Seismic Tomography Using Variational Inference Methods

摘要: Seismic tomography is a methodology to image the interior of solid or fluid media, and often used map properties in subsurface Earth. In order better interpret resulting images it important assess imaging uncertainties. Since significantly nonlinear, Monte Carlo sampling methods are for this purpose, but they generally computationally intractable large datasets high-dimensional parameter spaces. To extend uncertainty analysis larger systems we use variational inference conduct seismic tomography. contrast sampling, solve Bayesian problem as an optimization problem, yet still provide probabilistic results. study, applied two methods, automatic differential (ADVI) Stein gradient descent (SVGD), 2D problems using both synthetic real data compare results those from different methods. The show that can produce accurate approximations at lower computational cost, provided gradients parameters with respect be calculated efficiently. We expect fruitfully many other types geophysical inverse problems.