Denoising teleseismic data for deep Earth studies using a supervised deep-learning auto-encoder: a case study of diffracted waves from ULVZs

Stuart Russell , Katrin Hannemann, Carl Martin, Bennet Lindhorst, Waleed Esmail, Christine Thomas

Seismic noise, particularly the microseism generated by the oceans, is a fundamental limitation on using short period (1 - 10 s) teleseismic data to study the deep Earth. Deep learning auto-encoders have proven effective at denoising seismic data in other applications. Here, we provide a demonstration of their potential for improving the quality of deep Earth seismic data, using Sdiff post-cursors generated by ultra-low velocity zones (ULVZs) as a case study. We present a computationally efficient and accessible workflow, utilising a supervised learning process with 1D synthetics and real, global seismic noise. This workflow does not require real teleseismic signals, which are limited in number and quality, and is computationally efficient, not requiring expensive 3D spectral-element simulations. We present a series of example applications, finding the denoiser to be effective at removing noise and extracting Sdiff signal despite having not been trained with real teleseismic signals. Furthermore, we provide recommendations on how performance could be improved, including demonstrating transfer learning as an efficient and effective way to improve extraction of specific signals. Lastly we apply denoising to seismic data sampling known ULVZs, identifying previously unreported short-period signals from several different ULVZs and obtaining observations of the Hawaiian ULVZ along a novel source-receiver geometry from the East Pacific Rise to Japan, which is perpendicular to previous ray paths. These observations reveal previously undocumented frequency- and azimuth-dependent Sdiff waveforms associated with ULVZs, demonstrating that denoising can expand observational constraints on deep Earth structure.