By refining an artificial intelligence approach to predicting earthquakes in the laboratory, or labquakes, engineers at Penn State are paving the way to one day help forecast natural earthquakes.

“We are a long way off from predicting natural earthquakes, but understanding the physics of labquakes and how they evolve allows us to better understand the mechanics of real earthquakes,” said Parisa Shokouhi, Penn State professor of engineering science and of acoustics and corresponding author of the work published in Scientific Reports.

“We can study the precise conditions under which a gently creeping fault suddenly becomes unstable and triggers an earthquake, such as the amount of stress, the roughness of the fault or the role of small loose rock particles at the interface, to name a few possibilities.”

To monitor earthquakes in the field, instruments are placed at the Earth’s surface, far away from the depth where earthquakes typically occur, which Shokouhi said forces scientists to make simplified assumptions. Labquakes, on the other hand, are produced under tightly controlled conditions, allowing scientists to make detailed measurements on every aspect of the experiment.

Researchers create labquakes by sliding together blocks of rock—known as friction experiments—to generate the laboratory equivalent of earthquakes, or stick-slips, that they then monitor with ultrasonic transducers.

The team developed a machine learning model for labquake prediction that can also automatically retrieve specific parameters—known as rate and state friction parameters—from the ultrasonic monitoring of stick-slip experiments. The rate and state friction parameters define the mechanics of the labquakes; they determine the strength of the fault, signaling how close it is to failure.

To estimate these parameters, the team developed a physics-informed neural network (PINN) model—a modified machine learning algorithm that incorporates the rate and state friction law—to predict when the experimental fault might fail and produce a labquake.

The PINN model has the same or better accuracy as networks that do not incorporate the rate and state friction law, as well as the ability to predict labquakes further into the future. This is because, according to the researchers, the broader understanding of physics informs a wider interpretation than limiting it to the specifics of a particular experimental set up.

“We show that PINN models provide accurate predictions with a smaller amount of training data and that transfer learning—when trained models are applied to a new, related task—is greatly enhanced in these models,” said co-author Jacques Rivière, assistant professor of engineering science and mechanics at Penn State.

“That provides a very nice connection to what is essentially the million-dollar question for labquake prediction: How can it be extended to the prediction of real earthquakes? Physics-informed neural networks and transfer learning are likely to be major factors in developing models that can move us toward earthquake prediction.”

The ultimate goal, researchers said, is to develop similar models based on these methods in order to predict earthquakes in the field.

To develop and train the physics-informed neural networks, Prabhav Borate, a graduate student in engineering science, used labquake data collected in the Rock Mechanics Laboratory of co-author Chris Marone, professor of geosciences in the College of Earth and Mineral Sciences.

“We created the PINN models by training them to follow the rate and state friction laws,” Borate said. “This was achieved by designing the model to penalize itself whenever the predictions didn’t match the law. This approach proved effective in accurate prediction of labquakes using smaller datasets while providing invaluable information about the earthquake mechanics through the extracted friction parameters.”

This paper builds on previous work by the same research group, which was published in Nature Communications in June 2023.