Benchmarking and Adapting Neural Point Processes for Operational Earthquake Forecasting in California and China

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

Submitted September 7, 2025, SCEC Contribution #14827, 2025 SCEC Annual Meeting Poster #TBD

Recent advances in machine learning have produced Neural Point Processes (NPPs) with the potential to improve earthquake forecasting over classical statistical models such as the Epidemic-Type Aftershock Sequence (ETAS) model. EarthquakeNPP is a benchmarking platform designed to evaluate such models under consistent, reproducible conditions, hosting a suite of NPP implementations alongside ETAS and evaluating them using log-likelihood and CSEP catalog-based metrics.

Building on this platform, we investigate how different, seismologically-informed design choices affect the performance of the Deep Spatiotemporal Point Process (DeepSTPP) model using the China Earthquake Networks Center (CENC) catalog. Specifically, we evaluate including event magnitude, extending the visible history, incorporating an auxiliary training window, and varying magnitude thresholds. These domain-informed configurations improve forecasting in certain regimes, particularly for immediate aftershocks, though architectural limitations in memory and attention limit the gains from event magnitude and longer-range triggering.

We also present preliminary results from the first comparison between NPP models and operational USGS forecasts during the Ridgecrest and Puerto Rico earthquake sequences. While newer, generative NPPs consistently underperform with respect to catalog based evaluations, log-likelihood based NPPs demonstrate marginal temporal information gains over USGS forecasts. These findings motivate the continued adaptation and evaluation of NPPs to advance their readiness for operational earthquake forecasting.

Key Words
forecasting, neural networks, machine learning, operational forecasting

Citation
Werner, M. J., Stockman, S., Tian, W., & Zhang, Y. (2025, 09). Benchmarking and Adapting Neural Point Processes for Operational Earthquake Forecasting in California and China. Poster Presentation at 2025 SCEC Annual Meeting.

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