Forecasting earthquakes with neural and bayesian point processesproject period: 8 1 2025 through 7 31 2026AWARD purpose: artificial intelligence (AI) and machine learning are transforming society…

The U.S. Geological Survey (USGS) awarded $218,952.63 to the University of Bristol under the Earthquake Hazards Program Assistance (CFDA 15.807) for the period August 1, 2025 through July 31, 2026. This project grant funds the development and benchmarking of neural point process models for earthquake forecasting. The research will advance artificial intelligence and machine learning applications in seismology by creating a community benchmarking exercise to evaluate neural point processes against established seismological standards, including the Epidemic Type Aftershock Sequence (ETAS) model and the USGS Uniform California Earthquake Rupture Forecast 3 (UCERF3-ETAS) model. The deliverables will include enhanced neural point process models capable of making spatial and magnitude forecasts in addition to temporal forecasts, with comparative performance evaluations conducted using multiple datasets through the Collaboratory for the Study of Earthquake Predictability (CSEP).

This project directly supports the USGS program priority of developing advanced methods for earthquake hazard assessment and characterization. By establishing rigorous benchmarking protocols and performance metrics aligned with seismological community standards, the research will help validate the operational forecasting potential of neural point process models and provide a foundation for their potential integration into real-time earthquake forecasting systems used worldwide.

Activities

(1) develop a community benchmarking exercise for existing neural (and other) point process models. (2) extend the identified neural models to be able to ingest and forecast magnitude information, which is the most critical next step to improving the models. We will extend the models architectures and RE-assess their predictive skills. Finally, armed with credible space-time-magnitude neural point processes,(3) compare them against the UCERF3-etas model, as well as a recent approximate bayesian etas model. The latter has not yet been deployed at regional and multi-decadal scale, but IT is currently the only bayesian space-time-magnitude etas model sufficiently fast to process large catalogs and yet offer some forecast robustness to parameter uncertainties. Deliverables and outcomes: deliverable 1 is a community earthquake forecasting benchmark for neural and other point process models using multiple datasets in california. Deliverable 2 comprises version-controlled public codes to fit and run space-time-magnitude neural point processes, extended from best-performing space-time neural models. We expect substantial predictive gains over their baselines. Deliverable 3 consists of a workflow to compare the new neural models with UCERF3-etas as well as the approximate bayesian model. Broader outcomes include a more AI-trained earthquake forecasting community, strengthened ties to the AI community, and the potential for better operational forecasting.

Intended beneficiaries

Those benefiting from improved operational forecasts, such as for improved situational awareness and decision-making during earthquake sequences. We aim to meaningfully connect rapid advances in AI with urgent application needs in seismology, ensuring that future AI applications to earthquake forecasting can be meaningfully assessed for potential incorporation into operations. Subrecipient: N a