Earthquakes are among the deadliest natural disasters that strike human societies. Major quakes in the last decade alone have killed tens of thousands of people, including an earthquake in 2023 on the Turkey-Syria border that alone resulted in 50,000 deaths. Among the most difficult problems in seismology, the study of earthquakes, is not just dealing with the causes and effects of earthquakes; it’s predicting when they will hit in the first place.
Earthquakes are, broadly, the shaking of the Earth’s surface when built-up energy is suddenly released, which causes seismic waves. Earthquakes primarily happen along geological faults but also can be caused by major geological events like volcanic eruptions and landslides. Because earthquakes are the sudden result of a complex interplay of geological factors, some of which aren’t fully understood, the U.S. Geological Survey (USGS) goes so far as to say on its website in no uncertain terms that earthquakes cannot be predicted.
In an FAQ section that asks “Can you predict earthquakes?” the USGS writes, “No. Neither the USGS nor any other scientists have ever predicted a major earthquake. We do not know how, and we do not expect to know how any time in the foreseeable future. USGS scientists can only calculate the probability that a significant earthquake will occur (shown on our hazard mapping) in a specific area within a certain number of years.”
The agency then goes on to explain that an earthquake prediction must include three elements: the earthquake’s date and time, the location, and the magnitude. Those who claim they can predict earthquakes are wrong for a few reasons, says the USGS, including: their predictions are not based on scientific evidence, they do not define all three elements of a successful prediction, and/or “their predictions are so general that there will always be an earthquake that fits.”
That’s an authoritative statement from the top U.S. geology agency, but that may need an update if a new research breakthrough in earthquake prediction can be replicated consistently at a global level. That’s because researchers at The University of Texas at Austin have used artificial intelligence (AI) to correctly predict 70% of earthquakes over seven months in an area of China just a week before they happened.
While the technique needs to be tested in other regions, it offers a glimpse of a possible solution to a problem previously thought impossible: accurately predicting when an earthquake will occur over a short period of time.
Better data, better predictions
The research team used AI to analyze real-time seismic data over a seven-month trial in areas of southwestern China with substantial seismic activity. This data was provided by something called the AETA system (which stands for “Acoustic and Electromagnetism to AI”), an earthquake monitoring system in China developed by the Earthquake Monitoring and Prediction Technology Research Center at Peking University.
The AETA system is a network of 150 stations equipped with acoustic and electromagnetic sensors, two types of sensors useful for measuring the signals that typically precede earthquakes. Acoustic sensors pick up ground vibrations, while electromagnetic sensors pick up electromagnetic disturbances that may signal seismic activity.
The AETA system represents a vast repository of historical data on earthquake signals and occurrences. The researchers used this historical data to train AI to forecast the location and magnitude of earthquakes that “may occur next week, given the data of the current week,” according to their 2023 paper in the Bulletin of the Seismological Society of America.
They did this by continually extracting 95 features from every 10 minutes of data recorded by the AETA network, and providing that data to their AI. This approach generated a huge amount of data, almost 100,000 features in total each week. The researchers then used a mathematical technique called principal component analysis (PCA) to reduce the number of features, while keeping the most important information provided by the data collection.
Once that was done, the researchers applied their in-house AI models, based on a classic random forest method, to the data. There were two main models, one to predict if an earthquake would occur in the next week and another to predict the magnitude of said earthquake. These models were trained on years of earthquake data, then tested in real time on data being collected. When tested, the system was able to accurately predict 70% of earthquakes happening a week out, correctly forecasting 14 earthquakes over the trial period.
What’s astonishing about the result is that a long-thought impossible problem was solved without any fundamentally new approach.
“Fundamentally speaking, there are actually no new algorithms or techniques invented in our work,” says Yangkang Chen, one of the researchers on the project and the project’s AI lead. “We just use existing mathematical algorithms in a smart way.”
A novel combination of existing factors led to success. The team had access to an incredibly rich dataset in the form of the AETA network’s historical and real-time data collection. They were then able to apply existing, but powerful, AI models to that data. And they creatively used PCA to simplify that data and extract signal from the noise.
“Artificial intelligence and deep learning are game changers in analyzing massive seismic data,” says Sergey Fomel, another researcher on the project. “Although AI’s abilities are sometimes overhyped, this technology has made it possible to solve problems previously considered impossible.”
A competitive element also helped, says Chen. The University of Texas at Austin made the breakthrough as a team participating in an international competition in China, where 600 other designs were also tested.
“We cracked this problem via competition,” Chen says. “Based on the same data and the same criterion to measure the success, we compete by designing approaches to obtain the highest score.” In this case, the “highest score” awarded to competing AI approaches is the prediction accuracy.
Replicating the results
There’s no question that, at least for the area of China in which it was tested, the work of the University of Texas at Austin researchers is a breakthrough. But the next question to answer, for both the researchers and seismologists at large, is: Does this research still represent a breakthrough when applied to other areas around the world?
The team behind the project readily admits in their paper that the method now needs to be tested in other areas to see if it can produce comparable results. That may be easier said than done.
“It is hard to tell how generalizable the early results are,” says Fomel. “We know that progress will depend on rigorous, impartial testing, such as the Chinese contest that our team won. The seismological community must replicate testing using data from other seismically active areas.”
That would require having the right data—and the right amount of it—in other areas, then applying the existing AI to that data. Right now, the data on which the model worked so well is highly specific to the regions of China in which it was tested. Other areas with heavy seismic activity, say California or Japan, don’t have the same pattern of earthquakes, types of earthquakes, or data generated from those earthquakes. The AI system used in China may not work at all in other areas—or may need to be heavily modified to make accurate predictions based on available data.
However, Chen emphasizes the team’s work can and should be tested elsewhere.
“The application to anywhere else is straightforward,” he says. “On the one hand, we leverage all existing data in the target area, apply our in-house prediction technique, and perform a prediction.” On the other hand, he notes, researchers could also attempt to optimally mimic what his team did in China, deploying the same type of stations to measure new data and make predictions based on the same methodology.
“It is just a matter of funding and support,” he says.
But that also could present problems. Chen says this type of fundamental research is done in the spare time of the seismologists who undertake it. It’s often risky, so it is extremely difficult to get funding specifically for this type of thing, which is why this particular research resulted from a competition.
So it may still be a while before the USGS updates that statement on its website. In fact, the agency responded to a request for clarification about its statement in light of the new research, saying:
“The work recently published by Saad et al. in the Bulletin of the Seismological Society of America is a proper first step in proposing and vetting a possible earthquake prediction technique. This publication now allows other researchers to examine their technique, critique it, and/or improve it. And as the others themselves wrote in their article, it needs to be tested in other regions. It is too soon to say whether this is a breakthrough or a promising result that fails to lead to an actual, useful, technique for reducing earthquake risks and losses.”
However, Fomel is hopeful.
“It is good to be cautious and state that all predictions can be only probabilistic, similar to extreme weather events,” he says. “Our results show that making short-term earthquake probability estimates less random might finally be possible.”
Further Reading:
Saad, O.M., et al., Earthquake Forecasting Using Big Data and Artificial Intelligence: A 30‐Week Real‐Time Case Study in China, GeoScienceWorld, Sept, 5, 2023, https://pubs.geoscienceworld.org/ssa/bssa/article-abstract/113/6/2461/627949/Earthquake-Forecasting-Using-Big-Data-and
University of Texas at Austin, Artificial Intelligence Predicts Earthquakes With Unprecedented Accuracy, SciTechDaily, Aug. 20, 2023, https://scitechdaily.com/artificial-intelligence-predicts-earthquakes-with-unprecedented-accuracy/
U.S. Geological Survey, Can you predict earthquakes?, https://www.usgs.gov/faqs/can-you-predict-earthquakes
Logan Kugler is a freelance technology writer based in Tampa, Florida. He is a regular contributor to CACM and has written for nearly 100 major publications.