Verification of a real-time ensemble-based method for updating earth model based on GAN

The complexity of geomodelling workflows is a limiting factor for quantifying and updating uncertainty in real-time during drilling. We propose Generative Adversarial Networks (GANs) for parametrization and generation of geomodels, combined with Ensemble Randomized Maximum Likelihood (EnRML) for rapid updating of subsurface uncertainty. This real-time ensemble method is known to be approximate for non-linear forward models and might therefore produce inaccurate and/or biased posterior solutions when combined with a highly non-linear model arising from the neural-network modeling sequences. This paper illustrates the predictive ability of EnRML on several examples where we assimilate local extra-deep electromagnetic logs. Statistical verification with MCMC confirms that the proposed workflow can produce reliable results required for geosteering wells.

Automated summary

This paper proposes a method that uses GANs for parameterization and generation of geomodels, combined with EnRML for rapid updating of subsurface uncertainty. It illustrates the predictive ability of EnRML on assimilating well log data through several examples and verifies the results statistically using MCMC.