Enhancing Stuck Pipe Risk Detection in Exploration Wells Using Machine Learning Based Tools: A Gulf of Mexico Case Study

Abstract

Abstract We present a case study on the utilization of a machine learning (ML)-based computational tool for detecting stuck pipe risks early in live operations. The system was used in two Gulf of Mexico (GoM) wildcat exploration wells. The risk detection approach is based on a novel technology using physics-informed machine learning models to analyze real-time data and detect potential stuck pipe incidents in live operations. The ML models were pre-trained on a variety of wells from different fields. The system was designed for out-of-the-box usage, which supports operational monitoring for exploration wells without pre-training on offset well data. The methodology and the process of integrating the computational tool into live operations, and the flow of data between the tool and the drilling operation is described. Additionally, the paper delves into drilling practices that helped to prevent stuck pipe and examine specific incidents that were unavoidable. The application ran stably throughout the operations, with high uptime and few false warnings in both wells; on average, fewer than one false alert was observed per day of operations. The pre-trained models proved effective, requiring no additional training; this generalizability is an important prerequisite for utility when applied to exploration wells, where offset data may be unavailable. However, due to lack of personnel to follow up the system's outputs in real-time, the benefits were limited. The first well was drilled without stuck pipe incidents. Some sticking risk symptoms were identified during the operation, especially in a fault zone. The post-well analysis indicates that good drilling practices were enough to mitigate the risks. The drilling practices responsible for the success of the operation will be discussed. In the second well, there were stuck pipe incidents. The application provided some indications of stuck symptoms but with some limitations for how far in advance the risk could be detected. The causes of the stuck incidents, the challenges in avoiding them, and updates to the risk detection system for identifying these, will be explored. Based on the experience described in the paper, the authors will offer recommendations for optimal technology utilization both from the application's and organizational perspectives.