Scientists have spent years trying to predict exactly when and where a major earthquake will occur and how big it will be. Although research has been insufficient so far, there are promising developments.
A geophysicist at Los Alamos National Laboratory in New Mexico, Dr. Paul Johnson says his team is leading the way in developing a tool that can make earthquake prediction a dream come true. As with many scientific initiatives today, the research team’s approach is based on artificial intelligence (AI) in the form of machine learning (ML).
But the use of machine learning for seismologists is still in its infancy. The lack of quantitative data (such as earthquake magnitudes, shaking intensities) for hold-release earthquakes, which are the most common earthquake events, adds to the difficulty.
Major earthquakes are caused by the movement of geological faults at or near the boundaries between Earth’s tectonic plates, and researchers are basically looking for data at these points. But for hold-release earthquakes, the process before destructive slip takes a very long time, and there is little movement on a fault as strain builds up. To properly study earthquakes, researchers need to document the moment they occur. Due to the difficulty of this, the dataset remains quite limited.
Dr. Johnson, on the other hand, turns to a different type of seismic activity: slow-slip earthquakes. Similarly, these events, which are caused by the movement of tectonic plates, spread over hours, days and even weeks, as opposed to seconds in hold-release events. The slowness of these events can be a great treasure for researchers. From these long processes, it was possible to generate a set of data points that could better train the neural network to predict seismic activity.
The research team’s machine learning system demonstrated predictive capabilities in the Pacific Northwest’s Cascadia Subduction Zone. Listening to 12 years of seismic audio recordings emanating from slow fault movements, the system was able to look for patterns to reflect (reconstruct) past slow slide events based on the seismic signals that preceded them. With this projection, it was shown that the team could predict what would happen a week or so later.
Dr. While Johnson’s work shows that machine learning techniques can indeed be used in seismic events (slow slip), the data gap will need to be compensated to extend this prediction to earthquakes (hold-release). To address this shortcoming, the researchers simulated miniature earthquakes to mimic hold-release events in a laboratory. Using the collected data, a fine-tuned numerical simulation of the laboratory earthquake was created and then combined with data from real events.
The result is an efficient machine learning model that is effective in predicting when a laboratory earthquake will occur.
Dr. Johnson’s team plans to apply earthquake predictions to a real geological fault, possibly the San Andreas fault, in the future. The combination of data from the numerical simulation of the fault and data from real earthquakes will be used to train ML systems.
More research and work will be required to see how accurately the model can be used to predict seismic events not included in the training data. But if all goes well, seismologists may soon have enough accurate tools to predict earthquakes.
20360
