Advancing Earthquake Forecasting with Machine Learning

Probabilistic earthquake forecasting has made significant strides in the past decades, to the degree that government agencies around the world have implemented public, real-time systems. The underlying models are largely parametric and statistical, and comprise variants of the self-exciting Hawkes point process, such as the popular Epidemic Type Aftershock Sequence (ETAS) model. ETAS models have gained trust also as a result of their good relative performance in prospective forecast experiments by the Collaboratory for the Study of Earthquake Predictability (CSEP, cseptesting.org) in various tectonic settings around the globe. In recent years, however, machine learning variants of point processes have become available that offer significant advantages: they are much more flexible in their probabilistic description of earthquake interaction, and they are much faster. In this talk, I will review recent applications of neural point processes to seismicity forecasting around the world, which demonstrate distinct advantages and some (moderate) improvement in predictive skill. I will also argue that a clear community benchmarking process is required to make transparent and robust progress. Finally, I will present ongoing model enhancements of neural point processes and preliminary results from benchmarking in California. Machine learning has the potential of transform earthquake forecasting, but progress must be demonstrated in a robust and transparent manner.