AI Finds Six Unmapped Structures at Earth's Core–Mantle Boundary

Roughly 2,900 kilometres straight down, the solid silicate mantle that slowly churns beneath our feet gives way to Earth's liquid iron-nickel outer core. This core-mantle boundary, or CMB, is one of the most physically extreme and least directly observable regions of the planet. A new study shows that machine learning can now read its fine-scale structure in far greater detail than manual methods ever allowed.

The work comes from the Institute of Geology and Geophysics of the Chinese Academy of Sciences. Lead author Yurui Guan and colleagues trained deep-learning classifiers on more than two million seismic waveform records drawn from roughly 5,000 earthquakes of magnitude 6.0 and above, recorded worldwide between 1990 and 2024. The models were searching for a specific, notoriously faint kind of signal: PKP precursors. These are weak seismic arrivals that reach detectors a few seconds ahead of the main PKP phase, after being scattered by small-scale irregularities sitting just above the core-mantle boundary. Because they are easily drowned in noise, such signals had previously been retrieved by hand in small, geographically patchy batches.

The scale of the result is what sets the study apart. The system identified 174,929 PKP precursor signals, about ten times the combined total of every earlier global precursor catalogue. From that much larger sample the team built what they describe as the densest and most spatially complete global map of CMB scattering produced so far. Where previous, sparser datasets showed isolated patches of heterogeneity, the new catalogue connects many of them into broad, continuous belts running along the base of the mantle.

Stretching the interpretation further, the authors flag six regions of especially high scattering potential that had not appeared in prior literature. Coverage of the work places these zones beneath the North Atlantic, northern Eurasia, the South Atlantic, Southern Africa, the Pacific, and around Antarctica. The team labels them B1 through B6 and treats them as priority targets for future, higher-resolution imaging rather than as fully resolved objects. It is worth being precise about wording here: the maps show probable scattering regions, not the exact outlines of hidden structures, and no single seismic station can pin down a scatterer's location on its own.

Methodologically, the advance is as much about process as discovery. The researchers paired the neural network with iterative, human-in-the-loop validation: model outputs were repeatedly checked and corrected by hand, then used to retrain the system and suppress false positives. They also introduced a dual-probability framework that combines how often precursors appear at a given location with where a strong scatterer is most likely to sit, which lets them separate robust global patterns from noise.

On composition, the study is cautious. The authors argue that the patterns are consistent with thermochemical piles built from a mix of ancient subducted oceanic slabs, chemical segregation, and localized partial melting, possibly interacting with the large low-shear-velocity provinces long known beneath Africa and the Pacific. Several of the high-probability scattering areas overlap with ultra-low velocity zones already mapped independently by other seismic phases, which the authors take as supporting evidence. More dramatic origins, such as material left over from the Moon-forming impact, remain speculation the data cannot test.

The authors are explicit about the limits. Station and earthquake coverage is uneven across the globe, so some regions are far better sampled than others, and the probability maps trace where scattering is likely rather than the precise shape of any buried feature. They suggest that combining this catalogue with other seismic wave types, and deploying denser arrays, is the natural next step toward sharper images.

Analysis

Taken together, the six new zones matter less as individual discoveries than as evidence that the lowermost mantle is far better connected than sparse sampling suggested. The real scientific dividend is the catalogue itself: a tenfold-larger, consistently labelled dataset that future inversions can train on and cross-check. For a field historically bottlenecked by slow manual picking, that reproducible, machine-readable foundation may age better than any single map. One open question is whether the apparent belts are genuinely continuous structures or artefacts of where earthquakes and stations happen to sit; resolving that will require the multi-phase joint inversions the authors call for, not just more PKP data.

What this does, and does not, tell us about the surface is worth stating plainly. Better maps of CMB heterogeneity sharpen our estimates of the mineralogy, temperature, and flow at the base of the mantle, and those constraints feed models of mantle convection and the long-term behaviour of upwelling plumes. They do not, however, let anyone forecast earthquakes, volcanic eruptions, or tsunamis, and the study makes no such claim. The boundary's general influence on heat flow and mantle circulation is long established; linking these specific six zones to any surface hazard would be a step the evidence does not support.