Integrated Detection and Classification of Low-Frequency Earthquakes Using a PhaseNet-Based Deep Learning Framework

Deep beneath the Earth's crust, where rock is hot and pressure is immense, a quiet kind of shaking often goes unnoticed. These are low-frequency earthquakes, events that release energy in a slow, rumbling pulse rather than the sharp, high-pitched snap of a typical quake. While ordinary earthquakes might vibrate at a frequency similar to a high note on a piano, these deep tremors hum at a much lower pitch, often with a start that is so gradual it is hard to pinpoint exactly when they begin. Scientists have long suspected that these subtle shivers are not random; they are likely the result of fluids, such as water or magma, moving through cracks in the rock or slipping along fault lines without causing a violent rupture. Understanding these hidden movements is crucial because they may hold clues to how stress builds up in the Earth, potentially leading to larger, more destructive earthquakes. However, because these signals are weak and blend easily with background noise, traditional methods of listening to the Earth often miss them entirely, leaving a significant gap in our knowledge of deep crustal processes.

To fill this gap, a team of researchers developed a new way of listening using artificial intelligence. They trained a computer model, based on a system originally designed to find standard earthquakes, to recognize the unique signatures of these low-frequency events. Instead of just looking for a sharp start and an end, the model was taught to identify five different possibilities in the seismic data: the start of a normal earthquake, the end of a normal earthquake, the start of a low-frequency event, the end of a low-frequency event, and background noise. By feeding the system thousands of recorded waveforms from Japan, the researchers taught it to distinguish between the clear, sharp arrivals of ordinary quakes and the fuzzy, prolonged arrivals of the low-frequency kind. Once trained, the model acted as a highly sensitive ear, scanning continuous recordings of the Earth's vibrations and picking out arrivals that human observers or older software would have ignored.

The researchers then applied this new tool to a specific area of intense seismic activity in northern Yamaguchi Prefecture, where a swarm of earthquakes had been occurring at depths greater than 20 kilometers. The results were striking. The new method detected more than twice as many earthquakes as the official catalog maintained by the Japan Meteorological Agency. It found hundreds of tiny events that were previously invisible, including many that were too small to be recorded by standard instruments. More importantly, the system did not just count the events; it classified them. By analyzing the specific probability scores the model assigned to each arrival, the researchers created a simple index to decide whether a detected event was a low-frequency earthquake or a standard one. This allowed them to map out the swarm with a level of detail that was previously impossible, revealing that the low-frequency events were not scattered randomly but were concentrated in specific, deeper zones within the swarm.

However, the study also revealed a complication. The researchers found that the new index was not perfect at distinguishing natural low-frequency earthquakes from human-made explosions, such as those from quarry blasting. In fact, some of the events the system flagged as low-frequency earthquakes turned out to be blasts from a nearby quarry, which happen to produce similar low-frequency rumbles. This finding serves as a reminder that while the technology is powerful, it requires careful interpretation to separate natural geological processes from human activity. Despite this limitation, the ability to detect such a vast number of previously hidden events provided a clearer picture of the swarm's behavior. The data showed that the earthquakes migrated over time, moving from deeper parts of the crust toward shallower levels in a pattern that closely matched the movement of fluids diffusing through rock.

The spatial distribution of these events offered further evidence for this fluid-driven theory. The low-frequency earthquakes were clustered in the deepest and westernmost parts of the swarm, a region where the rock structure showed signs of high fluid content. This alignment suggests that the movement of hot fluids from the upper mantle or lower crust was likely the engine driving the entire swarm. The researchers observed that the migration of the earthquakes did not follow a simple, steady path; instead, it accelerated and slowed down, sometimes even reversing direction, which implies a complex interaction between fluid pressure, heat, and the rock's ability to fracture. By combining the new detection method with detailed maps of the underground rock structure, the study suggests that the swarm was sustained by a continuous supply of fluids that kept the rock weak and prone to slipping.

Ultimately, this work demonstrates that deep low-frequency earthquakes are far more common and widespread than previously thought, hiding in plain sight within the noise of standard seismic records. The new framework provides a robust way to monitor these events automatically, offering a window into the deep Earth that was previously closed. While the method still needs refinement to perfectly filter out human-made noise, it has already reshaped the understanding of a specific earthquake swarm, showing that what looks like a chaotic jumble of small quakes is actually a structured, fluid-driven process. As the technology improves, it promises to reveal the hidden rhythms of the deep crust, helping scientists better understand the slow, fluid movements that shape our planet's seismic future.