Machine learning-based prediction of cutting concentration for evaluating hole-cleaning efficiency in directional wells

This study explores the cutting transport mechanism in horizontal and directional well drilling by examining fluid rheology, cutting bed characteristics, wellbore features, and equipment interactions under high-pressure high-temperature (HPHT) conditions. The objective is to predict cutting concentration and assess hole-cleaning efficiency using supervised machine-learning techniques. Key input parameters include fluid density, pipe rotation, flow rate, temperature, inclination, hole eccentricity, plastic viscosity, and yield point. Three models, that is, Random Forest Regression, K-Nearest Neighbor Regression, and Support Vector Regression, were employed to derive predictions based on a dataset of 116 experimental values, divided into training and testing sets with an 80%–20% split. Analysis revealed that density, pipe rotation, and flow rate significantly impacted cutting concentration. The models were evaluated using MSE, MAE, RMSE, MAPE, and the coefficient of determination (R 2 ), with nested k-fold cross-validation (k = 5 and k = 10) used to obtain stable performance estimates. Under cross-validation, Random Forest and Support Vector Regression performed comparably (R 2 = 0.6885 vs. 0.7267 at k = 5; 0.6213 vs. 0.6471 at k = 10), while Random Forest significantly outperformed K-Nearest Neighbors (paired t-test, p < 10 −5 ) and was thus retained as the preferred model for its native feature-importance interpretability and competitive predictive accuracy. These results highlight the effectiveness of machine learning in monitoring cutting concentration, enhancing operational efficiency, and improving hole-cleaning in HPHT drilling environments.