Oral Presentation

Seismology is witnessing explosive growth in the diversity and scale of earthquake catalogs owing to improved seismic networks and increasingly automated data augmentation techniques. A key assumption in this community effort is that more detailed observations should translate into improved earthquake forecasts. Current operational earthquake forecasts build on seminal works designed for sparse earthquake records that combine the canonical statistical laws of seismology. This parsimonious approach is remarkably robust and stubbornly difficult to improve upon. Advances in the past decades have mainly focused on the regionalization of the models, the recognition of catalog peculiarities, and the extension to spatial forecasts; but have failed to leverage the wealth of new geophysical data. Here, we develop a neural-network-based earthquake forecasting model that leverages the new data in an adaptable forecasting framework: the Recurrent Earthquake foreCAST (RECAST). We benchmark temporal forecasts generated by RECAST against the widely used Epidemic Type Aftershock Sequence (ETAS) model using synthetic and observed earthquake catalogs. We consistently find improved model fit and forecast accuracy for Southern California earthquake catalogs with more than 10,000 events. The approach provides a flexible and scalable path forward to incorporate additional data into the earthquake forecast.