Application of Machine Learning Method for Modeling Settling Behavior of a Spherical Particle in Fibrous Drilling Fluids

Abstract

In the oil and gas industry, several challenges are possibly encountered during drilling operations which can lead to an increase in the non-productive time. One of these problems is the wellbore cleanout process in extended horizontal wells. Counting on the adjustment of fluid properties to robust the wellbore cleanout becomes a privileged solution while drilling in depleted and deep formations. Fibrous sweep fluid has been exploited to effectively clean the horizontal segment of the wellbore. Understanding the settling of cuttings in drilling fluid is a crucial factor for successfully preparing fibrous drilling fluid. This study aims to model the sedimentation behavior of a particle in the fibrous fluid using an artificial intelligence technique. The model will significantly assist the engineers in designing drilling fluid formulation by finding the optimum base fluid properties and fiber concentration. In this study, a total of 1012 data points of settling velocity measurement were collected from the literature. The data possess a diversity of seven input features including particle size, particle density, fluid density, fluid rheological properties, and different fiber concentrations with their corresponding measured settling velocity. The database was graphically and statically analyzed to draw insights into the dataset. Five different supervised regression machine learning algorithms (Random Forest, Support vector machine, CatBoost, Extra Tree, and Gradient Boosting) were utilized to develop a settling velocity model. In addition, the best-performing model is compared to two existing mechanistic models. The feature variable-importance analysis is implemented to identify the most crucial parameters affecting the settling velocity. The results of this study disclosed that CatBoost has a superior performance among the tested models for predicting the settling velocity. The next best accuracy is attained by the Extra Tree model. However, both models (Catboost and Extra Tree) exhibit a slight reduction in their accuracy which drops from 99% on the training set to 95% on test data sets. Moreover, Random Forest and Gradient Boosting exhibit generalized abilities and are less influenced by data outliners. The relative feature importance analysis reveals that the influence of the input feature on the settling velocity is ranked from highest to lowest as follows particle density, particle diameter, fluid consistency, fluid density, yield point, fluid behavior index, and fiber concentration. The model comparison study finds that Elgaddafi's model is more generalized compared to CatBoost and Xu's models. A new ML model for accurately predicting a particle settling velocity in the fibrous drilling fluid is presented in this study. The developed model overcomes the constraints of a numerical iteration process for the mechanistic models and the uncertainty of empirical correlations. An accurate prediction of the settling velocity leads to enhancing the wellbore cleanout in the most challenging operation.