Presenter
Hibat Errahmen Djecta from University of Stavanger and NORCE
Co-authors
Ressi Bonti Muhammad (University of Stavanger), Sergey Alyaev (NORCE), Yasaman Cheraghi (University of Stavanger), Kristian Fossum (NORCE), Reidar B. Bratvold (University of Stavanger), Apoorv Srivastava (Stanford University)
Abstract
This work showcases the development and rigorous validation of Pluralistic, an advanced automated geosteering robot designed to revolutionize well-placement operations. By integrating a Dual-Network Reinforcement Learning (RL) framework with a probabilistic interpretation approach, “Pluralistic” enhances decision-making processes, allowing for more accurate and efficient drilling outcomes. The system dynamically combines real-time data assimilation through Particle Filtering (PF) with adaptive steering adjustments driven by RL, ensuring optimal trajectory planning even in complex geological environments.
In the initial testing phase, the Pluralistic robot employed a simple reinforcement learning algorithm within a synthetic environment modeled after the setup of the unconventional round of the ROGII Geosteering World Cup (GWC) 2021. During this phase, the robot demonstrated its capability to make rapid, near-instantaneous decisions, operating within approximately 4 seconds – well below the time constraints of real-time drilling operations. In GWC 2021, the Pluralistic robot achieved an impressive reservoir contact rate of 77.3%, ranking it among the top 14% of participants of the unconventional round. This testing validates the Pluralistic approach as a capable technique for geosteering under high uncertainty in limited-data environments, which performs on par with and sometimes exceeds human experts.
Building on these successful initial results, we enhanced the Pluralistic robot with a more sophisticated Dual-Network RL framework. This upgraded system incorporates uncertainty-aware exploration strategies, allowing the robot to respond better to subsurface conditions’ dynamic and often unpredictable nature. We expect that these improvements should lead to advancements in the precision and reliability of well-placement decisions, positioning “Pluralistic” as a potentially strong contender in competitive geosteering.
The upcoming participation of the Pluralistic robot in the GWC 2024 will serve as a critical real-world benchmark for its capabilities. This event will test the robot against the current level of challenges, providing a definitive assessment of its performance in a competitive, high-stakes, unknown environment. The results from this participation will further solidify the robot’s role in advancing the field of assisted and automated geosteering, potentially setting new benchmarks for future technological innovations in the industry.
Biography
I am a Ph.D. student and research fellow at NORCE in Bergen, specializing in geosteering and data-driven decision-making. I hold a master’s degree in Artificial Intelligence and Data Systems from the University of Paris PSL-Dauphine, with a strong background in machine learning, deep learning, and data science. My experience includes working as a data scientist and machine learning engineer on projects in Earth observation, IoT networks, and adversarial machine learning. I am proficient in Python, PyTorch, TensorFlow, and various AI disciplines, and I am committed to advancing AI technology in the energy sector.