Deep‑Learning Scan Uncovers Six Anomalous Features at Earth’s Core‑Mantle Edge

Researchers have applied artificial intelligence to seismic data and identified six distinct, previously unrecognized structures at the boundary separating Earth’s liquid outer core from its solid mantle, a region located roughly 2,900 kilometres beneath the planet’s surface.

The discovery, detailed in a paper published in Journal of Geophysical Research: Solid Earth, stems from a novel approach that trains deep‑learning algorithms on thousands of earthquake recordings. By teaching the model to recognize subtle variations in the speed and direction of seismic waves, the team was able to map heterogeneities that conventional methods had missed.

Because no technology can physically reach the core‑mantle boundary, scientists rely on seismic waves generated by earthquakes to infer its properties. As these waves travel through Earth’s interior, they are altered by changes in material composition, temperature, and phase. The new AI‑driven analysis highlighted six zones where the waveforms deviated consistently, suggesting the presence of unusual structures such as localized compositional anomalies or thermal plumes.

These findings add to a growing body of research that portrays the core‑mantle interface as a dynamic, heterogeneous zone rather than a smooth, uniform surface. Prior studies have identified large low‑shear‑velocity provinces and ultra‑low velocity zones, but the six features reported here appear smaller and more discrete, hinting at previously unappreciated complexity.

The implications extend to several geophysical puzzles, including the mechanisms that drive mantle convection, the generation of Earth’s magnetic field, and the long‑term thermal evolution of the planet. Anomalies at the boundary could influence how heat is transferred from the hot core to the overlying mantle, potentially affecting the vigor of mantle plumes that surface as volcanic hotspots.

Lead authors caution that while the deep‑learning model provides a robust statistical signal, further validation with independent datasets and alternative analytical techniques will be necessary. Future work may involve integrating data from dense seismic arrays, such as those deployed in the United States and Europe, to refine the spatial resolution of the identified structures.

The study underscores the growing role of machine learning in Earth sciences, offering a powerful complement to traditional seismology. As computational tools become more sophisticated, researchers anticipate uncovering additional hidden features deep within the planet, advancing our understanding of the processes that shape Earth’s interior.