A New Method for Drill Cuttings Size Estimation Based on Machine Learning Technique

The information of solids content in a drilling fluid is one of the most important aspects for proper control of the drilling fluid properties and solid control equipment efficiency determination. Improper knowledge of solids size in a drilling fluid can cause adverse effects on the drilling performance and consumption of shaker screens. As a result, solids in a drilling fluid degrade in size with time and become extremely difficult to be removed by traditional means. The objective of this paper is to introduce a new method based on a machine learning approach to estimate undesirable solids (i.e., drill cuttings) size using solely operational drilling data for the correct selection of shaker screen size to maintain drilling fluid within its desired properties. The drilling data were obtained from the actual 5-7/8-inch hole section of four vertical wells and included surface drilling parameters, drilling fluid parameters, and drilling bit parameters. To obtain the drill cuttings size, we proposed to have a sampling line from the fluid return line and a drill cuttings sampling device in a drilling fluid laboratory. Additionally, a liquid particles analyzer can be used to obtain the particle size distribution of the collected drilling fluid sample at the depth of interest. To validate our proposal, synthetic data of the drill cuttings with a size of 330–410 µm, based on our experience in the area of interest, were produced to represent the drill cuttings size corresponding to the drilling data. The collected dataset was structured and split with a 70/30 training-to-testing ratio for building and assessing the models. Three machine learning models, i.e., random forest, gradient boosting, and extreme gradient boosting, were used to develop the models. Well-5 was used to conduct another experiment to further understand the performance of each model. Visual graphical analysis and statistical measurements, i.e., root mean square error, correlation coefficient, residual error, Nash–Sutcliffe efficiency, and variance account for, were used to verify the performance of the developed models.