Abstract
This paper aims to demonstrate the application of a new automatic geosteering method that combines probabilistic interpretation with artificial intelligence (AI) for look-ahead decision-making. We expand on our previous synthetic workflow by integrating the geosteering "robot" into a commercial cloud-based geosteering environment through its web application programming interface (API). We bench- mark the robot against 100 active human participants of the ROGII Geosteering World Cup (GWC) 2021.
Our automatic geosteering method combines a Reinforcement Learning (RL) algorithm with the Particle Filter (PF) method. PF continuously assimilates real-time log measurements obtained during geosteering operations, producing hundreds of most likely geology interpretations. Simultaneously, RL uses the information gathered from PF outputs to optimize steering decisions. The robot implemen- tation automatically collects the new well trajectory and logs and passes the latest data through the PF. The RL uses the most likely interpretations to balance the short- and long-term steering priorities and outputs a single recommendation that the robot sends back to the cloud.
Our combined PF and RL ("PLuRalistic") method achieves a remarkable reservoir contact percentage of approximately 80 % for thin and faulty target layers in our synthetic environments. The "PLuRalistic" robot expands this promising methodology to the commercial cloud environment. As part of our results, we provide a detailed account of the integration process to the cloud environment via the Solo Cloud Python SDK. This SDK is the conduit for retrieving real-time log measurements and delivering automated decisions, enabling a closed-loop geosteering decision-making framework for GWC and real geosteering in the future. The operation of our robot significantly surpasses real-time operation requirements, making one steering decision in approximately 4 seconds, far below the two-minute-per- stand drilling time allocated for the GWC. With the adjustments of the robot to pre-drill geology and GWC operational constraints, it managed to achieve 74.8% percent reservoir contact and top-quartile performance among human geosteerers.
The fully automated decision-making robot represents a radical innovation in geosteering workflows. High-fidelity simulation of the GWC gives a unique opportunity to verify and improve the AI technology. More importantly, the simulated environment with tools familiar to experts allows testing and improving user-system interaction. In particular, we utilize population data from experts for the proposal distribution of geology for the PF and evaluation of the decisions generated by RL.
