PESGB 18: Stochastic Seismic Waveform Inversion Using Generative Adversarial Networks As A Geological Prior
Talk, 1st PESGB/EAGE Workshop on Machine Learning in Geoscience 2018, London, UK
We present a probabilistic framework to solve ill-posed inverse problems governed by partial differential equations where a deep generative model is used as a prior on the coefficients governing the evolution of the solution space. A geophysical, seismic inversion problem is presented where the aim is to recover the spatial distribution p-wave velocities of a synthetic subsurface model, given observed acoustic waves at discrete recording stations on the surface. Posterior samples are acquired using a gradient-based approach based on stochastic gradient langeving dynamics. Talk PDF here