Enhancing Drilling Performance Using Real-Time Machine Learning Models: A Framework for Predicting Formation Transitions During Drilling Operations

Abstract

Accurate real-time identification of formation transitions is essential for improving drilling efficiency and reducing nonproductive time, particularly in geologically complex environments. Accurate estimates of sonic and bulk density logs are crucial inputs for subsurface formation identification; however, acquiring these logs at high resolution in real time during drilling remains a significant challenge because the sensors are typically located more than 98 ft from the drill bit. To overcome this limitation, this study proposes a machine learning-based framework to predict lithological changes using near-bit measurements. The approach leverages logging-while-drilling (LWD) data and real-time drilling parameters to enable more reliable and proactive subsurface evaluation. In this study, 55,945 datasets from fifteen wells in the Volve Field were processed and analyzed using a systematic workflow. Twelve wells provided LWD and drilling parameters, while three included wireline logs and core data. The selection of input features was initially based on proximity to the drill bit and their dominant influence on the target variable. First, lithofacies were classified using an optimized k-nearest neighbors (KNN) algorithm based on Gamma Ray (GR), interval transit time (DT), and bulk density (RHOB) logs, and then calibrated against core-derived mineralogy for formation type identification. Subsequently, an XGBoost-based classification model (XGB-C) was developed to predict formation types using the logs. To address depth and time offsets, an optimized XGBoost-based regression model (XGB-R) was also trained to predict DT and RHOB logs using real-time drilling parameters (such as ROP and WOB) and GR measurements. Data from ten wells were used for model development (80% training, 20% testing), while two additional wells served as blind validation sets. Model evaluation results indicate that the proposed framework reliably forecasts lithology and formation transitions during real-time drilling. The XGBoost-based supervised classification (XGB-C) model achieved blind-test accuracy exceeding 83%, demonstrating robustness in lithofacies prediction. For sonic and density log prediction, the optimized XGBoost-based supervised regression model (XGB-R) demonstrated superior generalization performance compared to benchmark models, including Random Forest (RF) and CatBoost (CB). The XGB-R model accurately captured depth-dependent variations in both sonic and density logs in offset wells, achieving maximum mean absolute percentage errors (MAPE) of 7.67% and 6.84%, respectively. This study presents a robust, reliable machine learning framework for real-time prediction of formation transitions, reducing reliance on subjective human judgment. The approach offers a scalable solution for intelligent lithological interpretation and can be seamlessly integrated into real-time drilling advisory systems. Its implementation improves drilling efficiency in complex environments and substantially reduces non-productive time and overall operational costs.