Seismologists are making significant progress in using machine learning to enhance earthquake forecasts, with three new studies demonstrating the effectiveness of deep-learning models in comparison to conventional approaches.
These advancements are preliminary but offer hope for improving earthquake predictions in specific scenarios, such as assessing the risk of aftershocks following a significant earthquake.
Earthquake forecasts do not provide precise predictions of when, where, and with what magnitude an earthquake will occur, as such predictions have proven unreliable. Instead, seismologists focus on statistical analyses to understand broader trends, like predicting the likelihood of aftershocks in the days to weeks after a major earthquake.
Machine learning seems well-suited to improving earthquake forecasts, as it excels at analyzing vast datasets and recognizing patterns. Seismology benefits from extensive global earthquake data. However, extracting meaningful patterns from earthquake data has been challenging due to the rarity of large earthquakes.
In recent years, machine learning has aided in identifying previously unnoticed small earthquakes in seismic records, expanding earthquake catalogs and enabling further machine-learning analysis.
The three new studies employ neural-network-based approaches to improve earthquake forecasting compared to conventional models. One study focused on southern California earthquake data from 2008 to 2021, another on earthquakes in central Italy in 2016-17, and the third on Japanese earthquake data spanning 30 years. These models outperformed traditional methods in forecasting earthquake occurrences and magnitudes.
While these models represent a promising step forward, they are not considered breakthroughs in their current form. However, they show potential for incorporating machine learning into daily earthquake forecasting. As earthquake datasets continue to grow, machine learning’s capacity to work with extensive data makes it a valuable tool for the field.
Seismologists anticipate that agencies like the US Geological Survey (USGS) will begin using machine-learning models alongside traditional ones, potentially transitioning to machine learning entirely if it proves superior. This could lead to improved forecasts in situations like aftershocks, which can disrupt lives for extended periods.
Despite these advancements, experts emphasize the importance of earthquake preparedness and building safety, as machine learning models are not a substitute for proper precautions.