Cemil Yavas knows firsthand the devastation that earthquakes can bring to communities, which led him to research how to predict the phenomenon for many years.
“My interest in earthquake prediction research stems from a deeply personal experience,” Yavas said. “I was in Istanbul during the devastating Aug. 17, 1999, earthquake, a disaster that claimed nearly 18,000 lives and left hundreds of thousands homeless. Among those affected were some of my closest friends. The chaos that followed—limited resources, strained rescue efforts, and the haunting sound of people trapped beneath rubble—left a lasting impression on me.”
Cemil Yavas, collaborating with Yiming Ji, Ph.D., Lei Chen, Ph.D., and Christopher Kadlec, Ph.D., faculty and researchers in the Department of Information Technology at the Allen E. Paulson College of Engineering and Computing, has used his personal motivation to build upon previous research and enhance earthquake prediction models.
Using groundbreaking machine-learning algorithms and neural network techniques, they can now predict the highest magnitude of potential earthquakes in Los Angeles from as long as a month out with 97.97% accuracy.
“The inspiration for this research stemmed from the urgent need to improve earthquake prediction accuracy for seismically active urban areas like Los Angeles,” Yavas said. “Earthquakes pose significant risks to densely populated regions, impacting public safety, infrastructure and the economy.”
In a joint 2023 report by the U.S. Geological Survey and the Federal Emergency Management Agency, earthquake damage was estimated to cost the U.S. nearly $14.7 billion annually, with California shouldering nearly two-thirds of that burden, at around $9.6 billion annually.
Yavas and his mentors are optimistic that innovation in prediction techniques, such as machine learning, specifically neural networks — a program that aims to mimic the human decision-making process — can mitigate its destructive effects. Using a layered, connected structure similar to the neurons in our brains, the network can detect patterns, weigh options, and arrive at conclusions.
These technologies may not typically be top of mind when studying and predicting earthquakes. According to Ji, seismology was solely the domain of geography and the other earth sciences until recently.
“At its core, machine learning involves teaching computers to find patterns in data, often revealing insights that may be too subtle or complex for humans to detect easily,” Ji said. “In seismology, this means we can analyze vast amounts of earthquake data—from ground vibrations to atmospheric conditions—and uncover patterns that might hint at when and where future earthquakes are likely to happen.
“By feeding seismic data into machine-learning models, we enable these systems to ‘learn’ from past earthquakes and make predictions about the magnitude and likelihood of future seismic events,” Ji continued.
This isn’t the team’s first time tackling earthquake prediction. They used similar techniques in previous research and achieved an accuracy rate of 69.14%. Subsequently, their work expanded to include other seismically active sites like Istanbul and San Diego, where their results improved drastically. They scored rates of 91.65% and 98.53%, respectively.
Their years of experience and previous attempts only enhanced their current research.
“Our team officially came together to focus on this research over the past year, but the journey began much earlier,” Chen said. “We each have a substantial background in machine learning. Before starting this project, we independently explored methods for earthquake prediction, gaining insights that ultimately contributed to our collaborative approach. This research represents a culmination of those years of groundwork, enriched by the collective expertise we’ve each brought to the table.”
The project, partially funded by the National Science Foundation’s S-STEM program and supported by the Allen E. Paulson College of Engineering and Computing, allowed the foursome to build upon momentum. They adopted a comprehensive data set that included all earthquakes since 2012 and used well-established variables in earthquake prediction.
Eventually, they developed a feature matrix and evaluated 16 different machine-learning and neural network algorithms for their accuracy in determining the highest magnitude of potential earthquakes within 30 days. The “Random Forest” model emerged as the top performer, achieving an accuracy level of 97.97%.
This research was recently published in the multidisciplinary journal Nature.
Though the research is restricted to the Los Angeles area, according to Kadlec, it can potentially improve prediction methods in other areas.
“While we trained our initial model on data specific to Los Angeles, the techniques and methodologies we developed are versatile,” Kadlec said. “With localized data—such as geological, atmospheric, and seismic information from other regions—our model could be adapted and retrained to forecast earthquakes elsewhere.”
He added something indicative of the team’s larger purpose.
“Ultimately, our goal is to make machine learning a central tool in earthquake forecasting and to inspire continued advancements in AI applications for natural disaster preparedness worldwide,” Kadlec said. “This research is just the beginning, and we hope it motivates others to push even further in developing tools to keep communities safe.”