In September 2017, about two minutes before an 8.2 magnitude earthquake struck Mexico City, alarms began to sound, warning local residents of the impending disaster.

These alarms are now present in countries such as the United States, Japan, Turkey, Italy, and Romania, and they have changed our perception of the threat of earthquakes.

Earthquake early warning systems can issue alerts through mobile phones, or send loud signals to the affected areas three to five seconds after the start of a potentially destructive earthquake.

First, seismographs near the fault will detect the occurrence of the earthquake and estimate the magnitude of the quake through a special algorithm.

Advertisement

If it is a medium or large-scale earthquake, the speed of the alarm propagation will be faster than the earthquake itself, so people can be notified in advance by a few seconds to a few minutes.This time window is crucial: in these seemingly brief moments, people can cut off electricity and gas pipelines, drive fire trucks onto the streets, and find safe places.

However, these systems have limitations, and false alarms occur from time to time. More importantly, they only respond to earthquakes that have already occurred, so we cannot predict earthquakes like we predict the weather.

Many earthquake-prone areas will always be in a state of "anxiety and unease." An accurate prediction can allow us to do more risk management work, including shutting down power grids and evacuating residents.

In 2013, when I began pursuing my Ph.D. in seismology, the topic of earthquake prediction was considered unserious, just like searching for the Loch Ness Monster, not belonging to the mainstream field of research.

But just seven years later, the situation has changed a lot. When I started my second postdoctoral research in 2020, I observed that scientists in this field have become more open to the concept of earthquake prediction.At that time, I was involved in a project titled "Tectonic," which was using machine learning to advance the development of earthquake prediction. The European Research Council had full confidence in its potential and thus provided the project with a four-year grant worth 3.4 million euros.

Today, many respected scientists are taking the prospect of earthquake prediction seriously and making progress in their respective sub-disciplines.

Some are studying a type of "slow earthquake" activity along fault lines, which could be a useful indicator that a terrifying destructive earthquake is imminent.

Others hope to find clues from other data, such as seismic noise, animal behavior, and electromagnetic signals, to explore the possibility of issuing warnings before an earthquake begins.

Seismology seems particularly opaque. Astronomers can observe stars, and biologists can observe animals. But those of us who study earthquakes cannot see underground, at least not directly.Instead, we use "substitutes" to understand what happens inside the Earth during crustal movements.

Seismology, which aims to study the seismic waves generated by the motion of the Earth's interior; geodesy, which aims to use tools such as GPS (Global Positioning System) to measure how the Earth's surface changes over time; paleoseismology, which aims to study the traces of past earthquakes hidden in geological strata.

There is still much that we do not know. In the 1960s, decades after the theory of plate tectonics was widely accepted, our understanding of the causes of earthquakes was far from beyond the idea of "stress accumulation to a critical threshold." The idea is that when stress accumulates to a critical threshold, it is released through earthquakes.

Different factors can make faults more likely to reach the threshold. For example, fluids have a huge impact: in the past decade, the injection of wastewater produced in the process of oil and gas production has led to a significant increase in tectonic activity in the central United States.

But when it comes to understanding what is happening on a specific fault line, humans are basically ignorant. We can construct an approximate map of the fault by using seismic waves and mapping the location of earthquakes, but we cannot directly measure the stress experienced by the fault, nor can we quantify the threshold of ground movement.For a long time, the best thing humans could do in terms of prediction was to understand the frequency of earthquakes occurring in specific areas. For example, the last earthquake that ruptured the entire San Andreas Fault in Southern California occurred in 1857.

It is estimated that the average interval for a major earthquake there is between 100 and 180 years. According to rough calculations, the earthquake may already be "overdue." However, as the estimation range shows, the interval between the recurrence of earthquakes can vary greatly and may be misleading.

Our sample size is limited to the scope of human history and the scope we can still see in the geological record. Therefore, the geological record only represents a small part of the earthquakes that have occurred in the history of the Earth.

In 1985, scientists began to install seismographs and other earthquake monitoring equipment in the Parkfield section of the San Andreas Fault in central California.

Compared with earthquakes along other faults, this section has had six earthquakes that occurred at an unusually regular interval, so scientists from the United States Geological Survey (USGS) were confident in predicting that the next earthquake of similar magnitude would occur before 1993.This experiment is largely considered a failure, as the earthquake did not occur until 2004 (of course, we do not wish for earthquakes to happen at any time).

Places including Hawaii have also noticed regular time intervals between earthquakes of similar magnitude, but in fact, they are exceptions rather than the rule.

More commonly, the predicted time intervals for the recurrence of earthquakes are often accompanied by large errors. For regions prone to major earthquakes, the intervals may be around several hundred years, and the errors can span several hundred years as well. Clearly, this method of prediction is far from an exact science.

Tom Heaton, a geophysicist at the California Institute of Technology and a former senior scientist at the United States Geological Survey, is skeptical about whether humans can predict earthquakes.

He largely regards them as random processes, which means we can only calculate the probability of earthquakes occurring, but we cannot accurately predict them.He said: "In terms of physics, this is a chaotic system." Everything behind this is related to the important evidence that proves the behavior of the Earth is orderly and deterministic. However, without understanding what is happening underground, it is impossible to intuitively think of this regularity.

But as scientists increasingly understand what is happening inside the crust, their tools are also becoming more advanced, and we have reason to expect their predictive ability to improve.

Considering that we can hardly quantify what is happening inside the Earth, earthquake prediction has long been considered almost impossible. But at the beginning of the 21st century, two scientific discoveries opened up this possibility.

First, seismologists discovered a strange low-amplitude seismic signal in a tectonic area in southwestern Japan. It would last for several hours to several weeks and occur periodically. This was different from anything they had found before, and it was subsequently called tectonic tremor.

At the same time, geologists studying the Cascadia subduction zone found some evidence that part of the crust sometimes moves slowly in the opposite direction to the normal direction.This phenomenon is known as a slow slip event, occurring in a thin region of the Earth's crust, located beneath the region that generates regular earthquakes. There, higher temperatures and pressures have a greater impact on the behavior of rocks and the way they interact with each other.

Scientists studying the Cascadia subduction zone have also observed the same type of signals found in Japan and determined that they coincide with the timing and location of these slow earthquakes. Thus, a new type of earthquake has been discovered.

Like regular earthquakes, slow earthquakes also redistribute stress within the crust, but they can occur on various timescales ranging from a few seconds to several years. In some cases, such as in Cascadia where they occur frequently, while in other regions they are isolated events.

Scientists then discovered that during slow earthquakes, the risk of regular earthquakes increases, especially in the subduction zone. The fault segments that generate earthquakes are subjected to two types of stress at the same time: regular plate movement and the irregular periodic reverse movement generated by slow earthquakes, which is deeper than the depth at which the earthquake begins.

These elusive slow earthquakes became the subject of my research during my Ph.D., but as is often the case in research work, I did not solve this problem. To this day, we are still unclear about the exact mechanism that drives this activity.Despite this, can we predict regular earthquakes through slow earthquakes? Since their discovery, almost every major earthquake has been followed by several papers indicating that a slow earthquake had occurred before it.

Before the 9-magnitude earthquake in Japan in 2011, there were two slow earthquakes. However, there are exceptions. For instance, despite many attempts to find evidence, we still cannot prove that a slow earthquake occurred before the 2004 Sumatra earthquake in Indonesia. That earthquake caused a devastating tsunami, resulting in the deaths of more than 200,000 people.

More importantly, a slow earthquake does not necessarily lead to a subsequent earthquake. At present, it is unknown whether there is something that can distinguish which slow earthquakes may lead to subsequent earthquakes and which cannot. It may be that in the hours before a major earthquake, some unique process occurs along the fault line.

In the summer of 2022, my former colleague Quentin Bletery and his colleague Jean Mathieu Nocquet published an analysis of crustal deformation data a few hours before 90 major earthquakes.

They found that about two hours before the earthquake, the crust along the fault line began to deform in the direction of the earthquake rupture at a faster rate. This discovery tells us that before the earthquake motion, there was an acceleration process along the fault, sometimes similar to a slow earthquake.He said: "This does indeed support the hypothesis that something has happened before, so we have a method of explanation. But in reality, we cannot predict in the physical world because we do not have the instruments."

In other words, earthquake precursors may exist, but we are currently unable to measure their presence well enough to identify them before an earthquake occurs.

Bouet and La Rocque conducted their research using traditional statistical analysis of GPS data. These data may contain information beyond the scope of our traditional models and reference frames.

Overall, seismologists are now applying machine learning in unprecedented ways. Although it is still early days, machine learning methods can reveal hidden structures and causal relationships that would otherwise appear as a pile of chaotic data.

Earthquake researchers have applied machine learning in various ways. Mustafa Mousavi and Gregory Beroza from Stanford University in the United States have studied how to use it for seismic data from individual seismic stations to predict magnitude, which is very useful for early warning systems and may also help to understand the factors that determine the scale of earthquakes.Harvard University Professor of Earth and Planetary Sciences, Brendan Meade, uses neural networks to predict the location of aftershocks.

Zachary Ross from the California Institute of Technology and his team are using deep learning to extract seismic waves from data, which may enable us to detect more earthquakes even in environments with high noise levels.

During my first postdoctoral research, I met Paul Johnson from the Los Alamos National Laboratory in the United States, who is both my mentor and friend. He is applying machine learning to help understand the seismic data generated by the laboratory.

There are many ways to create earthquakes in the laboratory. A relatively common method is to place rock samples inside a metal frame, subject it to confining pressure, cut the center to simulate a fault, and then use local sensors to measure what happens when the sample deforms.

In 2017, a study from Johnson's laboratory showed that machine learning can very accurately predict how long it will take for a laboratory-made fault to start shaking.Unlike many human methods for predicting earthquakes, this method does not use historical data but relies solely on the vibrations of the fault. Importantly, what researchers considered to be low-amplitude noise is actually the signal that allows machine learning to make predictions.

In this field, Johnson's team applied these findings to seismic data from the Cascadia subduction zone, where they identified a continuous acoustic signal from the subduction zone. This signal corresponds to the speed at which the fault moves during slow earthquake cycles.

He said, "(Machine learning) can let you establish correlations that you never knew existed. In fact, some of them are quite surprising."

Machine learning can also help us create more data for research. Beroza, Mousavi, and researcher Margarita Segou from the British Geological Survey believe that machine learning can help create a more powerful database of earthquakes that have occurred, as the number of earthquakes it identifies in seismic data may be 10 times what we know. Their related paper was published in Nature Communications in 2021.

These improved datasets can help us better understand earthquakes. Johnson said, "You know, in our community, there are many skeptics, and there are good reasons for that. But I think it allows us to see and analyze data and realize what the data contains in ways we never imagined."Although some researchers rely on the latest technology, others choose to look back at history and devise some rather radical research based on animals.

In the more than 10 years I have attended the Geophysical Conference, I have collected many shirts, one of which is printed with a namazu, a huge mysterious catfish that the Japanese believe causes earthquakes by swimming beneath the crust.

This creature is the unofficial mascot of seismology. Before the Great Edo Earthquake in Japan in 1855, a fisherman recorded some unusual activity of catfish in the river.

In a paper published in Nature in 1933, two Japanese seismologists reported that catfish in a sealed glass tank would show increasingly intense signs of restlessness before an earthquake. It is said that the accuracy of predicting earthquakes through this phenomenon is 80%.

Catfish are not the only animals with reactions. Records dating back to 373 BC show that many species, including rats and snakes, left a city in Greece a few days before it was destroyed by an earthquake. Reports indicate that before the 1906 San Francisco earthquake, some horses were whinnying in the early morning, and some even fled the area.The research director of the Max Planck Institute for Ornithology in Germany, Martin Wikelski, and his colleagues have been studying the use of domesticated animals' behavior to help predict the possibility of earthquakes.

In 2016 and 2017, in central Italy, the team installed motion detectors on dogs, cows, and sheep. They thus established a baseline level of animal activity and set a threshold indicating excessive behavior: an increase in movement of 140% or more relative to the baseline level for a period lasting more than 45 minutes.

They found that in 9 earthquakes of magnitude 4 or above, these animals became restless and agitated in advance 8 times, including the 6.6 magnitude earthquake in Norcia in 2016.

Moreover, there were no false alarms; when the animals did not exhibit excessive behavior, no earthquakes occurred. They also found that the closer the animals were to the epicenter, the more their seemingly panicked behavior (like an alarm) would appear earlier.

Wikelski has a hypothesis about this phenomenon: "My view on the whole thing is that the cause may be something in the air. The only thing I can think of is ionized (charged) particles in the air."Electromagnetism is not a strange theory. During or prior to multiple earthquakes, including the 2008 Wenchuan earthquake in China, the 2009 L'Aquila earthquake in Italy, the 2017 Mexico City earthquake, and even the September 2023 earthquake in Morocco, similar aurora-like earthquake lights have been observed.

Scientists at NASA's Ames Research Center, such as Friedemann Freund, have been studying these lights for decades and attribute them to charges activated by fault movement in certain rocks, such as gabbro and basalt. It's like rubbing socks on a carpet, which "releases" electrons (static electricity).

Some researchers propose different mechanisms, while others do not believe that earthquake lights have any connection with earthquakes.

Unfortunately, measuring the electromagnetic fields of the crust or surface is not easy. We do not have instruments that can sample electromagnetic fields over large areas. If we do not know in advance where the earthquake will occur, sampling is difficult or even impossible because we do not know where to install the instruments.

Currently, the most effective method for measuring the underground magnetic field is to set up probes in places with continuous groundwater flow. Scientists have done some work, such as looking for electromagnetic and ionospheric disturbances caused by earthquakes and pre-earthquake activities in satellite data, but the research is still in a very early stage.Some of the biggest paradigm shifts in the scientific community began without any understanding of the underlying mechanisms.

For example, the German geologist Alfred Wegener proposed the theory of continental drift in 1912, which is a fundamental phenomenon at the core of plate tectonics.

His theory was primarily based on the observation that the coastlines of Africa and South America matched, as if they could be pieced together like a jigsaw puzzle. However, it was met with much controversy.

His discovery lacked the crucial elements required by the modern scientific spirit. It wasn't until the 1960s, after evidence of the creation and destruction of the Earth's crust was discovered, that the theory of plate tectonics formally emerged, and the mechanism behind this phenomenon was eventually understood.

In the years from the proposal of the viewpoint to its confirmation, more and more people began to look at the issue from different perspectives. A paradigm shift occurred, and Wegener set the wheels of change in motion.Perhaps the same change has also occurred in earthquake prediction. It may take us several decades to look back on this period of earthquake research with certainty and understand its role in advancing the field.

However, Johnson and others are hopeful. He said, "I do believe this could be the beginning of something, like the plate tectonics revolution, and we might see something similar."