Enhancing wireline formation testing with explainable machine learning: Predicting effective and non-effective stations

Abstract

Formation pressure testing acquisition is a fundamental input to reservoir management and evaluation since pressure measurement is used in several steps of wells exploration and production, such as drilling, reservoir modeling, and connectivity. The formation pressure testing campaign may represent hours to days spent in the open hole well-logging programs; especially in the deep-water reservoirs, in which major oil companies constantly seek optimization reducing the operational time, and increasing well safety. This research develops a novel explainable supervised machine learning model approach in selecting optimal stations to perform wireline formation testing (WFT), assisted by auxiliary well-log features, and evaluated by Shapley value analyses. Six supervised models were evaluated, based on the typical five well-logging programs executed in the Pre-salt carbonates, comprising thirty machine learning models implemented. Among all of them, the CatBoost model trained with 33 well-logs returned the most robust results. It attained above 0.9 of classification accuracy, the area under curve, precision, and recall when applied in the two wells data for tests. Innovative results suggested that at least 8 h of operational time could be saved by avoiding non-effective stations, showing possibilities for improvement in this data acquisition. Moreover, well-log features of nuclear magnetic resonance, density, neutron, iron, gadolinium, aluminum, and sulfur elemental yields, played an important contribution to the output prediction. Finally, predicted results lesser equal to the model's mean expected Shapley values are classified as effective, as well as predicted results greater than the model's mean expected Shapley values are classified as non-effective.