New research shows AI earthquake tools forecast aftershock risk in seconds

Current methods used to forecast aftershocks – secondary quakes that can prove more deadly than initial earthquakes – can take several hours or days, but now, new machine learning models have been developed that can forecast where, and how many, aftershocks will take place following an earthquake in close to real-time.

Researchers from BGS, the University of Edinburgh and University of Padua created the AI-driven forecasting tools. They were developed by training machine learning models on earthquake data from California, New Zealand, Italy, Japan and Greece – parts of the world that regularly experience earthquakes.

The rapid forecasts produced by AI-powered tools could help authorities with decision-making about public safety measures and resource allocation in disaster-hit areas. The team analysed AI models’ ability to produce forecasts of how many aftershocks will take place within the 24 hours following earthquakes of magnitude 4 or higher.

They compared the performance of their AI models with the most widely used forecasting system – known as the Epidemic-Type Aftershock Sequence (ETAS) model – which is used operationally in Italy, New Zealand and the US.

While both model types show similar performance at forecasting aftershock risk, the ETAS model took much longer to produce results – up to several hours or days on a single mid-range computer – as it involves running a large number of simulations, the team says.

By training the AI tools on records of past earthquakes from regions with different tectonic landscapes, researchers say their models could be used to forecast aftershock risk in most parts of the world that experience earthquakes.

The research, published in the journal Earth, Planets and Space, was supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie SPIN Innovative Training Network.

This study shows that machine learning models can produce aftershock forecasts within seconds, showing comparable quality to that of ETAS forecasts. Their speed and low computational cost offer major benefits for operational use: coupled with the near real-time development of machine learning-based high-resolution earthquake catalogues, these models will enhance our ability to monitor and understand seismic crises as they evolve.

PhD student Foteini Dervisi, of BGS and the University of Edinburgh’s School of GeoSciences, who led the study.