Real-time predictive capabilities of analytical and machine learning rate of penetration (ROP) models

Abstract Real-time drilling optimization consists of selecting operational parameters that maximize a desirable measure of drilling performance. Drilling optimization efforts often aspire to improve drilling speed, commonly referred to as rate of penetration (ROP). Three operational drilling parameters may be constantly adjusted at surface to influence ROP: weight on bit (WOB), drillstring rotational speed (RPM), and drilling fluid (mud) flow rate. In the traditional, analytical approach to ROP modeling, inflexible equations relate WOB, RPM, flow rate and/or other measurable drilling parameters to ROP and empirical model coefficients are computed for each rock formation to best fit field data. Over the last decade, enhanced data acquisition technology and widespread cheap computational power have driven a surge in applications of machine learning (ML) techniques to ROP prediction. Machine learning algorithms leverage statistics to uncover relations between any prescribed inputs (features/predictors) and the quantity of interest (response). The biggest advantage of ML algorithms over analytical models is their flexibility in model form. With no set equation, ML models permit segmentation of the drilling operational parameter space. However, the increased model complexity diminishes interpretability of how an adjustment to the inputs will affect the output. There is no single ROP model applicable in every situation. This study investigates real-time predictive capabilities of analytical and ML ROP models in a continuous learning setting. Data are constantly recorded as a well is drilled, and it is postulated that ROP models can learn and adapt in real-time to become more accurate. Novel learning metrics demonstrate that machine learning models reduce test error much more effectively than analytical models with incremental training data availability. Both analytical and ML model performance improves with shortened retraining intervals, defined either by length or number of data points. Cross-validation is investigated as a methodology to select the best performing ROP model in real-time.