Is it really impossible to predict earthquakes? The "whispers" just before an earthquake captured by AI ─ Kyoto University's challenge to the "world a few seconds before"

Can AI Detect Earthquakes "Seconds Before"?

—Kyoto University's Visualization of "Precursors" and the Hope and Caution That Lie Ahead

"Being able to predict earthquakes in advance"—this is a dream humanity has held for many years.
However, in reality, many signs previously considered as "precursors," such as abnormal animal behavior and "earthquake clouds," lack scientific backing. Current seismology believes that "major earthquakes occur with almost no warning."PhysOrg

Challenging this conventional wisdom is the latest research announced by a team from Kyoto University in November 2025. They successfully detected extremely weak signals that appear only just before an "earthquake in the laboratory" by using a large-scale rock experiment apparatus and feeding the data into machine learning.PhysOrg

Experiments with Meter-Scale "Artificial Faults"

The stage for the research is a meter-sized rock testing machine designed to replicate the behavior of real faults. Rock blocks are pressed together strongly and slowly shifted to create a state where they are less likely to slip due to friction. Eventually, when the limit is exceeded, the rock suddenly slips, generating vibrations. This is the "stick-slip earthquake" recreated in the laboratory.PhysOrg

The apparatus is equipped with numerous high-sensitivity sensors that record rock deformation, acoustic emissions, and changes in shear stress with high temporal resolution. The research applied machine learning to explore "where in the data characteristic changes are hidden in the seconds to tens of seconds before an earthquake occurs."PhysOrg

The Key is Stress Changes in the "Quietly Slipping Area"

Previously, it had been reported that machine learning could accurately predict earthquake timing in small centimeter-scale experiments. However, natural faults are much larger and structurally complex. Therefore, there was significant doubt about whether predictions could be made similarly on a larger experimental scale.PhysOrg

The Kyoto University team compared experimental data with numerical simulations of physical models to investigate what machine learning was using as "clues." They found that changes in shear stress in the "creep area," which slowly slips, rather than the average stress of the entire fault, serve as crucial signs just before failure.PhysOrg

In other words, AI is detecting the gradual approach of stress to its limit in areas that appear quiet on the surface—such an image is emerging. This insight is shifting the perspective of earthquake research from "viewing faults as a whole" to focusing on "where and how they are slipping," emphasizing spatial disparities.PhysOrg

Why We Can't Say "Earthquake Prediction is Achieved"

It's important to note that this achievement does not mean "we can now know where a major earthquake will occur tomorrow."

• The experiment was conducted within an artificial rock apparatus

• The scale is meter-level, still small compared to actual faults

• It's not guaranteed that the same subtle signals can be observed in real faults tens of kilometers underground

There are many hurdles, and researchers themselves position this as "an important step in deepening the physical understanding towards short-term earthquake prediction," not as "immediate application."PhysOrg

Nevertheless, the fact that "the immediate physical processes can be explained by both AI and physical models" is significant. By linking the often "black box" reasoning of machine learning to physical mechanisms, discussions can proceed in a way that is easier for seismologists to accept.PhysOrg

Voices of "Expectation" and "Caution" on Social Media (Anticipated Reactions)

When this news is reported, various reactions are likely to fly around on social media. Here, we organize some typically imagined responses (these are hypothetical comments to indicate trends, not actual posts).

1. The Positive Group Speaking of Hope

• "AI is making earthquake prediction more realistic...! Proud that it's a Japanese study."

• "Even if it's just tens of seconds before, knowing could improve the precision of stopping elevators and slowing down bullet trains."

For those with the earnest desire to "know even a second earlier," stemming from experiences like the Great Hanshin-Awaji Earthquake and the Great East Japan Earthquake, this research is seen as a strong beacon of hope. Particularly, the fact that a Japanese university is leading the effort adds to the expectation of fulfilling the responsibility as a disaster prevention nation.

2. The Cautious Group Wary of "Sensational Reporting"

On the other hand, such cautious voices are also imagined.

• "The headline 'AI Predicts Earthquakes' might start spreading on its own again."

• "I hope they clearly convey that the lab story and real earthquakes are completely different."

Given the sensitive theme that could directly connect to disaster prevention information from the government and municipalities, many are likely wary of "headlines that excessively raise expectations" spreading. While the keyword "AI" increases topicality, its magical image can also easily lead to misunderstandings.

3. Perspectives from On-Site Experts and Engineers

Among seismologists and disaster prevention engineers, more technical discussions are likely to heat up.

• "What kind of sensor network is needed to capture the same type of acoustic emissions in the real field?"

• "Can't we test similar models using data from existing seismic networks?"

• "It's good that they aligned with the physical model. Next, how to connect it at a multi-scale level."

Here, rather than "buzz," the practical perspective of "how to implement it into real observation networks and infrastructure" comes to the forefront.

4. Discomfort with "AI Omnipotence"

Additionally, alongside expectations for AI, voices expressing discomfort with treating it as an "omnipotent tool" are also imagined.

• "It's not that AI is amazing, but that the achievement is due to proper experimental data and physical models."

• "AI is an 'assistant to find hints,' and understanding earthquakes is the job of humans."

This research is like a "concert of AI and physics." The point is that machine learning does not run independently but explains "why it works" by comparing it with physical simulations. Many comments are likely to appreciate this sense of balance.

The Future This Research Opens Up

So, what path does this research lead to in the future?

1. Application to Observational Data
   First, it is expected that verification will proceed on whether the same method can be applied to observational data of natural earthquakes. In addition to existing seismic networks, combining more sensitive sensors and borehole observations will be key to detecting "precursor patterns seen in the lab" in reality.

2. Advanced Utilization of "Seconds to Tens of Seconds Before"
   Even if short-term precursors can be detected, it is likely that predictions will be on a very short time scale, such as "seconds to tens of seconds before." However, even in that brief time, there are many scenarios where it can be utilized on the side of social infrastructure, such as automatic stopping of bullet trains, emergency stopping of factory equipment, and responses during surgeries in hospitals.

3. Integration with Earthquake Risk Assessment
   In the future, a framework that combines long-term earthquake probability models with short-term precursor detection to multilayeredly evaluate "when, where, and how much damage might occur" will likely be explored. In doing so, the policy of "not using AI output as a black box" from this study will become important.
   
   
   

What We Should Know Now

Finally, let's organize what we can do "from now on" after reading this article.

• Not Jumping to Conclusions That "Earthquake Prediction is Achieved"
  The current achievement is at the stage of advancing the understanding of "precursor signals" at the laboratory level.

• Still, Correctly Evaluating the Progress of Research
  Combining AI and physical models to better understand fault rupture processes is a definite advancement and an important step towards future disaster prevention technology.PhysOrg

• Continuing to "Update" Disaster Prevention Actions
  No matter how advanced prediction technology becomes, the importance of basic preparations such as "securing furniture," "preparing emergency bags," and "confirming contact methods with family" remains unchanged. Rather, it is necessary to update our disaster prevention actions in line with technological advancements.

The era when AI can hear the "whispers" of earthquakes might be just around the corner. However, it is still humans who decide how to interpret those voices and how to incorporate them into societal systems. The achievement of Kyoto University can be said to be a promising material to advance that discussion one step further.