Global Adjoint Tomography with Source Encoding

Abstract

We present our latest results on the tomographic models and method development for adjoint tomography. We start with concluding our decade-long global adjoint tomography project using classical methods. The third-generation model, GLAD-M35, features improved constraints on the P wavespeed and deep mantle structure, thanks to a 50\% increase in the event numbers and a new P-wave category that compensates for the imbalance between P and S measurements in the dataset. From the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update history we are able to approximate the Hessian and its inverse and provide estimates on the resolution and uncertainty of the model through their low-rank representations. We then explore the application of a new technology called source encoding. It dramatically reduces the number of simulations required for an iteration, which can potentially bring the global tomography to a next level. We demonstrate the effectiveness of source encoding through regional experiments. With a dataset of 786 events and 9,846 stations, we are able to recover the target model well with a computational cost much less than classical methods. We obtain a new regional model SE100 by applying our experiments with actual data and the new features revealed from the inversion are consistent with our existing knowledge and provide new insights to the mantle structure of the region. Finally, we explore the application of source encoding on a global scale, which is more challenging due to steady state condition and data availability. To increase the ray coverage, we introduce a new trace processing technique that replaces part of observed data with synthetic data, enabling us to utilize some traces that are otherwise not suitable for source-encoded inversion. We show that Laplace domain is the only option for global source encoding, which makes the measurement mainly based on first arrival and surface waves. To compensate for the lack of coverage in the Southern Hemisphere and deep mantle due to this measurement, we explore options like epicentral distance weighting and a mask for regional Fourier domain measurement.