Towards Flexible and Scalable Earthquake Forecasting

Kelian Dascher-Cousineau, Oleksandr Shchur, & Emily E. Brodsky

Published August 16, 2021, SCEC Contribution #11588, 2021 SCEC Annual Meeting Poster #265

The growth in scale and diversity of earthquake catalogs currently outpaces the development of standard statistical models. Yet, the recognition of a full spectrum of slip calls for a model framework with the capacity to both incorporate or infer about the underlying nucleation processes. We present an application of a scalable, flexible, and statistically rigorous neural model for temporal point processes. We benchmark its performance against an ETAS model using both synthetic and real earthquake catalogs. Key features of this model framework complement current statistical and physics-based approaches: 1) the computational complexity of a neural point process model grows linearly with catalog size rather than the O(n2) growth of ETAS models, 2) the neural temporal point processing skill is not predicated on an observed branching structure, and 3) the model extends naturally to include additional covariates or features without restructuring the model architecture.

Key Words
Earthquake forecasting, Machine learning

Dascher-Cousineau, K., Shchur, O., & Brodsky, E. E. (2021, 08). Towards Flexible and Scalable Earthquake Forecasting. Poster Presentation at 2021 SCEC Annual Meeting.

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