Almost 35 years of seismic data. More than two million waveforms from approximately 5,000 earthquakes of magnitude 6 or greater. Manual analysis had examined this file for decades and kept missing something. A deep learning system developed by researchers at the Chinese Academy of Sciences did not. It identified approximately 174,929 weak seismic signals (more than ten times all previous catalogs combined) and mapped six zones at the boundary between Earth’s core and mantle that no one had documented before. The study appeared in the Journal of Geophysical Research: Solid Earth in late August 2026. A schematic showing how different seismic waves travel through the Earth, from the source of an earthquake (star) to a seismic station on the surface (triangle). The waves are deflected near the core-mantle boundary (CMB). PKP precursors pass through the outer liquid core, but not the solid inner core, and reach the detectors before the PKIKP waves. (Guan et al., J. Geophys. Res. Solid Earth, 2026) Training a classifier on the Earth’s heartbeat The AI method is as important as the discovery itself. The signals in question are called PKP precursors: weak echoes that are scattered in the irregularities where the solid rock of the mantle meets the liquid outer core of iron and nickel, approximately 2,900 km deep. They arrive just before the main seismic wave, weak enough to be drowned in the noise. Finding them manually was slow, subjective, and geographically patchy. The team trained a deep learning classifier on examples of human-labeled precursors, ran it across decades of archived recordings, and then used iterative human-in-the-loop validation to improve accuracy—the same methodology behind content moderation systems, except the training data is earthquake noise rather than flagged posts. Think of it like training Shazam on every song ever recorded, then playing it silently, listening for the ghost of a melody. What the system found were six previously undocumented regions of strong dispersal, labeled B1 to B6, beneath high-latitude Eurasia, Central Asia, the South Atlantic, and other undersampled areas. Previous studies have observed isolated and seemingly random anomalies. The new data set reveals something more surprising: larger, belt-like areas stretching across parts of the world. As the paper states: “We also discovered six areas that likely harbor significant heterogeneities that have never before been documented, providing clear priority targets for future exploration of the Earth’s deep interior.” Earthquakes with identified PKP precursors that traveled from the source (pink stars) to seismic array detectors (blue triangles). (Guan et al., J. Geophys. Res. Solid Earth, 2026) What’s really down there? Ancient subducted slabs, partial melting, and a highly speculative cosmic collision. The boundary between the core and the mantle is already strange territory. Solid but malleable mantle rock encounters liquid iron-nickel at about 5,500 °C, with a temperature jump of about 1,000 °C across the interface. Known features here (massive low shear velocity provinces beneath Africa and the Pacific) already feed competing geodynamic models. The heat flow at this boundary drives Earth’s geodynamo, the mechanism behind the magnetic field that shields from solar radiation. The six new zones are regions where temperature and composition likely differ markedly from the surrounding mantle material, which the researchers cautiously describe as deep heterogeneities rather than fully mapped structures. Possible origins include: Remnants of subducted slabs carried over billions of years Localized partial melting Mineral phase transitions Some coverage has raised a more dramatic hypothesis: that these zones could contain material from Theia, the Mars-sized object that collided with the early Earth to form the Moon. That’s still squarely in hypothesis territory. Seismic data confirm the difference in composition. It doesn’t say anything about the origin. The six new zones (B1-B6) identified in this analysis likely contain “deep-seated scatterers” that affect the way seismic waves travel through the Earth. (Guan et al., J. Geophys. Res. Solid Earth, 2026) The Bigger Picture of AI in Science The most important AI isn’t always the one in your pocket. Domain-specific machine learning continues to generate discoveries that consumer-facing AI products rarely touch. The catalog of PKP precursors will continue to grow. Researchers hope it will improve fine-scale models of the lower mantle and eventually enable similar methods on other planets, once enough seismic data exists — a pattern already seen when airborne LiDAR reshaped archaeological understanding, or when earthquake debris became a testing ground for sensor-driven robotics. Six new objectives. A method that scales. The planet beneath your feet is still revealing its secrets, and the tool that finally heard them was unremarkable. He was patient, precise and pointed to the correct file.
