Integrating Mechanics and Machine Learning for Build-Up Rate Prediction

Abstract

Accurate prediction of build-up rates is critical for selecting the bottom hole assembly (BHA), controlling wellbore trajectory, and geosteering in directional drilling operations. Traditional methods, constrained by model assumptions, often fail to fully capture the complexity of subsurface conditions. This study integrates mechanical models with machine learning by using mechanical parameters as key inputs and incorporating mechanistic constraints into the loss function to enhance prediction robustness. An incremental training mechanism is employed to dynamically update the models, allowing them to adapt to evolving drilling environments. An interpretability analysis is conducted to reveal the influences of different features on prediction results. Applied to drilling data from the Junggar Basin, China, the models show significant improvement, with the mean absolute error (MAE) averaging 0.50°/30m and the maximum error (ME) averaging 3.06°/30m, representing decreases of 16% and 23%, respectively. After model updates, the average MAE dropped to 0.34°/30m, a reduction of 32%, while the R² exceeded 0.90, indicating excellent adaptation to complex drilling environments. The interpretability analysis confirmed the significance of mechanical parameters in influencing predictions, providing valuable insights for engineers in trajectory control and feature selection. This approach offers a novel perspective for build-up rate prediction and can be effectively employed to enhance the precision of trajectory control, thereby facilitating safe and efficient drilling operations.

Keywords: build-up rate prediction, mechanics-data fusion modeling, deep learning, dynamic update, real-time prediction