A Novel Artificial Intelligence-Based Cement Slurry Design Recommendation System

Abstract

Abstract Cementing is a critical stage in the well construction process, providing zonal isolation, support, and protection for casing. Cement slurry design is a key element in cementing and represents a complex chemical process. Designing slurries with numerous chemical properties is challenging as it must account for often-contradictory requirements. Conventional slurry design relies heavily on using traditional search methods, based on applying numerous filters containing slurry composition, test conditions, additives, etc. This approach is highly ineffective, tedious, and time consuming, and can lead to suboptimal results, because of an engineer’s subjective experience and biases. In this paper we present an artificial intelligence-(AI-)based cement slurry design recommendation system that recommends the most relevant slurry designs based on slurry composition, well conditions, and laboratory test results. The original database contains ~230,000 formulations and ~2 million tests completed since 2018. These data are cleaned, preprocessed, and fed into the recommendation system, which is built in two steps. First, vectorization of all historical slurry records is performed with consistent feature engineering. Key slurry features include slurry composition, well conditions, and laboratory test results. Second, these vectors are used to compute similarity metrics between slurry records applying AI-based algorithms, such as clustering, exact, and approximate nearest-neighbor methods. In the inference phase, the system uses the computed similarity metrics to recommend the most relevant slurries for a set of design requirements. The AI-based system is data-driven and objectively recommends the most relevant slurries for a set of design requirements, from the database. A comparative study of the tested AI algorithms and their corresponding outcomes is presented. Specific evaluation metrics are proposed to evaluate the recommendation results. The recommended slurries are visualized in both tabular and graphical forms, for user-friendly analysis. In addition, several examples are provided that demonstrate how this innovative approach improves the slurry design methodology. The proposed approach enables retrieval of the relevant slurry candidates from a huge database in seconds, compared to performing manual search and analysis, which can take hours. The recommended slurries form a solid foundation for later stages of the slurry design process. Furthermore, smart selection of slurries saves many hours of expensive laboratory testing.