Joint facies and reservoir properties estimation by trans-dimensional seismic inversion using machine learning based priors

Abstract:

Markov chain Monte Carlo algorithms, a popular class of Bayesian inference techniques, are employed for accurate estimation of posterior probability distribution of model parameters in seismic inversion. We implement a trans-dimensional Bayesian inversion approach that infers continuous reservoir properties, discrete facies models, layer boundary locations, and the associated uncertainties from pre-stack seismic data. RJMCMC (reversible jump markov chain monte carlo) is an effective tool to solve such trans-dimensional problems. We have extended the RJMCMC algorithm to simultaneously invert for discrete parameters (facies) and continuous parameters (elastic and petrophysical properties). At each location, the target reservoir properties have multimodal and non-parametric distributions. Our method iteratively samples the facies, by moving from one mode to another, and reservoir properties, by sampling within the same mode. The integration of facies classification in the reservoir characterization process aims to provide a geologically consistent relation between the elastic and petrophysical properties and their uncertainties. However, because of limited data bandwidth, data noise and imperfect model parameterization, the inversion of seismic reflection data is an ill-posed problem from which the reservoir properties cannot be uniquely recovered. To reduce this non-uniqueness and promote faster convergence of MCMC algorithms, optimal proposal distributions need to be determined. Here we use statistical learning to generate priors for our facies variable. The proposal distribution for continuous reservoir properties are non-parametric and facies dependent and are generated from the well-logs. Synthetic and field data inversions demonstrate that seismic inversion using our methodology can generate high resolution reservoir models of facies and reservoir properties.