SCEC Award Number 25227 View PDF
Proposal Category Individual Research Project (Single Investigator / Institution)
Proposal Title Space-time neural point process for earthquake forecasting
Investigator(s)
Name Organization
Kelian Dascher-Cousineau Utah State University
SCEC Milestones D1-1, D2-1, D3-1 SCEC Groups EFP, Seismology, FARM
Report Due Date 07/31/2026 Date Report Submitted 10/09/2026
Project Abstract
 This project extended the Recurrent Earthquake foreCAST (RECAST), a neural temporal point process, to joint space-time forecasting and tested it against a space-time Epidemic-Type Aftershock Sequence (ETAS) model. With a few months of postdoctoral effort, we (1) implemented ST-RECAST, in which the location of the next event is conditioned on its waiting time rather than assumed independent of it; (2) built a pipeline to train it and compare it with a space-time ETAS benchmark; and (3) ran head-to-head tests. On synthetic ETAS catalogs, ST-RECAST captures the timing of the clustering but issues overly diffuse spatial forecasts; its skill improves steadily with training data, yet ETAS still performs better on most test sequences with about 6000 training sequences. Our working hypothesis is that space-time forecasting requires more training data than is available locally. We therefore built machinery to train one RECAST model on 3000 one-year sub-catalogs drawn around roughly a thousand M6.5+ mainshocks in the global ANSS catalog, each with its own completeness level and a balanced contribution to the loss. In pseudo-prospective 14-day forecasts issued 10 minutes after M6.5+ mainshocks, the global model (not yet spatio-temporal) is better calibrated and more accurate than ETAS fitted to each local history, reducing the median relative error by about a quarter on the validation set. This setup mirrors operational aftershock forecasting and offers a route to the data volume that space-time models appear to need.
Intellectual Merit The project produced a space-time extension of RECAST in which locations are conditioned on event timing, and a framework for testing it against space-time ETAS, addressing SCEC milestones D2-1 (benchmarking machine learning against standard models) and D3-1 (alternatives to ETAS). Synthetic tests, in which ST-RECAST improves steadily with training data but remains below ETAS, support our hypothesis that data volume limits neural space-time forecasts within a single region. The global framework addresses this: one model learns aftershock behavior from hundreds of sequences worldwide and already improves on locally fitted ETAS for temporal forecasts.
Broader Impacts The award supported postdoctoral training (Daniel Wells) in neural point-process forecasting and helped launch a new earthquake forecasting group at Utah State University, where the PI started as an assistant professor in January 2025. The models and testing code are openly available on GitHub (danedwells/recast_2) for the community to build on and test independently. Because the global model does not, in principle, need a rich regional catalog, the approach offers a path toward aftershock forecasts in sparsely monitored regions, in the format of current operational forecasts.
Project Participants Kelian Dascher-Cousineau (PI; Assistant Professor, Department of Geosciences, Utah State University): project design, model development, analysis and reporting.
Daniel Wells (Postdoctoral Researcher, Utah State University): model development and testing.
No partner organizations or external collaborators were involved.
Exemplary Figure Figure 3. Global RECAST versus ETAS aftershock forecasts, issued pseudo-prospectively 10 minutes after the first M6.5+ mainshock in each validation sub-catalog (14-day counts of M4.5+ events within 250 km). (A) Calibration: quantile of the observed count within each forecast distribution; a calibrated forecast is uniform (dotted line), and the end bins collect outcomes in the forecast tails. (B) Accuracy: ratio of observed to median forecast counts. One RECAST model trained on global sub-catalogs is better calibrated and more accurate than ETAS fitted to each local history. Credit: K. Dascher-Cousineau (Utah State University).
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.