A team of geologists used a deep learning system to analyze more than two million seismic records and identified six continuous bands of anomalies at the boundary between Earth's core and mantle — at a depth of roughly 2,900 kilometers. The study's results were published in the journal Journal of Geophysical Research: Solid Earth (2026, DOI: 10.1029/2025jb033195). Direct drilling to this depth remains technologically impossible, so the only "window" into the planet's interior is provided by seismic waves generated by earthquakes around the world.

How a Neural Network Processed Two Million Seismograms

To study the core–mantle boundary in detail, the researchers used so-called "wave precursors" — weak seismic signals that arrive a few seconds before the main waves and make it possible to detect small heterogeneities in the planet's deep layers. Manually analyzing such signals across thousands of records would have required enormous computational resources and time. To overcome this limitation, the scientists applied a deep learning system that processed seismograms collected by observatories around the world between 1990 and 2024. It was precisely this neural-network-level automation that made it possible to analyze a two-million-sample dataset and to identify patterns invisible to traditional methods.

Where the Six Anomalous Bands Are Located

The detected zones of heterogeneity cover extensive regions of the planet: beneath the North Atlantic, northern Eurasia, the South Atlantic, southern Africa, the Pacific Ocean, and the Antarctic region. Several of the identified signal-scattering zones coincide with previously known ultra-low velocity zones (ULVZ), which confirms the validity of the method. The core–mantle boundary plays a key role in regulating heat flow, mantle circulation, and the formation of volcanic plumes, so the discovery of new structural anomalies in this zone has direct implications for understanding geodynamic processes.

What Lies Behind the Anomalies: Plates, Chemistry, and Melting

In the authors' view, the complex structure of the lower mantle in the identified zones formed as a result of several processes: the mixing of subducted ancient tectonic plates, chemical segregation (the separation of elements by density and chemical composition), and localized melting of rocks. This zone contains both large-scale formations and small local patches with deviating temperature or chemical composition. The authors emphasize that the map they created shows probable areas of seismic signal scattering, not the exact boundaries of subsurface structures, and that further combining this catalog with other types of seismic waves will allow the details of Earth's internal structure to be refined.

Contradictory Data

The provided materials contain a discrepancy in the terminology of the seismic phases. The main text of the study and the RBC-Ukraine overview article use the term "PKP wave precursors" (P-waves that have traveled through the outer core), whereas the caption to the illustration of the global distribution of entry and exit points in the journal's publication refers to the PKIKP phase (P-waves that have traveled through the inner core). These are technically distinct seismic phases, and the precise identification of the phase used affects the interpretation of the depth and nature of the detected anomalies. In addition, a number of Russian media outlets (in particular, TASS) present the study's results as a "map of iron leaks from the core into the mantle," whereas the authors in JGR: Solid Earth frame their findings as the detection of zones of heterogeneity and scattering, without reducing them exclusively to iron transport. This difference in emphasis could mislead readers about the scale and nature of the structures identified.

Prospects and Limitations

The study demonstrates that combining large seismic datasets with machine learning methods opens access to layers of the Earth that previously remained beyond the reach of detailed mapping. However, the authors honestly point out the limitations: the map reflects probable scattering areas rather than exact structural boundaries, and interpreting the physicochemical nature of the anomalies requires additional data. Future work will involve integrating the catalog of PKP/PKIKP precursors with other types of seismic waves (S-waves, normal modes, diffracted phases), which will make it possible to build a more detailed three-dimensional model of the lower mantle and, possibly, to revise existing views on the planet's geodynamic history.