DISTINGUISH Workflow: a New Paradigm of Dynamic Well Placement Using Generative Machine Learning

Optimization of well placement in changing conditions during drilling aims to ensure maximum return on investment while constraining risk. This real-time process of drill-bit navigation, called geosteering, is crucial for hydrocarbon extraction and emerging directional drilling applications such as geothermal energy, civil infrastructure, and CO2 storage solutions. Traditional geosteering relies on rapid, manual interpretation of Logging-While-Drilling (LWD) data, manual comparison with offset wells and pre-drill uncertainty assessments. Given the everaccelerating pacing of drilling operations, time constraints are gradually making this approach more challenging and error-prone. Thus, the geo-energy industry strives for a geosteering workflow that continually captures and updates the subsurface uncertainty.

We propose “DISTINGUISH”: a real-time, AI-driven workflow designed to transform geosteering by integrating Generative Adversarial Networks (GANs) for geological parameterization, ensemble methods for model updating, and global optimization for complex decision-making during directional drilling operations. The DISTINGUISH framework relies on offline training of a GAN model to reproduce relevant geology realizations and a Forward Neural Network (FNN) to model LWD tools’ response for a given geomodel. The online phase of the workflow starts with an ensemble of Gaussian model vectors, mapped to geomodel realizations by the GAN and then converted to synthetic LWD measurements by the FNN. The ensemble of predictions is then compared to the actual LWD data and subsequently updated by an ensemble-Kalman filter-like method. A discrete dynamic programming (DDP) algorithm, informed by GANs, enables non-convex decision optimization beyond simple decision trees and greedy choices.

The main technical contribution of this paper is the complete workflow. It features the stepwise reduction of GAN-geomodel uncertainty around and ahead of the bit, informed by real-time LWD data with a DDP-optimization-based decision support system. This iterative updating process enhances predictive models of geology ahead of drilling and ultimately leads to better steering decisions. We share an open-source workflow implementation, including a benchmarking dataset built around an outcrop-based fluvial-type reservoir model. The package and the achieved performance target of the current model shall be a starting point for further method development and workflow improvement.

DISTINGUISH workflow represents a new and different approach to digitalizing geosteering, combining advanced modelling, real-time optimization, and AI to improve subsurface navigation and optimize well placement. Its ability to update and refine uncertainty in geomodels positions it as a critical technology for the future of drilling operations across various domains, fulfilling industry expectations for enhanced decision-making workflows.