Assessment of machine-learning techniques in predicting lithofluid facies logs in hydrocarbon wells

Abstract

Wireline log interpretation is a well-exercised procedure in the oil and gas industry with all its added value from exploration to production stages. It becomes even more important when it is one of only a few available alternatives to compensate for the lack of core samples in a study of lithologic and fluid variations in a well. Yet, as with other purely expert-oriented interpretational techniques, there is always a considerable risk of subjective or technical errors. We have adopted a hybrid approach that links a machine-learning (ML) algorithm to the log interpretation procedure to solve these problems. We have applied this approach to two different hydrocarbon (HC) fields with the aim of predicting the HC-bearing units in the form of lithofluid facies logs at different well locations. The values of these logs are labels of classes that are separated based on their lithologic and fluid content characteristics. After training different MLs on the designed lithofluid facies logs, we chose a bagged-tree algorithm to predict these logs for the target wells due to its superior performance. This algorithm predicted HC units in an accurate interval (above the HC-fluid contact depth), and it showed a very low false discovery rate. The high-accuracy rate, speed of analysis, and its generalization ability, even in data-deficient cases, accentuate why including ML algorithms can improve the understanding of the subsurface at every phase of the exploration and production process. The proposed approach of using ML algorithms, trained and tuned based on the expert's knowledge of the reservoir, can be modified and applied to future wells in a HC field to significantly minimize the risk of false HC discoveries.