Machine Learning Modeling of Drill Cuttings Transport Based on Computational Fluid Dynamics Simulation Data

Abstract

ABSTRACT Drill cuttings transport is a critical operation in drilling deep wells. Inefficient removal of cuttings can lead to both hydraulic and mechanical problems during drilling operations. High‐fidelity computational fluid dynamics (CFD) simulations provide detailed insight but are computationally expensive. Alternative models often oversimplify the relationships between operational parameters and particle transport, thereby failing to predict critical quantities. These quantities include transport efficiency or cuttings concentrations. In this study, surrogate models are developed using a common data set generated by validated multiphase CFD simulations. The data set consists of 269 CFD simulations varying the Reynolds number, Taylor number, particle concentration, and particle size. Five surrogate models, namely, Decision Trees (DTs), Random Forest (RF), Extra Trees (XTRs), Extreme Gradient Boosting (XGBoost), and Gaussian process regression (GPR), were trained to predict the domain‐averaged particle velocity. The models are evaluated in terms of prediction accuracy, learning behavior with respect to training data size, robustness through bootstrap resampling, and physical consistency of the response trends. All trained models achieved very high predictive accuracy (), confirming that CFD‐based particle transport can be reliably emulated using data‐driven surrogates. The learning‐curve analysis showed that performance improves rapidly with increasing data, saturating beyond approximately 60% of the data set. Despite the statistical accuracy, tree‐based ensembles produced piecewise‐constant, nonsmooth response surfaces. However, GPR generated smoother and more physically consistent trends. Sensitivity analysis revealed that the Reynolds number dominates the system behavior, explaining more than 95% of the output variance. The high predictive accuracy observed across all surrogate models (), with most predictions within of CFD reference values, reflects the low‐dimensional character of the system, where the output is dominated by the Reynolds number. The trained models enable global sensitivity analysis through Monte Carlo simulation and real‐time prediction of cuttings velocity, supporting continuous monitoring, control, and optimization of borehole cleaning in deep drilling operations, while demonstrating how information from expensive high‐fidelity CFD simulations can be retained in fast, computationally cheap surrogates.