A Sequential Decision Framework for Geosteering: Trajectory Selection Using Reinforcement Learning and Multi-Objective Optimization

Abstract

Abstract Geosteering trajectory decisions are made sequentially as new logging-while-drilling data arrives, and each decision affects both the immediate trajectory and the information available for every decision that follows. Current practice handles this through manual, human-in-the-loop reinterpretation of a static geomodel at each new reading, a process with no repeatable decision criterion. This work develops a closed-loop framework that keeps the subsurface model current as new data arrives and hands trajectory selection to an explicit decision engine. Two engines are implemented and compared under identical conditions: NSGA-II, a multi-objective genetic algorithm, and a tabular Q-learning agent trained to maximize expected reward over each stretch of the well path. The framework is tested on a synthetic case with a fully known ground truth and on a real case study from well F-11B in the Hugin Formation, Volve field. Q-learning tracks its reference trajectory substantially more closely than NSGA-II in the synthetic case, cutting positional error by more than half (RMSE 19.5 m vs. 48.1 m), and remains the more accurate of the two in the Hugin case study as well (RMSE 38.6 m vs. 47.5 m). Despite using none of the real-time saturation or drainage information available to the original F-11B drilling team, both methods place the well within a percentage point of the field's own porosity exposure at first production. Because both decision engines run through the same closed-loop conditioning and post-processing pipeline, the framework extends directly to additional wells, formations, and reward formulations as more field data becomes available, offering a practical path toward broader, data-driven geosteering decision support.