Interpretable Automated Machine Learning Workflow for Intelligent Drilling in the Petroleum Industry: Case Study on Rate of Penetration Prediction

Summary Accurate rate of penetration (ROP) prediction is essential for optimizing the drilling process and improving drilling efficiency. Traditional empirical models often struggle to capture the complex nonlinear relationships between ROP and various engineering parameters. In this study, we introduce a research workflow that integrates automatic machine learning with model interpretability. The workflow based on this collaborative prediction strategy realizes automated data processing and automatic stacking integration of multiple machine learning models, including gradient boosting machine (GBM), extremely randomized trees (XRT), distributed random forest (DRF), deep learning, and generalized linear model (GLM). This strategy leverages the strengths of each model to enhance predictive performance and robustness. The optimal model is adaptively selected after integrated validation and evaluation without human intervention, which greatly reduces the model selection and hyperparameter optimization time for ROP prediction. The results indicate that the stacked model, following 10-fold cross-validation, demonstrates the highest model prediction accuracy with a root mean squared error (RMSE) of 1.92 and an R2 value of 0.92. Remarkably, the model achieves this high performance with a training time of only 120 seconds, showcasing its efficiency and scalability. This performance was achieved after applying a series of preprocessing steps, including outlier handling, normalization, and Savitzky-Golay (SG) filtering with a window length of 7 and a polynomial order of 2. These preprocessing techniques effectively reduced noise, smoothed the data, and improved the model’s ability to capture underlying patterns, contributing to the high predictive accuracy observed. Shapley additive explanations (SHAP) provide global interpretation of the effects of drilling parameters on ROP prediction as well as local explanations of the contribution of each feature to the local prediction. In addition, the overall trend and individual differences of ROP in relation to input features are analyzed through partial dependence plots (PDPs) and individual conditional expectation (ICE) plots. This makes the predictions more convincing and interpretable. In this research data set, torque and rotary revolutions per minute (RPM) have a directly positive impact on ROP, while drilling fluid density (DFD) exhibits clear interval thresholds, beyond which it exerts a negative influence. This helps to understand from a data perspective the specific range of ROP fluctuations influenced by drilling parameters. This study provides an automated machine learning (AutoML) workflow to accurately predict ROP, enabling researchers to concentrate on practical business scenarios without extensive machine learning expertise or frequent manual interventions. It is expected to accelerate the process of intelligent transformation of the scientific research paradigm in the petroleum sector.