Performance Assessment of Conventional and Machine Learning Models for Downhole Thermal Profiling during Drilling Operations

Abstract

Bottom Pipe Temperature (BPT) refers to the dynamic interplay between circulating drilling fluids, formation heat, and operational parameters at the bottom of a well. Proper estimation of the BPT is crucial to ensuring optimal operations, given the importance of mud, high drilling efficiency, equipment integrity, and efficient downhole surveillance, particularly in deep-water drilling, where mud management options are limited. This paper examined field data from a well in the Gulf of Mexico with a total depth of approximately 8,449 ft, comprising 8,259 drilling records covering 89 drilling and formation-related parameters, including flow rate, rotary speed, mud characteristics, formation depth, and geothermal gradient. The drill string temperature ranged from 95 to 171°F. Conventional thermal estimation methods, such as the Geothermal Gradient and API Circulating Temperature models, cannot accurately model the intricate thermal behavior during drilling, resulting in low predictive performance (R² of 0.2041 and 0.2010; MAE of 22.12 °F and 22.20 °F, respectively). Conversely, after normalization, missing value imputation, and train/test splitting, the dataset was trained and validated using machine learning models, specifically the Gradient Boosting Regressor, K-Nearest Neighbors (KNN), and a deep learning network developed with TensorFlow. Gradient boosting Regressor was the most accurate (R2 = 0.9906, MAE = 1.17 °F), then KNN (R2 = 0.9853, MAE = 0.77 °F), and the deep learning model (R2 = 0.9590, MAE = 2.89 °F). Hence, machine learning solutions, especially ensemble methods, are suitable for modeling nonlinear relationships between operational and geological variables and for accurately predicting BPT in real time with the highest reliability. The application of these models in drilling processes can enhance decision-making, improve drilling effectiveness, and minimize the risk of downhole temperature uncertainty.