Few months go by without another devastating earthquake somewhere in the world reminding us how we all remain at the mercy of major seismic events that strike without warning. But a new branch of geophysics powered by machine learning is uncovering fresh insights into the earth’s slipping faults that often trigger these catastrophic earthquakes.
Machine learning, which often goes by the catchier moniker of artificial intelligence, has captured the public’s imagination with its promises of fully autonomous cars and the approaching “singularity” when machines out-think people. The current state of the art, however, shows little signs of true intelligence, such as the ability to abstract the principles behind a given phenomenon. In image recognition, AI systems learn from rote memorization to identify objects and are, therefore, often fooled. For these reasons, machine learning remains a more appropriate term for this branch of computational science.
Many of the recent headline-grabbing developments in machine learning hinge on an approach called deep neural networks. Yet a simpler and more transparent form of machine learning called decision trees is unlocking impressive new scientific discoveries. In the case of our earthquake research at Los Alamos National Laboratory, a machine-learning process involving decision trees has revealed previously unsuspected physics principles that a deep neural network would have obscured and humans poring over data sets probably never would have noticed. To our surprise—and delight—this approach has led to a breakthrough in probing the mechanics of earthquakes, which will certainly advance our pursuit of the holy grail of geoscience: earthquake forecasting.
Machine learning is not magic. It does, however, bring an unprecedented ability to sift through vast streams of information. Deep neural networks have proved capable at challenging tasks like facial recognition, but they also are constrained by their black-box nature. Neural networks might give the right answer, but you won’t know how they got there. They perform computations on the data that cannot be examined from the outside by the scientist—if you want Siri to find the name of a song, you don’t care how the machine finds it. If you wish to understand some physical process, you certainly care. You can’t build a theory based on digital thin air.
In the machine-learning approach we use, decision trees establish a set of questions about the statistical aspects of the information contained in an earthquake’s acoustic signal. Based on one decision, the machine-learning program branches to another decision, and so on—a diagram of the process ends up looking like a tree, thus the name. And this kind of machine learning program is not opaque. Like good high school math students, the decision trees “show their work” every step of the way. That lets us examine each decision in a whole forest of decisions to see why and how it was made.
In geoscience and beyond, machine learning is changing the nature of scientific inquiry, as big-data crunching continues its fitful march across the disciplines. Traditionally, science proceeds from formulating a hypothesis to conducting an experiment, analyzing the data, refining the hypothesis and experimenting again. While stunningly successful, this method is vulnerable to bias—a biased or simply misguided hypothesis will compromise the results and conclusions of a research project.
In our team’s work, we minimize our preconceived bias by letting the machine sift through massive seismic data sets seeking connections between the data and some important statistical aspect of a slipping geologic fault, such as friction or fault displacement. During this procedure the machine-learning program builds itself. Then we test those connections by mapping them onto data the program has never seen.
We started by analyzing the acoustic signals from a laboratory earthquake machine at Penn State University. Between artificial earthquake events, the machine learning discovered a continuous signal that had previously been dismissed as useless noise and ignored for years by all, including me. Remarkably, no matter when the signal is read, it lets us diagnose the condition of the fault and reveals the time remaining before the fault slips in our laboratory earthquake experiments.
After we detected the telltale signal in the lab, we used the same machine-learning technique to study real-world seismic data from “slow slip” events in major earthquake zones in North America, South America and New Zealand. Slow slip, which can be imperceptible to humans, is the gradual shifting of the earth adjacent to the locked zone of a fault. As pressure builds in the locked zone, friction eventually yields to a mega-earthquake.
We found that these geologically active areas continuously broadcast the same signal through a slow-slip cycle. The signal diagnoses precisely, at any instant in time, the fault movement in that area and how soon it might slip. That information may allow us to determine how and when a slow slip leads to a mega-earthquake.
In the future, we hope to demonstrate that the basic physics beneath this phenomenon applies to any kind of material failure, whether it is an avalanche or a fracturing steel beam on a bridge. They may all emit a characteristic acoustic signal revealing their current state and how soon they will fail.
Machine learning might not magically answer every scientific question, but it gives us a tremendously powerful and flexible addition to our tool kit. We had little chance of discovering this revelatory acoustic signal without machine learning. It has opened up a revolutionary way of “hearing” the earth and uncovered a new natural law. It may enable geophysicists to take significant steps toward earthquake forecasting.
Similar breakthroughs will occur in other scientific fields. We are still feeling our way through the early stages of machine-learning–powered science. We can hardly imagine where it will take us in the next few years, but it surely will surprise us.
The author works with a team that includes Bertrand Rouet-Leduc, Claudia Hulbert and Ian McBrearty at Los Alamos. Collaborator Chris Marone operates the earthquake machine at Penn State. The team’s recent papers on machine learning and earthquake behavior can be found here.