Machine Learning Classification of Global Seismic Phases: A Cross-Validated, Physics-Informed Evaluation Using the IMS2.0 Automatic Arrival Bulletin
This study evaluates a Random Forest classifier trained on physical and statistical metadata from the IMS2.0 automatic arrival bulletin, demonstrating an 87% accuracy in identifying global seismic phases and highlighting the potential of combining machine learning with seismological principles to enhance CTBTO monitoring capabilities.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Earth is constantly speaking, sending out ripples of energy whenever the ground shifts, whether from a distant earthquake, a volcanic rumble, or the deep rumble of the ocean. To listen to this conversation, scientists have built a global network of sensitive instruments called the International Monitoring System. This network, operated by an organization dedicated to verifying a treaty against nuclear testing, keeps a twenty-four-hour watch on the planet. Its job is to catch these vibrations, figure out where they came from, and decide what kind of event caused them. The system is incredibly busy, generating a constant stream of data that includes the time a wave arrived, how strong it was, and the direction it came from. However, sorting through this flood of information is difficult. The system must distinguish between different types of waves, such as the fast-moving primary waves that travel through the planet's core, the slower waves that skim the crust, and the mysterious waves that travel through the ocean. Sometimes, the signals are so faint or mixed with background noise that even the best automatic computer programs struggle to tell them apart.
A team of researchers from the University of Nairobi decided to test whether modern computer learning could help solve this sorting problem. They focused on a specific, high-speed snapshot of data collected by the monitoring system on a single day in August 2026. This snapshot contained nearly fifteen thousand individual wave detections from stations all over the world. Instead of feeding raw sound recordings into a computer, which requires massive amounts of data and processing power, the researchers used the summary notes that the system already creates for every detection. These notes include simple numbers like the direction the wave came from, how fast it seemed to be moving, how loud it was, and how long it lasted. They trained a computer model, known as a Random Forest, to look at these numbers and guess the type of wave. The goal was to see if a machine could learn the subtle patterns in these notes better than the current rules used by the monitoring system, and to understand exactly where the machine succeeded and where it got confused.
The results showed that the computer model was remarkably good at identifying the most common and distinct types of waves. When the signals were clear and traveled long distances across the globe, the model correctly identified them nearly ninety-five percent of the time. It was particularly successful at spotting the primary waves that travel through the deep Earth and the regional waves that move through the crust. The researchers found that the model relied heavily on two specific clues: the direction the wave came from and its speed. This makes perfect sense, as these two factors are the most reliable way to tell different kinds of seismic waves apart. The computer learned that waves traveling at certain speeds from specific directions almost always belong to a specific category, confirming that the machine was not just guessing but was actually learning the physical laws that govern how energy moves through the Earth.
However, the study also revealed where the technology still hits a wall. The model struggled significantly with two specific types of detections: the T-phase, which are waves that travel through the ocean, and the N-class, which are signals that look like random background noise. The computer often confused these two, or mislabeled them entirely, achieving a much lower success rate than it did for the other waves. The researchers investigated why this happened and found that the problem was not a flaw in the computer's brain, but a limitation in the information it was given. The summary notes used for the T-phase and noise detections did not contain enough detail to separate them. The waves that travel through the ocean and the random noise often look very similar in terms of their speed and direction, and the notes lacked the deeper details of the wave's shape that would allow the computer to tell them apart.
This distinction is crucial for the future of global monitoring. The study proves that for the vast majority of seismic events, the existing data collected by the monitoring system is already rich enough to support a highly accurate, automatic classification system. A computer can now be trusted to sort through the clear signals and flag the important ones, potentially saving human analysts from having to review every single detection. Yet, the research also clearly defines the boundary of what is possible with current data. To improve the identification of ocean waves and faint background signals, the system would need to move beyond simple summary notes and analyze the actual shape of the sound waves themselves. Until that happens, the computer will remain excellent at listening to the Earth's clear voice, but it will still need human help to make sense of its whispers.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.