Apples to Apples: Impartial Assessment of Drilling Technologies through Big Data and Machine Learning

Ensuring a proper apple to apple comparison is a challenge in drilling performance evaluation. When assessing the effect of a particular drilling technology, such as bit, bottomhole assembly (BHA) or mud type, on the rate of penetration (ROP) or other drilling performance criteria, all other factors must be fixed to truly isolate the effect. Traditionally, performance evaluation starts with manual identification of reasonably similar entities, such as drilling runs or well sections by means of numerous selection criteria; e.g., location, depths, inclinations, drilling conditions, tools, etc. The selected drilling performance metrics are then compared using statistical analysis techniques with various extents of thoroughness. Such analyses are laborious and are usually limited to just a handful of cases due to practical reasons and time constraints. Furthermore, the analyses are difficult to apply to large data sets of hundreds or thousands of wells, and there is always a risk of missing an important combination of factors where the effect is important. Therefore, conclusions based on these analyses may well be insufficiently justified or even confirmation biased, leading to suboptimal technical and business decisions. This paper presents a combined machine learning and statistical analysis workflow addressing these challenges. The workflow a) discovers similar entities (wells, intervals, runs) in big datasets; b) extracts subsets of similar entities (i.e., "apples") for evaluation; c) applies rigorous statistical tests to quantify the effect (mud type, BHA type, bit type) on a metric (ROP, success rate) and its statistical significance; and, finally, d) returns information on areas, sets of conditions where the effect is pronounced (or not). In the statistical analysis workflow, the user first specifies the drilling technology of interest and drilling performance metrics, and then defines factors and parameters to be fixed to better isolate the effect of the drilling technology. The historical data on thousands of entities are then preprocessed, and the entities are clustered by similarities in the multitude of factors by the k-means algorithm. Statistical tests are performed automatically on each cluster, quantifying the magnitude of technology effect on performance criteria, and calculating p-values as the measure of statistical significance of the effect. The results are presented in a series of clustering observations that summarize the effects and allow for zooming into the clusters to review drilling parameters and to perform further in-depth analysis, if necessary. All steps of the workflow are presented in this paper, including data processing details, and reasons for selecting specific clustering algorithms and statistical tests. Several examples of the successful applications of the workflow to actual drilling data for thousands of wells are provided, focusing on the effects of BHA, steering tools, and drilling muds on drilling performance. This unique approach can be used to improve other drilling performance evaluation workflows.