Identifying Trends and Predicting Mechanical Properties in Large Scale Cement Datasets with Machine Learning

Abstract

Abstract With the growing demands of challenging well construction operations in the oil and gas industry, cementing operations have become increasingly important. While oilwell cement properties in the short term are largely understood, longer term properties are largely ignored due to difficulties in measuring them. This is problematic because the lifetime of oilwells has grown as technology has improved, with some wells experiencing decades of life. Several of these physical and mechanical properties are dependent on the formulation of the cement – especially the composition, water content, curing conditions as well as conditions downhole in the wellbore. Using limited data available from experimental evaluations, it is possible to evaluate these properties longer term using machine learning approaches, as well as identify possible patterns in the dataset. This paper tests this by subjecting a dataset of representative cement properties which were collected from previous experimental evaluations to different machine learning algorithms such as K-Means and Support Vector Machines (SVM) to create a predictive model. Although there is a lot of work being done on machine learning and evaluating cement characteristics and properties, a lot of it is focused on the construction industry, with little work focusing on oilwell cement. Use of clustering and predictive algorithms can help solve and classify data in real-world oil and gas applications when a large amount of unlabeled field data pertaining to cements is available. The dataset used for the machine learning evaluations comprised of laboratory testing results of over 1100 distinct samples of Class G, H, and C cement, of different formulations and aged for periods ranging from a few days to several months and cured at 25 and 75 degrees Celsius. Among the mechanical and physical properties measures, of note were the densities, unconfined compressive strengths (UCS), pulse velocities (UPV) as well as physical dimensions of the samples. While generating the ML model, the dataset is split into two groups, with 30% of the datapoints used as a validation subset. Once the models are trained and tested, blind analysis is performed to determine possible trends in the cement types, as well as possibly predict the UCS using the available data. Given the availability of sufficient datapoints, machine learning techniques demonstrate promise in properly estimating cement's UCS as well as identifying broad trends in the formulation of the cement samples. When using the K-Means algorithm to identify trends in the cement dataset, the model correctly classified the available datapoints into five separate classes – each corresponding to the class of cement used, as well as the ageing period of the samples. The accuracy of the clustering was verified using blind data as well as by using a K-Nearest Neighbor algorithm to determine the accuracy metrics. UCS of samples was also reliably estimated using the SVM model, which showed excellent error margins and R2 values between actual and predicted datapoints. Optimal analysis of properties for any cement slurry will come from a combination of these approaches and computing the statistical confidence of all predicted datapoints.