Machine Learning Workflow for the Determination of Hole Cleaning Conditions

Abstract

Abstract With the proliferation of directional and horizontal wells, modeling the mechanics of cuttings transport from its downhole origin to the shale shaker is essential in drilling operations since problems observed in vertical trajectories tend to become more complex when encountered in deviated sections. A good example is the efficacy of hole cleaning operations, a straightforward affair in vertical wells. However, when the well's orientation is inclined, it quickly becomes a complex interplay of fluid rheology, cutting bed features, wellbore properties, and in-situ drilling hardware. Inefficient hole cleaning can lead to problems such as reduced rate of penetration (ROP), stuck pipe due to cuttings buildup, high torque, bit balling, excessive equivalent circulating density, and poor cementing. This is why efficient cuttings removal remains a crucial strategy in optimizing the economics of drilling operations. Using experimental data from Yu et al. (2007) at the University of Tulsa, this work set out to model cuttings concentration from fluid rheology and drilling parameters using several machine learning (ML) techniques accessible as open-source packages in the Python environment. Fluid density, yield point, plastic viscosity, flow rate, temperature, inclination, hole eccentricity, pipe rotation, and ROP were all controllable parameters in this experiment. Our exploratory analysis sought to understand relationships among these drilling parameters and the degree of correlation between these features and the objective. Pearson's coefficient and the Gini impurity coefficient values suggest that fluid density, flow rate, and pipe rotation were the focal variables influencing our predictions of the concentration of cuttings. The investigated algorithms included ridge regression, support vector machines, several ensemble approaches, and neural networks. Several base models were created in this study using default parameters to establish baseline performances that hyperparameter tuning approaches such as RandomizedSearchCV and GridSearchCV could improve. Mean absolute error (MAE) and correlation coefficient (R) score served as performance metrics to evaluate model performance, with lower MAE and higher R values indicating superior performance. The Random Forest, Gradient Boosting, Adaptive Boosting, and a stacked regression model of the former three models showcase the workflow implemented in this study since they exhibited the most reliable performance across the training and test datasets. When applied to the entire dataset, the stacked model had the best performance, with an MAE of 2.13% cuttings concentration and a correlation coefficient of 0.946, demonstrating the ability of machine learning models to infer wellbore cuttings concentration reliably. Compared to the performance of empirical and machine learning models in the literature surrounding this work, this demonstrates an improvement in our ability to model downhole hole cleaning efficiency vis-a-vis cutting concentrations. Predictions from the model can help the drilling engineer make informed decisions about drilling fluid programs by allowing for a quick and accurate evaluation of hole cleaning conditions.