Harvard University and Google researchers described an AI system in August 2018 that predicted where aftershocks were likely to occur after an earthquake. The model learned from data on 199 earthquakes and more than 130,000 aftershocks.
The project illustrated a valuable use of machine learning in science: finding patterns in complex physical data that could help experts target attention after a disaster.
Aftershocks Created Ongoing Danger
An initial earthquake could be followed by smaller or sometimes damaging tremors. Emergency teams needed to assess where risk remained elevated while people inspected buildings, restored infrastructure and planned recovery.
Traditional forecasts were possible but could require significant time and simplifying assumptions. A more accurate map of likely aftershock zones could support better preparation.
The tool was designed to inform risk assessment, not to predict an earthquake with certainty.
The Dataset Captured Many Past Events
Training on 199 earthquakes and more than 130,000 aftershocks gave the model a broad set of examples. It could learn relationships between features of the initial rupture and the spatial distribution of later shocks.
Historical data did not contain every possible geological setting. A model might perform differently in a region with fault structures unlike those represented in the training set.
Scientists needed regional validation before relying on predictions for public decisions.
Forecasts Extended Beyond the Immediate Hours
The contemporary reporting said the system could identify aftershock risk up to a year after the initial event. This challenged the public assumption that all relevant danger was confined to the first short period.
Longer forecasts could help agencies plan inspections and communicate ongoing risk. They also carried uncertainty that needed to be explained carefully.
People should not interpret a probability map as a guarantee that a location was safe or unsafe.
Computational Earth Science Bridged Fields
The research involved Harvard’s Phoebe DeVries, Google recruiting lead Brendan Meade and Google machine-learning researchers Martin Wattenberg and Fernanda Viégas. The team described the work as computational earth science.
Such collaborations could pair domain knowledge with modern modelling methods. The strongest projects treated seismology as central, rather than assuming a generic AI technique could solve a specialised scientific problem without expert guidance.
Interdisciplinary work made the assumptions visible.
Evaluation Needed Scientific and Practical Metrics
Accuracy could be measured against known aftershock locations, but practical usefulness required more. Did the model improve on existing methods? Did it generalise to unseen earthquakes? Did its maps help emergency planners make better decisions?
False confidence could be costly. A model that missed a high-risk zone or produced excessive warnings could affect evacuation, inspection and public trust.
Evaluation needed to include uncertainty, comparison baselines and expert review.
Communication Was Part of Safety
Risk information could influence anxious communities after a disaster. Visualisations needed to be understandable, while avoiding alarmist or overly reassuring messages.
Scientists and officials had to explain what a forecast did not say. A higher probability of an aftershock was not a prediction of a particular time, magnitude or individual outcome.
Responsible communication protected both public safety and confidence in science.
Open Science Could Improve Resilience
The research was published in Nature, while debates about paywalls and access to machine-learning work were active. Open methods, data where appropriate and reproducible evaluation could help earthquake researchers worldwide build on the result.
Disaster-prone regions with fewer resources should not be excluded from advances funded by global scientific knowledge.
Sharing had to account for safety and data-quality issues, but openness could accelerate validation and adaptation.
AI Did Not Replace Physical Preparedness
Prediction tools could guide attention, but earthquake resilience still depended on building standards, emergency response, public education and infrastructure investment.
An accurate map could not repair an unsafe building or ensure an evacuation route remained open. Technology was one layer in a larger system of preparedness.
This perspective prevented an attractive model from becoming an excuse to neglect proven safety measures.
The Partnership Made a Grounded Contribution
Harvard and Google’s work used AI to answer a defined scientific question with meaningful public value. It showed how patterns across historical events could improve a difficult forecast.
The lasting lesson was that disaster AI had to be calibrated, transparent and embedded in expert practice. The model was most useful when it helped seismologists and emergency teams see risk more clearly—not when it was presented as a substitute for scientific judgment or preparedness.
