Comparison of random walk metropolis and adaptive metropolis MCMC algorithms in deep learning enabled Bayesian inference

Abstract

Different Markov chain Monte Carlo (MCMC) sampling algorithms are compared in this work. The parameter being estimated is injection rate and the forward model is coupled flow and geomechanics in the presence of fault. The forward model is reduced using deep learning in the form of LSTM autoencoder, and this reduced order model is used in the evalation passes in the MCMC algorithm.