New AI tools predict aftershock risks within seconds of a major earthquake

A research team from the University of Edinburgh’s School of GeoSciences, the British Geological Survey and the University of Padua has developed artificial intelligence tools that can predict aftershocks almost as soon as the ground stops shaking.

The models were trained using detailed earthquake records from regions with different tectonic environments, including California, New Zealand, Italy, Japan, and Greece. These areas are among the world’s most seismically active, providing millions of data points from decades of recorded tremors.

Using this global dataset, the team built deep learning systems capable of forecasting where aftershocks will occur and how many will follow within the first 24 hours after a M4.0 or stronger earthquake. For context, M4.0 quakes typically release as much energy as 15 tons (13.6 tonnes) of TNT.

The new models were tested against the current operational standard, known as the Epidemic-Type Aftershock Sequence (ETAS) model, which is widely used in Italy, New Zealand, and the United States. The AI systems achieved comparable accuracy while producing forecasts in seconds instead of hours.

“Machine learning models can produce aftershock forecasts within seconds, showing comparable quality to that of ETAS forecasts,” said lead author Foteini Dervisi of the University of Edinburgh and the British Geological Survey. “Their speed and low computational cost offer major benefits for operational use.”

Why speed matters when the ground keeps shaking

Aftershocks often pose the greatest danger following a major earthquake. They can collapse structures weakened by the initial event, trigger landslides, and further damage infrastructure.

Traditional forecasting systems such as ETAS simulate thousands of possible aftershock patterns, a process that can take several hours or even days on a typical computer. In fast-changing disaster zones, those delays can cost lives.

By contrast, the new AI models generate results almost instantly once they receive seismic data from sensors. The researchers note that this enables near real-time updates, allowing authorities to direct rescue teams, issue evacuation warnings, or close high-risk infrastructure without waiting for manual calculations.

The speed advantage also allows the models to be retrained continuously as new data becomes available. Each additional earthquake recorded improves the system’s understanding of seismic patterns, making it more accurate over time.

According to the study, such rapid modeling could be integrated with automatic alert systems, giving emergency coordinators access to dynamic risk maps as an earthquake sequence unfolds.

The science behind the models

The researchers tested two deep learning architectures: the Small Attention U-Net, a compact convolutional neural network, and Earthformer, a transformer-based model designed for spatiotemporal data. Both process sequences of daily earthquake activity maps to forecast how seismicity will evolve in the following 24 hours.

Each model receives information from the previous seven days, including the number of earthquakes per location, the largest magnitude, and the average depth. By learning the relationships among these factors, the AI can identify patterns invisible to human analysts.

Once trained, the models instantly transform new seismic input into a detailed map of likely aftershock zones. The spatial resolution is 0.1 degrees of latitude and longitude, equivalent to about 11 km (6.8 miles) at the equator.

Tests showed that both architectures performed similarly in forecasting accuracy, but the smaller U-Net variant required far less computation. This efficiency means it can be updated more often and deployed on modest hardware in field operations, where access to supercomputers may be limited.

Unlike earlier systems that depend on fixed formulas, these AI tools learn directly from historical data, capturing regional variations in fault behavior and geological complexity.

High-resolution earthquake catalogues

One of the breakthroughs enabling these models was the use of high-resolution earthquake catalogues. Recent advances in automated seismic monitoring allow computers to detect and record up to ten times more small events than traditional manual catalogues.

These detailed datasets provide a clearer picture of how stress transfers through the crust during and after a mainshock. Even earthquakes too small to be felt by people can signal where the next larger aftershock might occur.

The team fine-tuned its models using these enhanced catalogues from Southern California and central Italy. This training improved the accuracy of spatial predictions, particularly in regions where fault networks are complex or overlapping.

However, the researchers note a continuing challenge known as short-term aftershock incompleteness. Immediately after a large quake, overlapping seismic waves can obscure smaller signals, making it difficult to detect every aftershock in real time. Overcoming this limitation remains a priority for improving AI-based forecasts.

The integration of high-resolution data and machine learning may ultimately lead to models that not only predict aftershocks but also provide new insights into the physical laws governing earthquake sequences.

Comparing artificial and traditional forecasting

In direct comparisons, the AI models matched the performance of ETAS in both accuracy and spatial distribution while running thousands of times faster.

ETAS, a statistical model first proposed in 1988, treats earthquake sequences as branching processes in which each event can trigger others. It remains a standard in seismology but requires repeated simulations to estimate probabilities, consuming significant computational resources.

By contrast, once trained, the AI systems can produce forecasts instantly and repeatedly without additional simulations. This makes them more practical for operational settings such as national geological agencies or disaster management centers.

The researchers found that all models slightly underpredicted the total number of aftershocks, largely due to incomplete datasets, but the spatial forecasts were consistent with observed patterns. In most tests, both SmaAt-UNet and Earthformer correctly identified zones of heightened risk around the mainshock area.

The ability to deliver forecasts within seconds could enable continuous updates throughout an aftershock sequence, enhancing public safety and optimizing the deployment of emergency services.

Toward global seismic forecasting

Because the models were trained on data from diverse tectonic environments, they can generalize beyond any single region. This opens the possibility of using the same system in areas with limited seismic monitoring infrastructure, provided minimal data are available.

The approach is especially valuable for developing countries along major fault systems, where access to advanced computing or real-time analysis tools is often limited.

Future research will focus on integrating live data streams from global seismic networks, allowing the models to operate in fully automated mode. Researchers envision systems that can map risk within seconds of an earthquake, displaying likely aftershock zones on interactive dashboards accessible to civil defense authorities.

The study’s authors emphasize that artificial intelligence should complement, not replace, physics-based models. Combining statistical understanding with machine learning’s pattern recognition offers the most promising route to reliable, fast, and adaptive earthquake forecasting.

The road ahead

As global data quality improves and computing power becomes more accessible, the integration of AI into operational seismology is expected to accelerate.

The researchers argue that the next step is to pair AI models with real-time earthquake catalogues produced by machine learning detectors. These automated systems can now identify even the smallest earthquakes almost immediately after they occur, feeding data into forecasting algorithms in near real-time.

This combination could revolutionize the management of seismic crises. Instead of static hazard maps, authorities would have live, continuously updated forecasts showing how risk evolves hour by hour after a major event.

The implications extend beyond emergency response. By revealing the complex statistical relationships between mainshocks and aftershocks, these AI models could also deepen scientific understanding of earthquake physics.

Ultimately, the goal is a global, adaptive earthquake monitoring network capable of issuing accurate forecasts in seconds, providing communities with critical information when every moment counts.

References:

¹ AI quake tools forecast aftershock risk in seconds, study shows – The University of Edinburgh – November 25, 2025

² Towards a deep learning approach for short-term data-driven spatiotemporal seismicity rate forecasting – Dervisi, F., Segou, M., Poli, P. et al – Earth, Planets and Space – November 25, 2025 – https://doi.org/10.1186/s40623-025-02241-6 – OPEN ACCESS

Reet Kaur

I’m a science journalist and researcher at The Watchers, contributing to the Epicenter edition, where I cover peer-reviewed scientific research and emerging discoveries across Earth and space sciences. With a background in astronomy and a passion for environmental science, I’ve worked in shark and coral conservation in Fiji, conducting reef and shark-behavior research, contributing to mangrove restoration, and earning PADI Open Water and Coral Reef Certifications. I bring a blend of scientific rigor and storytelling to illuminate the discoveries shaping our planet and beyond.

This author does not have any more posts.