Geosteering for improved oil recovery

The placement of wells in a subsurface reservoir is a very important factor for the recovery of oil and gas. The process of adjusting the direction of the well during drilling addressing the newly gained insights in geology is called geosteering. Geosteering decisions are generally based on geological models. To make optimal decisions while drilling, it is necessary that geological models for the near-wellbore region are well-calibrated against measurements and that the uncertainty is quantified. The current work processes for geosteering suffer from shortcomings. (i) It is highly challenging to calibrate geo-models to Deep Electromagnetic (DeepEM) and Logging While Drilling (LWD) measurements for complex formations. (ii) There is a lack of flexibility in the current geo-models for support of structural updates and fast local updates. (iii) There is a lack of a transparent, systematic, and consistent workflow for quantifying complex geological uncertainties and using this information when making geosteering decisions. The primary objective of this project is to develop a methodology for geosteering by continuously updating geo-model based on LWD measurements including DeepEM measurements and making informed decisions. The first prototype of an interactive Decision Support System (DSS) has been developed and applied to synthetic cases in 2018. It uses a complete workflow with updating of multiple (ensemble) model realizations and the decision recommendations. The results show statistically optimal well landing in complex cases with several targets (https://doi.org/10.1016/j.petrol.2019.106381). To compare the DSS to decisions done by geosteering experts we have organized a geosteering competition during the 2019 Formation Evaluation and Geosteering workshop by NFES and NORCE. The system performed better than 95% of participants. The developed online competition platform can be further used as a benchmark for sequential decisions under uncertainty (see https://doi.org/10.1016/j.acags.2021.100072). The rest of the activities in the project are improving specific parts of the methodology and are thus addressing the secondary project objectives: (a) processing of realistic DeepEM measurements, (b) developing more flexible geo-model representation, and (c) improving the ensemble-based workflow. Real-time interpretation of DeepEM measurements requires fast and reliable modeling. Depending on the required update frequency, different levels of details can be required. In the project, we have developed a deep-neural-network-based model which is trained to approximate a commercial 1D Extra-Deep EM simulator and can simulate 5000 log positions per second (https://doi.org/10.1190/geo2020-0389.1). We have also developed a vendor-independent fast integral equation method that can solve full Maxwell's equation in 2D and 3D in seconds. The choice of models within the probabilistic ensemble-based workflow allows tweaking the DSS to the needs of different geosteering operations. Furthermore, we demonstrated that simultaneous assimilation of 'shallow' nuclear density logs enhances the look-ahead capability during geosteering (https://doi.org/10.30632/SPWLA-2021-0105). The transparent geosteering workflow requires geo-models that represent the relevant geological uncertainties and are suitable for real-time updates. For complex geological environments, we adopted generative adversarial networks (GAN) used in 'deep fakes' to parametrize, reproduce, and update realistic geology (https://doi.org/10.3997/1365-2397.fb2021051). We also extended the geo-model from the original workflow to parametrize a layered part of a 3D geo-model used for a real operation in the Barents Sea. With an ensemble of such geo-models, we can reduce the uncertainty in 3D even when using 1D measurements. For both cases the workflow can compute the probabilities of different geological objects, aiding decision support. The approximations in the geo-models and the fast modeling of measurements introduce several assumptions to work in real-time, which inevitably causes so-called model errors. Unlike the measurement errors, they cannot be addressed by standard ensemble-based update algorithms. To alleviate model errors during geosteering we introduce the flexible iterative ensemble smoother coupled with deep-neural-network-based DeepEM model. The new flexible algorithm qualitatively reproduces a state-of-the-art deterministic inversion for a real historical operation but also quantifies the uncertainties in the positions of geological boundaries in real-time.

The main outcome is the developed geosteering workflow that combines mathematics, computer science, decision theory, geology, and engineering. We also demonstrated several state-of-the-art machine learning techniques in geosteering. The research network and the ideas developed in the project contributed to establishing SFI DigiWells and other collaborations might also result in new research and innovation in the future. 19 academic papers are published (+3 evaluated) in international journals and conferences. The published research furthers computational geosciences, engineering, and decision theory. The academic software resulting from the project is used for further research by NORCE, University of Stavanger, Herriot Watt University, University of the Basque Country, Software Competence Center Hagenberg. There is a tentative plan to pilot the real-time geosteering workflow which uses deep learning and automatically accounts for modeling errors in Aker BP and Equinor in 2022.

Geosteering decisions are generally based on geological models. To make optimal decisions while drilling, it is necessary that geological models for the near wellbore region are well calibrated against measurements and that the uncertainty is quantified. The current work processes for geosteering suffer from shortcomings. (i) It is highly challenging to calibrate geomodels to deep Electromagnetic Measurements (EM) and Logging While Drilling (LWD) measurements for complex formations. (ii) There is a lack of flexibility in the current geomodels for support of structural updates and fast local updates. (iii) There is a lack of a transparent, systematic and consistent workflow for quantifying complex geological uncertainties in the geomodel, and considering them when making geosteering decisions. The primary objective of this project is to develop methodology for geosteering by continuously updating geomodel based on LWD measurements including deep EM measurements. To achieve this objective, we will address the following secondary objectives; - Ensemble-based geosteering workflow. Develop and improve an ensemble-based workflow for geosteering that would integrate seamlessly with reservoir management. - Processing deep EM measurements. Apply appropriate EM forward modelling tools and investigate efficient processing of EM measurements for integration in geosteeering workflow - Flexible earth model representation. Further develop a flexible representation of the geological model that supports local updates and uncertainty modelling of the geological structures and properties in real-time.