Using Deep Learning for Flexible and Scalable Earthquake Forecasting

• Published in: Geophysical research letters Vol. 50; no. 17
• Main Authors: Dascher‐Cousineau, Kelian, Shchur, Oleksandr, Brodsky, Emily E., Günnemann, Stephan
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 16.09.2023; Wiley
• Subjects: Benchmarks; Catalogues; Datasets; Deep learning; earthquake; Earthquake data; Earthquake forecasting; Earthquake prediction; Earthquakes; ETAS; Forecast accuracy; Forecasting; Machine learning; Mathematical models; Model testing; Modelling; RECAST; Seismic activity; Seismology; Statistical analysis; Statistical models; Synthetic data; Training
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2023GL103909

Seismology is witnessing explosive growth in the diversity and scale of earthquake catalogs. A key motivation for this community effort is that more data should translate into better earthquake forecasts. Such improvements are yet to be seen. Here, we introduce the Recurrent Earthquake foreCAST (RECAST), a deep‐learning model based on recent developments in neural temporal point processes. The model enables access to a greater volume and diversity of earthquake observations, overcoming the theoretical and computational limitations of traditional approaches. We benchmark against a temporal Epidemic Type Aftershock Sequence model. Tests on synthetic data suggest that with a modest‐sized data set, RECAST accurately models earthquake‐like point processes directly from cataloged data. Tests on earthquake catalogs in Southern California indicate improved fit and forecast accuracy compared to our benchmark when the training set is sufficiently long (>104 events). The basic components in RECAST add flexibility and scalability for earthquake forecasting without sacrificing performance. Plain Language Summary We explore the potential for deep learning in earthquake forecasting. Prior work has relied heavily on statistical models that do not scale to fully utilize the currently available large earthquake data sets. Here we build on recent developments in deep learning for forecasting event sequences in general to create an implementation for earthquake data. The new approach allows us to incorporate larger data sets, potentially with more information about each earthquake. We also avoid a specific functional form, so the method naturally adapts to additional information about events, like magnitude or variations in behavior over time. As we add more data, results show continued improvements. This ability to incorporate and improve continually as training data sets increase indicates that there is more information in the earthquake catalogs than has yet been used for earthquake forecasting. Key Points We introduce a deep learning model for earthquake forecasting and explore its performance on synthetic and regional earthquake data sets It is flexible in the sense that a predefined functional form is not required It is scalable in two senses: it is efficient on large data sets, and its performance relative to benchmarks improves with more training data