AI is changing our understanding of earthquakes

When the biggest earthquake in more than a decade rattled Russia’s remote Kamchatka Peninsula in July, seismologists around the world knew within moments. For earthquakes big or small, sensors around the globe detect the tremors and relay that information to researchers, who quickly analyze the observations and issue alerts.

Now artificial intelligence is poised to make almost everything about earthquake research much faster — and to rewrite researchers’ very understanding of how earthquakes happen.

By using a subfield of AI called machine learning, some scientists are identifying up to millions of tiny, previously unseen earthquakes in data gathered from seismically active places. These new and improved databases are helping researchers to better understand the geological faults along which quakes happen, and can help to illuminate the risks of future quakes. Some scientists are even using machine learning to improve their forecasts of how many aftershocks may rattle a location that has just experienced a large and damaging earthquake.

More broadly, researchers hope that machine learning, with its ability to crunch through huge amounts of information and learn from the patterns within, will reveal fresh insights into some of the biggest mysteries about earthquakes, including how a quake unfolds in its first devastating seconds.

“Machine learning opened a whole new window,” says Mostafa Mousavi, a seismologist at Harvard University.

Shaking earth, exploding data

Earthquakes happen when geological stress builds up in the ground, such as when two plates of Earth’s crust grind alongside one another, as they do at California’s San Andreas Fault. At some point, the stress reaches a critical threshold and the fault ruptures, breaking the rock and causing seismic energy to ripple outward and shake the ground.

That energy is recorded by seismometers and other instruments around the world, which are positioned in great numbers in geologically active areas like California and Japan. The data feed into national and international systems for tracking earthquakes and alerting the world. The amount of data has exploded in recent years as seismologists find new ways to gather information on ground movements — like detecting seismic signals over fiber optic networks, or using the accelerometers built into smartphones to create a phone-based earthquake warning network.

Just a decade or two ago, much of the analysis of seismic signals was done by hand, with scientists working as quickly as possible to assess recordings coming in from their observing networks. But today, there are just too many data points. “Now the only — almost — way that you can deal with the seismic data is to go to automatic processing,” says Mousavi, who coauthored a 2023 article in the Annual Review of Earth and Planetary Sciences on machine learning in earthquake seismology.

One of the most common uses of machine learning in seismology is measuring the arrival time of seismic waves at a particular location, a process known as phase picking. Earthquakes generate two kinds of seismic waves, known as P and S waves, that affect the ground in different ways and show up as different types of squiggles on a seismogram. In the past, a seismologist would analyze data arriving from seismic sensors and hand-select what they gauged to be the start of P waves or S waves on those seismogram plots. Picking the starts of those waves accurately and precisely is important for understanding factors such as where exactly the earthquake hit. But phase picking is very time consuming.

In the past few years, seismologists have been using machine learning algorithms to pick seismic phases much faster than a human can. There are a number of automated methods that can do phase picking, but machine learning algorithms, which have been trained on huge volumes of data on past quakes, can identify a wide variety of signals from different types of tremors in a way that was not possible before. The practice is now so standard that the term “machine learning” is no longer stated in the titles of research papers, says Mousavi. “By default, everybody knows.”

AI-based phase picking is faster than phase picking by humans and at least as accurate, Mousavi says. Seismologists are now working to expand these tools to other types of seismic analysis.

Expanding quake catalogs

One area that has already seen big discoveries is the use of machine learning to expand earthquake catalogs — basically, lists of what earthquakes happened where in a particular region. Earthquake catalogs include all the quakes that seismologists can identify from recorded signals — but AI can find exponentially more tremors than human scientists can.

Essentially, machine learning can trawl through the data to identify small earthquakes that people don’t have the ability or time to flag. “Either you don’t see them by eye, or there’s no time to go and look at all those tiny events,” says Leila Mizrahi, a seismologist with the Swiss Seismological Service at ETH Zürich. Often, these tremors are obscured by background noise in the data.

In a pioneering 2019 study in Science, researchers used an AI algorithm that matched patterns of seismic waves to identify more than 1.5 million tiny earthquakes that happened in Southern California between 2008 and 2017 but had not been spotted before. These are itty-bitty quakes that most people wouldn’t feel even if they were standing on top of them. But knowing they exist is important in helping seismologists understand patterns of behavior along a geological fault.

In particular, Mousavi says, tiny earthquakes are important as a window into how larger earthquakes begin. Large earthquakes may happen along a particular fault once every century or more — far too long a time period for scientists to observe in order to understand the rupture process. Tiny quakes behave much the same as big ones, but they happen much more frequently. So studying the pattern of tiny quakes in the newly expanded earthquake catalogs could help scientists better understand what gets everything going. In this way, the richer catalogs “have potential to help us to understand and to model better the seismic hazard,” Mousavi says.

Expanded earthquake catalogs can also illuminate the structure of geological faults below a region much better than before. It’s like going from a simplistic sketch of how the faults are arranged to a painting with more photorealistic details. In 2022, a team led by seismologist Yongsoo Park, then at Stanford University, used machine learning to build an expanded catalog of quakes in Oklahoma and Kansas between 2010 and 2019, many of them induced by oil and gas companies injecting wastewater into the ground. The work illuminated fault structures that weren’t visible before, allowing the scientists to map the faults more precisely and to better understand seismic risk.

Park and his colleagues showed that 80 percent of the larger earthquakes that happened could have been anticipated based on the smaller earthquakes that occurred before the big ones. “There is always a possibility that the next major earthquake can occur on a fault that is still not mapped,” says Park, who is now at Los Alamos National Laboratory in New Mexico. “Routinely capturing smaller earthquakes might be able to reveal such hidden faults before a major earthquake happens.”

Scientists are applying this approach around the globe. Researchers in Taiwan, for instance, recently used machine learning to produce a more detailed catalog of a magnitude 7.3 tremor in April 2024 that killed at least 18 people on the island and damaged hundreds of buildings. The study, reported at a seismology meeting in April 2025, found the AI-based catalog to be about five times more complete than the one produced by human analysts, and it was made within a day rather than taking months. It revealed new details on the location and orientation of geological faults — information that can help officials better prepare for how the ground might move in future quakes. Such catalogs “will become the standard in every earthquake-prone region in the future,” says team leader and seismologist Hsin-Hua Huang of Academia Sinica in Taiwan.

Forecasting is still a problem

So far, AI hasn’t been as successful in tackling another of seismology’s biggest challenges — forecasting the probability of future quakes.

The field of earthquake forecasting deals with general probabilities — such as the chances of a quake of magnitude X happening in region Y over time period Z. Currently, seismologists create quake forecasts using mathematical analyses of past earthquakes, such as a statistical method that relies on observations of how past earthquakes triggered subsequent quakes. This approach works well enough for specific tasks, like understanding how many aftershocks may rattle a region after a Big One. That sort of information can help people in a disaster zone know whether it’s safe to return to their houses or whether more aftershocks might be on the way, threatening to collapse more buildings.

But this kind of analysis can’t always accurately capture the real seismic risk, especially along faults that only rarely yield big quakes and thus aren’t well represented in the seismic record. Seismologists are testing AI-based algorithms for earthquake forecasting to see if they might do better, but so far, the news is tepid. In their best performances, the machine learning analyses are about as good as the standard methods of quake forecasting. “They are not outperforming the traditional ones yet,” says Mousavi, who summarized the state of the field in an August 2025 article in Physics Today.

In one of the more promising experiments, Mizrahi has been trying to use AI to speed up producing aftershock forecasts in the crucial minutes and hours after a large earthquake hits. She and a colleague trained a machine-learning algorithm on the older statistical method of quake forecasting, then unleashed it on its own to see how the AI would do. It did perform much faster than the older, non-AI approach, but there’s still more work to do. “We’re in the process of evaluating how happy we are with it,” says Mizrahi, who published the findings last year in Seismological Research Letters.

In the future, researchers hope to speed up these types of forecasting analyses. Other areas of seismology could eventually benefit, too. Some early research hints that machine learning could be used in earthquake early warning, for instance estimating exactly how much the ground will move in the seconds after an earthquake has started nearby. But the usefulness of this is limited to the few parts of the world that have early warning systems in place, like California and Japan.

Park also cautions about relying too much on machine learning tools. Scientists need to be careful about maintaining quality control so they can be sure they are interpreting the results of any AI analysis correctly, he says.

Overall, though, seismologists see a bright future in using AI to understand earthquakes better. “We’re on the way,” Mizrahi says.