Global Distribution of PKP Precursors Derived From Three Decades of Seismic Data With Deep Learning

Abstract Small‐scale lateral heterogeneities in the lowermost mantle are crucial for understanding mantle convection and core‐mantle interactions. Owing to their high sensitivity to fine scatterers, PKP precursors provide a powerful probe of deep Earth heterogeneities. However, manual identification of these precursors is inefficient, subjective, and insufficient for vast global seismic data sets. Here, we integrate deep learning with iterative manual optimization to process over 2 million waveforms from earthquakes (Mw ≥ 6.0) between 1990 and 2024, automatically identifying 174,929 precursors, an order of magnitude more than prior data sets, to build the most comprehensive global map of potential scattering distributions. We propose a dual‐probability framework integrating precursor occurrence and scatterer location probabilities to assess the global spatial stability of precursor observations and to constrain the probable locations of localized strong scatterers. By applying this extensive catalog within the dual‐probability framework, we achieved an order‐of‐magnitude improvement in global sampling density and a substantial expansion of spatial coverage over previous individual‐detection studies, connecting once‐isolated sampling zones (e.g., beneath Pan‐America) into continuous bands. Cross‐validation with independent seismic phases confirms the robust embedding of multiple ultra‐low velocity zones (ULVZs) within diverse velocity heterogeneity backgrounds, suggesting that these structures represent thermochemical piles shaped by some combination of multi‐episode subducted slab remnants, localized partial melting, and interactions with large low‐velocity provinces (LLVPs). Extension to undersampled regions reveals, for the first time, six high‐potential scattering zones, laying the groundwork for prioritized target areas in subsequent multi‐phase joint inversions and fine‐scale imaging of core‐mantle boundary (CMB) heterogeneity.