Probabilistic forecasting for geosteering in fluvial successions using a generative adversarial network

Quantitative workflows utilizing real-time data to constrain uncertainty have the potential significantly improve geosteering. Fast updates based on are particularly important when drilling in complex reservoirs with high uncertainties pre-drill models. However, practical assimilation of requires effective geological modelling and mathematically robust parameterization. We propose a generative adversarial deep neural network (GAN), which is trained reproduce geologically consistent 2D sections fluvial successions. Offline training produces fast GAN-based approximation geology parameterized as 60-dimensional model vector standard Gaussian distribution each component. Probabilistic forecasts generated using an ensemble equiprobable realizations. A forward-modelling sequence, including GAN, converts initial (prior) realizations into EM log predictions. An smoother minimizes statistical misfits between predictions data, yielding update vectors reduced around well. Updates can then be translated probabilistic facies resistivities. This paper demonstrates workflow for geosteering outcrop-based synthetic succession.In our example, method reduces correctly predicts most major features up 500 m ahead drill-bit.The condensed summary also submitted presentation at 3rd EAGE/SPE Geosteering Workshop held 2–4 November 2021, online.