This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Since scientists can't drill 2,900 km (1,800 miles) to the boundary between Earth's core and deep mantle, they must use information relayed by seismic waves recorded during earthquakes to study this region. In a new study published in the journal JGR Solid Earth, a group of scientists used deep learning to study a large set of a specific type of seismic wave.
The analysis revealed six continuous bands of irregularities at the core-mantle boundary (CMB) that had previously appeared only as sparse patches. The team involved in the new study used PKP precursors—faint seismic waves that typically arrive a few seconds before stronger seismic waves—to analyze what's happening at the CMB. The CMB helps control heat flow, mantle circulation and the rise of volcanic plumes, so insight into the region helps researchers understand the dynamics and evolution of Earth's deep interior.
The region contains large features but also much smaller, hidden patches with different temperatures or chemical makeup. PKP precursors are particularly good for analyzing these details. Normally, PKP precursors are weak and difficult to spot manually in earthquake records.
Researchers often try to improve analysis with techniques that enhance the signal-to-noise ratio, remove weaker signals and locate the source of wave scattering. But to be effective, researchers needed a scalable method to consistently analyze millions of seismic recordings. Manually, this would take a very long time.
But last year, another team successfully used a neural network framework to identify more than 30,000 PKP precursor signals using seismic data. However, that study depended more heavily on dense seismic arrays, limiting the coverage area. Still, this indicated that deep learning could be used to analyze PKP precursors that are normally too difficult to isolate.
For a more thorough investigation of PKP precursor signals and to map the distribution of deep-mantle heterogeneities, the researchers analyzed more than 2 million seismic recordings collected around the world from 1990 to 2024 using a deep learning system. The system filtered out poor-quality records and classified whether a waveform contained a PKP precursor. The team repeatedly checked and corrected samples by hand to improve the AI.
Extract — continue reading at the source.
