Dimensionality reduction and uncertainty quantification in seismic waveform inversion

Abstract

Randomized techniques have gained a tremendous increase in popularity with the ever-growing abundance of observational data, such as is the case in seismic inversion. In this research, the high-dimensional prior covariance matrix of Bayesian problems is estimated using a randomized singular-value decomposition. Furthermore, randomized trace estimation and stochastic batching methods are outlined against a new RSVD-inspired source-encoding algorithm in full waveform inversion.