40 km to 3.07 km. These examples emphasize the need to check consistency and leverage of new observations instead of assuming that adding stations always improves location accuracy. The update-stage analysis and the 15-record single-station analysis use different sampling units and should not be merged into a single performance metric.

A real-time example of the 20 April 2026 earthquake offshore eastern Japan demonstrates the full processing chain using station IU.MAJO at an epicentral distance of approximately 560 km. The prototype estimates M_cat as 7.43, compared with the catalog magnitude of 7.4. The latency from identified P-wave arrival to message dissemination is approximately 82 s. This end-to-end duration includes data transmission and packet overhead and is not merely the neural network inference time; it also does not represent warning lead time at a target site. Overall, the results validate the operability of the prototype processing workflow for the tested events and show the potential to retrieve early event-size and distance information from distant single-station P waves. However, the small set of regional and far-field case studies restricts statistical generalization. Further extensive testing, benchmark comparisons against conventional methods, and continuous waveform-based estimates of detection, false-alarm and missed-event rates, latency, and stability are required before operational performance can be confirmed.