Deep Learning Reveals Six Hidden Belts at the Boundary Between Earth's Core and Mantle

Scientists have used deep learning to map the boundary between Earth's core and mantle in unprecedented detail, revealing six continuous belts of unusual structure that had previously shown up only as scattered, isolated patches.

Since researchers cannot drill the roughly 2,900-kilometer depth to the core-mantle boundary, they rely instead on seismic waves generated by earthquakes. The team, whose findings appear in the journal JGR Solid Earth, focused on faint seismic waves seismologists have named PKP precursors, which arrive a few seconds before stronger seismic waves and are sensitive to small differences in temperature or chemical makeup at the boundary. These signals are normally too weak and scattered to analyze by hand at a large scale.

Building on an earlier neural-network approach that had identified more than 30,000 PKP precursor signals but was limited by its reliance on dense networks of seismic stations, the new team applied deep learning to more than 2 million seismic recordings collected worldwide between 1990 and 2024. Researchers repeatedly checked and corrected the system's classifications by hand to improve its accuracy at filtering poor-quality records and spotting the faint precursor signals.

The system identified 174,929 precursor signals, roughly 10 times the combined total of all previous global datasets. This expanded coverage connected previously isolated observations into six continuous belts of likely deep-mantle irregularity, located beneath the North Atlantic, northern Eurasia, the South Atlantic, Southern Africa, the Pacific, and the region around Antarctica. Several of the high-probability areas overlap with ultra-low-velocity zones already known from other methods, supporting the findings.

The researchers note that their maps show probable scattering regions rather than the exact outlines of hidden structures, and that coverage remains uneven because seismic stations and earthquakes are not evenly distributed around the globe.