Earthquake Forecasting Using Single-Station Waveform Detection Without Reliance on Event Catalogs

Yuriko Iwasaki, Emily E. Brodsky, & Kelian Dascher-Cousineau

Submitted September 7, 2025, SCEC Contribution #14304, 2025 SCEC Annual Meeting Poster #199

Earthquake forecasting directly from waveforms, bypassing traditional catalogs, promises enhanced efficiency since catalogs are derived from waveforms and the raw data inherently encapsulates a broader spectrum of information. For instance, utilizing the waveform itself offers the advantage of capturing the influence of minor events that might be overlooked in published catalogs due to low quality or being below the threshold of detectability. To incorporate these subtle events, we employed single-station waveform detection, which preserves critical information regarding phase arrival times and amplitudes, thereby enabling the calculation of origin times, distances, and magnitudes for small events that cannot be recorded by multi-station networks. The single-station detection methodology has had prior success in Guo et al. (2025) where it yielded enhanced results in dynamic triggering. Here we extend the approach to general earthquake forecasting.

As an initial proof-of-concept, we applied PhaseNet (Zhu & Beroza, 2018) to waveforms recorded at a single station to identify the P and S phases, determine hypocenter locations, and subsequently integrate this information into Recurrent Earthquake foreCAST (RECAST)—an advanced deep-learning framework that utilizes neural temporal point processes for earthquake prediction (Dascher-Cousineau et al., 2023). The results suggest that this method potentially achieves performance comparable to that of the RECAST implementation that utilizes a published catalog. Moreover, single-station predictions may offer the additional advantage of applicability in regions where instrumentation is limited.

Incorporating smaller events necessitates consideration of the quality of hypocentral information, as some minor events may be poorly resolved and should thus contribute with lower weight in predictions. By inputting the probabilities for the P and S phases calculated from PhaseNet into RECAST, we tested this method. The preliminary results suggest that earthquake forecasting can be effectively achieved using single-station waveform data alone.

Key Words
Earthquake forecasting, machine learning, single-station detection

Citation
Iwasaki, Y., Brodsky, E. E., & Dascher-Cousineau, K. (2025, 09). Earthquake Forecasting Using Single-Station Waveform Detection Without Reliance on Event Catalogs. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)