Researchers at Kyoto University have made significant strides in earthquake prediction by employing machine learning techniques to identify subtle signals that may precede seismic activity. This study, published in the journal Nature Communications on November 19, 2025, explores how advanced algorithms can analyze data from laboratory experiments to better understand fault behavior before earthquakes occur.
For years, predicting earthquakes has remained an elusive goal for scientists. Traditional methods have often relied on anecdotal evidence, such as unusual animal behavior, which lacks empirical support. As a result, seismologists generally conclude that earthquakes happen with little to no warning. Despite this, the team at Kyoto University sought to investigate whether machine learning could provide insights into the precursory phases of seismic activity, particularly noting that faults undergo processes like “micro-fracturing” and “slow slip” before a major quake.
In their research, the team conducted experiments simulating meter-scale earthquakes, which more accurately reflect the conditions found in the Earth’s crust compared to previous smaller-scale studies. Reiju Norisugi, the first author of the paper, explained, “We applied an advanced machine learning technique to data collected from a meter-scale rock-friction experiment that generates ‘stick-slip’ laboratory earthquakes.” This method allowed for the detection of numerous acoustic emissions, which serve as foreshock signals indicating that a fault is nearing failure.
The researchers discovered that machine learning models trained solely on foreshock data could effectively identify these subtle signals that emerge just prior to the onset of laboratory earthquakes. To validate their findings, the team compared the machine learning model’s performance with physics-based numerical simulations that replicate the experimental data.
One of the crucial insights from their analysis was that the evolution of shear stress on “creeping” areas of the fault provided more significant predictive power than the average stress across the entire fault. Yoshihiro Kaneko, the team leader, noted, “These localized stress changes provide more diagnostic information, underscoring the importance of monitoring spatially heterogeneous fault-slip behavior.”
By successfully demonstrating that machine learning can detect and interpret minor physical changes before rupture in realistic, large-scale laboratory fault systems, this research represents a meaningful advancement in the quest for short-term earthquake forecasting. The study bridges the gap between laboratory research and the complexities of natural fault systems, paving the way for future investigations into earthquake prediction.
The implications of this work could be profound, as understanding the final stages of fault loading is a vital component of developing effective forecasting methods. As researchers refine these machine learning models and their application to real-world scenarios, the hope is that they can improve public safety and preparedness in the face of natural disasters.
