Chinese Academy of Sciences researchers used a deep-learning model to scan over two million seismic records and identified 174,929 faint PKP precursor signals. The resulting map reveals six previously undocumented regions of heterogeneity near the core–mantle boundary, roughly 2,900 kilometers beneath the surface.
The catalogue, ten times larger than all previous efforts combined, was published in the Journal of Geophysical Research: Solid Earth. The map sharpens the picture of Earth’s deep interior, but what these structures are made of remains inference, not observation.
The numbers are what make the discovery legible. A team at the Chinese Academy of Sciences trained a deep-learning classifier to scan more than two million seismic records spanning 1990 to 2024. It flagged exactly 174,929 PKP precursors — faint echoes that scatter off features near the core–mantle boundary. The haul is ten times larger than every earlier catalogue combined. From that data emerged a map: six undocumented zones where the planet’s deep interior is uneven enough to scatter seismic waves. The signal has always been there. An algorithm trained to find it just read the whole archive for the first time.
Six zones the manual era missed
The classifier sifted through waveforms from roughly 5,000 earthquakes of magnitude 6 or higher, caught between 1990 and 2024. It identified 174,929 PKP precursor arrivals linked to small-scale heterogeneities — regions where temperature or composition depart sharply from the surrounding mantle. The team identified approximately 174,929 faint seismic signals. That single data point reframes the problem. Earlier manual studies had recovered isolated patches. The new catalogue reveals broad, belt-like scattering zones.
Six of those zones, labeled B1 through B6, had never been documented. They sit beneath high-latitude Eurasia, Central Asia, the South Atlantic, and other regions that were sparsely sampled before. The group includes possible signatures of ancient subduction, mineral phase changes, or localized partial melting — physical conditions that make the lowermost mantle a more complicated place than models have assumed.
The challenge of extracting these signals is acute. PKP precursors arrive just before the main compressional wave, weak enough to drown in noise. “PKP precursor work remains methodologically difficult because the signals are faint, sparse, and highly sensitive to noise and station coverage,” notes the seismological research community at IRIS. Manual review could never keep up. The deep-learning pipeline, by contrast, iteratively taught itself what a real precursor looks like, then flagged likely matches across the entire archive.
The process that turned two million waveforms into a catalogue of scattering regions didn’t happen in one pass. The sequence below tracks the logic: training, detection, validation, and mapping.
The results strengthen the case that some previously isolated anomalies are part of broader belt-like structures in the deep mantle. That matters for geodynamic models. A handful of isolated scatterers can be dismissed as noise. Continuous zones imply a global process — one that may be tied to how material cycles through Earth’s interior. The data provide a higher-resolution basis for studying fine structure and geodynamic state in Earth’s deep interior. What they do not yet give is a material diagnosis.
The archive that keeps learning
The real shift here is not the six zones but the method. For decades, the global earthquake archive sat partially read — too vast, too noisy, too slow to audit by hand. The Chinese Academy of Sciences team taught a classifier to recognise a faint seismic fingerprint, then ran it across the whole record. The result is not just a bigger catalogue. It is a demonstration that legacy data, when reprocessed with better algorithms, can quietly contain new discoveries.
The study appears in a respected specialty journal, and its scientific standing is that of a major catalogue expansion, not a settled explanation of origin. Independent teams will need to test the detection pipeline and verify the six-zone map. If the expected supplementary methods and robustness checks are published with the 2026 cycle, the door opens for consensus. If not, the result will stay influential but unconfirmed — a sharp map with no material legend.
The model’s real output is not the six zones. It is a method that turns the whole archive into a reusable signal layer — one that only gets richer the next time a better classifier is trained on it. That is what makes the 2026 catalogue a beginning, not an endpoint.
Beyond the headline
The Science Gap
The real gap is between detecting a signal and identifying what made it. This study improves the first part dramatically, but the second still depends on indirect inference from wave scattering, so the map is best read as a high-resolution clue set rather than a finished inventory of deep-Earth materials.
The Reach
The implication reaches geophysics software and workflow, not just Earth science theory. The mechanism is scalable machine learning on archival seismic data, and the non-obvious consequence is that similar methods can expose faint structure in other underused geophysical datasets that were previously too large to audit by hand.
The Bigger Picture
This is part of a broader shift in earth science from sparse manual interpretation to algorithmic recovery of hidden structure from legacy archives. The deeper change is not just better mapping; it is that the archive itself becomes richer every time better models are applied, meaning old data can keep producing new geological signals.
A catalogue that resets the deep-Earth research agenda
With a new map of deep scattering zones now in the open literature, researchers across four disciplines have a rare chance to act on it immediately.
- Seismologist or Geophysics Researcher
You should evaluate the deep-learning methodology for your own datasets. The Chinese Academy of Sciences research highlight and the full article in Journal of Geophysical Research: Solid Earth (DOI e2025JB033195) give you the detection pipeline and the map. Use them to re-examine existing traces for missed precursors and to test whether the six belt-like zones appear in independent station data.
- AI/Machine Learning Scientist in Geoscience
Study the model architecture and the iterative human‑validation loop. The classifier’s success on faint, sparse signals in a noisy archive suggests similar approaches can be ported to other geophysical problems — from low-amplitude seismic tremor to weak electromagnetic or gravitational anomalies. The method’s transferability is the near‑term prize.
- Earthquake Hazard Modeler
Monitor follow‑up work that links these heterogeneities to mantle convection and stress accumulation. While this study does not hand you a forecast parameter, it sharpens the base map of deep‑Earth structures that underpin long‑term tectonic models. When improved slab maps are released, compare them against the six zones to assess whether they align with known subduction geometries.
- Planetary Scientist or Geodynamicist
Integrate the new scattering regions into your geodynamic simulations. The belt‑like pattern and the depth constraints offer stronger boundary conditions for models of mantle plumes, slab descent, and core–mantle interactions. The open question — whether these heterogeneities are slab‑derived, partially molten, or something else — is now a testable hypothesis against the higher‑resolution map.
Explainer
- PKP precursors
- PKP precursors are faint seismic arrivals that appear just before the main PKP wave, which travels through Earth’s outer core. They are scattered off small-scale heterogeneities near the core–mantle boundary and reveal structure too fine for standard tomographic methods. Because they are extremely weak, cataloguing them has historically required tedious manual inspection that missed most events.
- Core–mantle boundary
- The core–mantle boundary lies roughly 2,900 kilometers beneath the surface, where solid silicate mantle rock meets the liquid iron‑nickel outer core. A temperature jump of about 1,000 degrees Celsius and sharp contrasts in density and composition drive much of the convection that powers Earth’s magnetic field. Direct sampling is impossible, so seismic waves are the only lens we have.
- Heterogeneities
- In deep‑Earth seismology, heterogeneities are regions where physical or chemical properties differ from the surrounding mantle — variations in temperature, mineral phase, or composition that scatter seismic waves. They can be remnants of subducted tectonic plates, pockets of partial melt, or reaction products from mantle‑core interactions. The term carries no assumption about origin, only that the material is uneven enough to produce a measurable seismic echo.
- Subduction
- Subduction is the process by which one tectonic plate slides beneath another and sinks into the mantle, carrying surface material deep into Earth’s interior. It is the primary engine of plate tectonics and the main way elements and minerals are recycled. Over billions of years, subducted slabs can reach the core–mantle boundary, where they may contribute to the heterogeneities mapped in this study.
RELATED STORIES
