Predicting Oilwell Cement Properties Downhole Using Existing Databases, Modified Neural Networks and Other Machine Learning Approaches

Abstract Cementing operations are one of the most critical tasks performed during drilling operations, and a proper cementing job ensures the long-term stability of the well, and continued operations at the field. However, not a lot is understood about the physical and mechanical properties of cement over longer periods, especially in difficult or adverse downhole conditions, mainly due to a lack of representative data. Machine Learning algorithms and data driven approaches can help solve these problems and help make informed well-operations decisions. During the operational life of a well, the oilwell cement present in the borehole is exposed to a range of downhole conditions, pressures, stresses and fluids, all of which cause mechanical, physical and chemical degradation of the cement over time. While evaluating the downhole properties of cement placed in the wellbore over time is challenging, experimental studies have shown that temperature, cement formulation and the presence of corrosive formation fluids all affect the mechanical properties, particularly the compressive strength of cement (UCS). This can lead to shearing, cracks and de-bonding causing the formation of micro-annuli and migration pathways for formation fluids. Lab-based studies have shown that oilwell cement exhibits a wide range of physical properties and compressive strengths at different temperatures and after exposure to corrosive agents over time based on their formulation. Class G cements exhibited the highest compressive strengths, followed closely by Class H cements. Higher temperatures experienced downhole can lead to faster setting due to rapid dehydration, but this also resulted in a lower compressive strength failure threshold. Geopolymers and specialized cements demonstrated better stability, making them more suitable for high-temperature and acidic environments. Studies indicate that acoustic logs can help evaluate downhole conditions of cement, providing a means to correlate field and experimental data with the cement's behavior and its changing mechanical properties over time using machine learning models. Compressive strength of cement was found to have a strong correlation with acoustic data, as well as other parameters such as temperature, age and density, which can be used to assess the integrity of operational, old or abandoned wells. Based on this relationship between the compressive strength and acoustic data, machine learning models, including both traditional algorithms and deep learning approaches, have proven effective at predicting the evolution of cement mechanical properties over time. Among the deep learning algorithms, the LSTM model had the highest prediction accuracy, while a supervised random forest model showed the best fit between actual and predicted data. A thorough understanding of oilwell cement's behavior and the changes in its mechanical properties is essential for ensuring safe and efficient operations. Data-driven models that link cement's mechanical properties to various parameters such as the acoustic data can be highly effective at predicting the overall long-term integrity of the cement sheath.