Coupling Bayesian inversion with machine learning for uncertainty quantification of dynamic reservoir property predictions and application to field data

Estimating reservoir property changes from 4D seismic data suffers from a large degree of uncertainty as the changes are often subtle and the seismic data are noisy and bandlimited. To use 4D seismic information reliably, we combine machine learning (ML) with a Markov chain Monte Carlo (MCMC) based Bayesian model. We adopt a stepwise ML to provide informative priors for unknown pressure and fluid saturation changes and develop efficient MCMC sampling methods to draw many samples from the defined joint posterior probability distribution. We demonstrate the effectiveness of this new approach using a field dataset from the Gulf of Mexico (GOM).