Deep learning for prediction of complex geology ahead of drilling

Abstract

During a geosteering operation the well path is intentionally adjusted in response to the new data acquired while drilling. To achieve consistent high-quality decisions, especially when drilling in complex environments, decision support systems can help cope with high volumes of data and interpretation complexities. They can assimilate the realtime measurements into a probabilistic earth model and use the updated model for decision recommendations. Recently, machine learning (ML) techniques have enabled a wide range of methods that redistribute computational cost from on-line to offline calculations. In this paper, we introduce two ML techniques into the geosteering decision support framework. Firstly, a complex earth model representation is generated using a Generative Adversarial Network (GAN). Secondly, a commercial extra-deep electromagnetic simulator is represented using a Forward Deep Neural Network (FDNN). The numerical experiments demonstrate that the combination of the GAN and the FDNN in an ensemble randomized maximum likelihood data assimilation scheme provides real-time estimates of complex geological uncertainty. This yields reduction of geological uncertainty ahead of the drill-bit from the measurements gathered behind and around the well bore.

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

“…For every realization in the ensemble, we use a three-step modelling sequence to convert the model vector to the EM-log forecasts, see In this paper, we use an iterative ES, namely, the approximate Levenberg-Marquardt ensemble randomized maximum likelihood method (EnRML) introduced by Chen and Oliver (2013), to condition the prior model ensemble to the real-time measurements. At every iteration, the method minimizes an objective function, representing a regularised statistical misfit between the expected and the actual measurements (Fossum et al, 2021). As a result, the ensemble of model vectors is updated to better match observations, see…”

“…We select a representative synthetic truth, which contains several channel groups surrounded by crevasse splays, see Figure 5. We make the simplifying assumption that the resistivity for each facies is known and constant, and the reservoir is oil-filled: background/ shale, 4.0 ohm m; oil-saturated channel sand, 171.0 ohm m; partially saturated crevasse splay sand, 55.0 ohm m. We use the same resistivity values and the same forward deep neural network model (Alyaev et al, 2021) to generate the true and the expected measurements from the ensemble.…”

“…These methods are specifically advantageous in real-time applications where prediction/evaluation performance is critical. Fossum et al (2021) proposed using a generative adversarial network (GAN) to produce complex geological realizations in real-time to aid geosteering. The authors verified that the chosen iterative ES reduces the uncertainty in the GAN-based multi-realization model.…”

“…Second, the geo-model is transformed into a series of 1D resistivity projections at measurement locations. Third, a forward deep neural network (FDNN) converts the 1D inputs to synthetic EM logs: a full suite of extra-deep EM log traces transmitted in real-time during drilling(Alyaev et al, 2021). The extra-deep EM logs include 22 measurements with a sensitivity of up to 20 meters from the tool(Larsen et al 2018).…”