Deep learning-based recovery of weak PcP phases from noisy seismic records

Detecting weak seismic phases in noisy recordings remains a major challenge in geophysics, especially for improving our understanding of the Earth’s deep interior. Among these phases, PcP, a P-wave reflection from the Core-Mantle Boundary (CMB), provides valuable constraints on structures such as the D” layer and ultra-low velocity zones. However, its inherently low amplitude and the masking effects of both irregular and random noise make its identification difficult, particularly in short-period records. To address this challenge, we adapt the Two-Step Deep Image Prior (TSDIP) method, originally developed for seismic reflection data, to the problem of detecting weak phases like PcP in global seismology. TSDIP takes advantage of the ability of deep neural networks to capture signal structure without supervision, beginning with an intentional overfitting stage that suppresses high-frequency noise. In the second stage, the same network is retrained on the initially denoised data to further reduce residual noise while preserving weak but coherent phases. Our implementation employs a fully connected neural network with carefully calibrated training epochs to balance noise suppression with signal preservation. Using both synthetic and field datasets, we demonstrate that TSDIP consistently outperforms Curvelet-based denoising, damped rank reduction, and the standard Deep Image Prior (DIP), yielding a more reliable recovery of PcP in heavily contaminated seismic records. This work establishes TSDIP as a powerful new tool for extracting weak seismic phases and improving imaging of deep Earth structures.