Global Mapping of Small-Scale Heterogeneities at the Core-Mantle Boundary: Insights from Deep Learning Analysis of PKP Precursors

Small-scale lateral heterogeneities at the lowermost mantle are fundamental to understanding mantle convection dynamics and core-mantle interactions. PKP precursors, generated by seismic scattering from fine-scale structures near the core-mantle boundary (CMB), provide a powerful yet underutilized probe for imaging deep Earth heterogeneities. However, the manual identification of these weak signals is inefficient, subjective, and inadequate for the vast volumes of modern seismic data.
We present a comprehensive analysis of global PKP precursor observations using a supervised deep learning framework combined with iterative human-guided optimization. Processing over 2 million vertical-component waveforms from earthquakes (Mw ≥ 6.0) recorded between 1990 and 2024, we automatically identified 227,770 high-quality PKP precursor signals—an order of magnitude increase compared to previous global compilations. This unprecedented dataset, termed DeepScatter-PKP, provides the densest and most spatially complete observational foundation for characterizing CMB scattering structures to date.
To systematically evaluate the stability and spatial distribution of scattering signals, we developed a dual-probability framework integrating precursor occurrence probability (Pocc) and scatterer location probability (Pscat). This approach enables simultaneous assessment of broad-area scattering stability and precise localization of strong scatterers. Our significantly enhanced sampling density and coverage connect previously isolated scattering patches into continuous anomaly belts, notably beneath the Pan-American region and the western Pacific margin.
Cross-validation with independent seismic phases confirms the robust embedding of multiple ultra-low velocity zones (ULVZs) within diverse velocity heterogeneity backgrounds, suggesting thermochemical origins involving remnants of multi-episode subducted slabs, partial melting, and interactions with large low-velocity provinces (LLVPs). Extension to undersampled regions reveals six previously unidentified high-potential strong scattering zones, including beneath the South Atlantic, high-latitude Eurasia, and circum-Antarctic domains.
Our results demonstrate that small-scale scatterers occur in both high-velocity and low-velocity domains, highlighting the diversity and independence of their origins beyond LLSVP boundaries. The DeepScatter-PKP dataset and dual-probability framework establish priority targets for future multi-phase joint inversions and high-resolution CMB imaging, offering new constraints on the thermochemical state and dynamic evolution of Earth's deep interior.