Hydraulics and Machine Learning Based Hybrid Approach to Cement Job Design

Abstract Cement job design involves multiple decisions. For example, fluids rheology, centralizers, plugs/darts, pipe movement, spacer volume, flow potential, displacement efficiency etc. These results in best practices for creating a tailored barrier that can survive for the life of the well. Each decision addresses a specific risk. If the decisions incur cost and additional tasks, it is important to demonstrate their need and value. Physics based cement job simulators are typically used to demonstrate how design decisions can mitigate the risks. While such physics-based simulators are useful, there is room to improve their recommendations using decisions adopted by successful historical jobs. This will result in a hybrid approach for design recommendations from both a scientific and an empirical standpoint. The current work demonstrates such a hybrid design process. It uses machine learning to identify relevant successful historical jobs and then to score their physics-based design decisions. Relevant jobs are identified from thousands of historical jobs using data clustering based on well conditions like Total Measured Depth, Total Vertical Depth, Casing Size, Mud Weight, Job Type, etc. The Design decision values of these historical jobs are proportionately weighted using predefined recommended values and total weights. A job score is calculated by adding weights of all the design decisions (i.e., best practices). The overall job score and the design decisions of an upcoming job are compared with those of relevant historical jobs. This allows engineers to identify specific design gaps and tailor the design to improve the overall score and therefore the ability to achieve operational success. While the underlying physics-based calculations provide scientific evidence, the relevant and successful historical jobs provide empirical evidence. The database of historical jobs will grow with time enabling job design tailoring using a continuously growing date base with improved executed jobs. Relevant jobs identified using data clustering comprised a healthy mix of both local and global jobs. This makes the job design learnings transferable across the globe. Extensive manual analysis has shown that the process of identifying relevant jobs is intuitive. For example, only extended reach wells are identified as relevant jobs when the well of interest is an extended reach well. These observations assert the use of data clustering to identify relevant jobs. The average job score (from thousands of jobs) is positively correlated to its complexity. For example, the average score of all land production casing jobs is greater than that of all land surface casing jobs. This aligns with the practice that a production casing cement job can be designed with a higher risk averse approach vis-à-vis a surface casing job. Within a job type, the job score increased with depth and temperature. Both these observations assert the usage of a job score-based approach to minimize risk and maximize the probability of operational success. In summary, a new cement job designed and tailored using the proposed approach has two advantages. Its design decisions are backed by both scientific and empirical evidence. It learns from a database which is a constantly growing database of continuously improving executed jobs.