Earthquakes are usually seen as rare events — felt only occasionally, and at random.   Scientists who study earthquakes and the movement of the Earth’s subsurface see them differently. In Southern California, where major population areas rest atop the highly active San Andreas Fault, Zachary E. Ross, assistant professor of geophysics and William H. Hurt Scholar at the California Institute of Technology, uses a novel AI model to study the nonstop chatter of small earthquakes.  

While most small earthquakes go unnoticed by the people living near them, a vast network of sensitive seismic monitoring equipment that spans California’s faults is constantly listening for and recording these tiny movements.   At a recent lecture held by the Distinctive Voices program of the National Academy of Sciences, Ross explained why small earthquakes hold major potential for geophysical research. “They are filling in the gaps between all the larger ones, and basically tell you a much more complete picture about how these sequences are evolving in space and in time.”  

“By reprocessing data with very sensitive techniques,” powered by AI, Ross and his research team have been able to confirm that “something like 2 million [seismic] events have happened in Southern California alone over the last decade, and they’re happening pretty much all the time.”

The long-standing challenge that scientists face when studying small earthquakes is that they present an enormous amount of data that is difficult and time-consuming to process, sort through, and understand. Legacy computer systems that process seismic data are taught to recognize what is an earthquake and what is not — but the “blips of energy” that usually define small earthquakes look very similar for vehicles, air traffic, and blasts made in quarries to these computer systems, Ross explained, which leads to false detection and limits researchers’ ability to make sense of large seismic datasets. “Since the late 1970s, and up until a handful of years ago, the core technology for doing this has not really changed much,” he said.  

Over the last decade, Ross and his colleagues started looking to advances in AI to help them overcome these hurdles, and ultimately turned to breakthroughs in deep learning — a type of machine learning that harnesses neural networks designed to “duplicate the decision-making of what humans can do,” he explained. Some commercial deep learning models can detect and name objects in photos, for example — a task that is easy for humans, but which can be done at far larger and faster scale by a neural network trained to do so. Ross’ model works much the same way as AI image detection, parsing the subtle differences between small earthquakes and other energy blips and making split-second decisions to categorize each event, enabling it to sift through enormous seismic data sets efficiently.  

One of the factors that makes AI a good fit for this area of research, Ross said, is that “seismologists have been collecting this data for a century,” often hand labeling which blips are an earthquake and which are not. “Today, we have millions of measurements that have all been scrutinized by hand.” A large amount of labeled data is a crucial ingredient for success in deep learning, as it gives the model ample information and examples to study and learn from.  

Ross’ new model has revealed previously unseen trends in small earthquake data, allowing his team to explore seismic events and faults in a new way. Earthquakes are really not random, he said. “There are very rich spatial and temporal patterns that demonstrate interactions between these events, so thinking about them as just one event is not necessarily the most complete picture of the physical processes that drive earthquakes.”

The Four-Year Earthquake

In 2016, residents of Mount San Jacinto, an area east of Los Angeles, reached out to Ross’ team about a series of small, repeated earthquakes, known to scientists now as the Cahuilla Swarm. “We first learned about this from somebody that lived around here and was feeling all these tiny events,” Ross recounted, “and they wanted to know what was going on, so we looked into this.”  

Ross put his AI model to work visualizing how the earthquake “swarm” had changed and moved both through time and through the Earth — distilling thousands of events and revealing that they all took place within “a single fault surface at depth,” Ross said. “It turns out that [the swarm] started from a spot about 5 miles below the surface that was something like 100 meters, and grew to something 30 times that size over four years. So, it’s moving very slowly through this fault system.”  

Through this new way of understanding the event, Ross and his team both recorded and discovered “a new type of slow-growing earthquake swarm,” he said. They also developed a theory to explain why the Cahuilla quakes unfolded as they did. “We think that there are basically pockets of hydrothermal water … that are trapped in the Earth,” he said. “Initially, this fluid is essentially sealed off” — but, over time, “the pressure is building up. Eventually, this seal breaks, and if there’s a fault nearby, this fluid now will start to enter the fault,” triggering many small earthquakes.

Improving Earthquake Forecasting

In 2020 near Westmorland, California, several magnitude 4 earthquakes — large enough to be felt by residents — hit over four days. Ross and his team examined data collected by specialized GPS sensors around the region. “These are much more sensitive than the ones you have in your phone,” Ross said. “They’re able to measure essentially deformation of the ground at basically the millimeter level or smaller.”  

Feeding this data into their AI model and expanding their scope to the days before the earthquakes began, they found “the ground actually started to deform about 12 hours before the earthquake started.”  

The model helped them understand that this small, early movement came from the same fault that later produced noticeable shaking, but at a rate too slow to radiate seismic energy, something Ross calls “slow creep.” The slow creep started 12 hours before the larger event in each of the three earthquakes they modeled — making this “silent signal” a candidate for further study as scientists seek to develop better earthquake forecasting abilities, Ross said.

“Of course, none of this tells you anything about the magnitudes of events. We’re not even that far along yet in terms of our abilities to model things,” Ross cautions. “But this at least is telling you something about expected rates and so forth, and how that can be tied to something that is observable.”  

“Tiny earthquakes have value,” Ross said. “It is a good tool for seismologists to better understand what is happening — to help us better inform our decisions and to provide information to the public.”

Related Resources

• Video recording of the lecture

• Distinctive Voices program

• Ross Lab at Caltech