Abstract

Bayesian inference has become a valuable tool to solve geophysical inverse problems. However, the commonly-used Markov chain Monte Carlo (McMC) methods are generally computationally intractable for large datasets and high-dimensional parameter spaces. In this talk, I introduce a new method called variational inference to solve geophysical inverse problems. Variational inference uses optimization to solve the inverse problem, yet still produces fully nonlinear, probabilistic results. By applying the method to a range of applications, including travel time tomography and full waveform inversion, I demonstrate that variational inference can produce accurate approximations to the results obtained using McMC with significantly reduced computational cost. The method therefore provides an efficient alternative to McMC methods and can be applied to a variety of geophysical inverse problems.