Case Study: Using Machine Learning and Ultra-Deep-Reading Resistivity for Better Reservoir Delineation

Abstract

Understanding reservoir fluid and facies distribution is crucial for optimal reservoir development. Ultra-deep, logging-while-drilling (LWD) resistivity measurements with a deep detection range into the formation have started a new chapter of formation evaluation. A hybrid inversion of statistical and deterministic approaches based on ultra-deep measurements has successfully determined formation resistivity profiles more than 100 ft away from drilled wellbores, providing proactive geosteering information for real-time well-placement decisions. However, the inversion sometimes produces artificial geological features because of so-called solution ambiguities attributable to lower measurement sensitivity in certain formation resistivity contrasts and reservoir geometries. Previously, geosteering geologists were trained to recognize such unrealistic geological structures based on multiple sources of information, rather than just the ultra-deep resistivity inversion results. This paper introduces machine-learning (ML) algorithms to evaluate the sensitivity of individual measurements, as well as to cluster the inverted models to acquire more geologically reasonable models of the surrounding formations. A case study shows significant improvement as a result of the ML algorithm in the structural consistency of the reservoirs. The boundaries were better determined with fine details using the ML algorithm, as compared to results from existing algorithms. The enhanced answer product enabled a better understanding of the formation properties surrounding the wellbore and retrieved several fine features that were not observed previously.