The complexity and computational cost of conventional geomodelling workflows limit their applicability for formation evaluation in real-time using data acquired during drilling. We propose a workflow replacing traditional geomodels with a Generative Adversarial deep neural Network (GAN) trained to reproduce sections of complex geology. Offline training produces a fast GAN-based approximation of, in our case, fluvial geology parameterized as a 60-dimensional model vector with standard Gaussian distribution of each component. The Ensemble Randomized Maximum Likelihood (EnRML) method reduces geological and petrophysical uncertainty by integrating real-time extra-deep EM measurements into GAN-model realizations. The ensemble of updated GAN-geomodel realizations provides probabilistic forecasts of facies around and ahead of the well.
The approximations in the EnMRL method combined with a highly non-linear GAN model may produce inaccurate or biased predictions. To test if the proposed workflow provides reliable results, we performed a statistical verification with an accurate but slow Markov Chain Monte Carlo (MCMC) method. In our testing, both the slow and the real-time methods reduce uncertainty and correctly predict major geological features up to 500 meters ahead of drill-bit. The results - demonstrated in a synthetic environment - indicate the method’s potential to improve assisted real-time formation evaluation in complex geology significantly.