Can Domain-Informed Design Improve Neural Spatio-Temporal Point Processes for Earthquake Forecasting?

Weixi Tian, Sam Stockman, Yongxian Zhang, & Maximilian J. Werner

In Preparation September 3, 2025, SCEC Contribution #14977

Statistical earthquake forecasting models, such as the Epidemic-type Aftershock Sequence (ETAS), have been developed through decades of empirical seismological studies, embedding formalized findings and assumptions about earthquake triggering, clustering, and catalog completeness. In contrast, recent neural Spatio-Temporal Point Process (STPP) models are often trained without such domain-specific considerations, treating seismicity as generic spatiotemporal data. In this study, we investigate how incorporating seismology-motivated inputs and training strategies into a neural STPP, namely the Deep Spatio-Temporal Point Process (DeepSTPP), affects forecasting performance, using the earthquake catalog provided by the China Earthquake Network Center. These domain-driven designs include introducing magnitude as an input feature, incorporating an auxiliary training period, extending the visible event history, and varying the magnitude threshold. We also benchmark DeepSTPP against ETAS and a homogeneous Poisson process to assess relative strengths and limitations. Our results show that certain domain-motivated configurations, such as including an auxiliary period, improve performance, particularly in forecasting immediate aftershocks. However, limitations in the model architecture, including short memory and attention dilution, constrain the benefit of features such as event magnitude and long-range history. Benchmarking across different magnitude thresholds demonstrates DeepSTPP's superior ability to handle complex real-world catalogs, likely due to its flexibility and generalization capacity. These findings highlight the value of integrating seismological practice into neural model design and point toward future architectures that can fully exploit earthquake catalogs.

Citation
Tian, W., Stockman, S., Zhang, Y., & Werner, M. J. (2025). Can Domain-Informed Design Improve Neural Spatio-Temporal Point Processes for Earthquake Forecasting?. Earth's Future, (in preparation).

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