Machine Learning Assists Bit Selection and Optimization

Abstract

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper IPTC 21299, “Drilling in the Digital Age: Machine-Learning-Assisted Bit Selection and Optimization,” by Peter Batruny, SPE, Hafiz Zubir, and Pete Slagel, SPE, Petronas, et al. The paper has not been peer reviewed. Copyright 2021 International Petroleum Technology Conference. Reproduced by permission. _ The project outlined in the complete paper describes machine learning as a powerful tool for bit selection and parameter optimization to improve drilling performance. Machine learning will become a significant part of well planning, design, and operations in the future. The study demonstrates how artificial neural networks (ANNs) can be used to learn from previous operations and influence planning decisions to improve bit performance. Introduction Multiple wells have been drilled in an onshore field in Iraq using different bit designs and with a variety of downhole conditions. To improve the rate of penetration (ROP) in a significant manner, a radical shift in how drill bits are selected, as well as a closer look at bit characteristics, is needed. This work aims to combine analytical and physical models to examine aspects of bit designs previously overlooked. Two mathematical models traditionally have been used for bit selection, the cost per foot (CPF) model and the mechanical specific energy (MSE) model. This work concludes that MSE is more effective than CPF in optimizing ROP and is a valid benchmark for comparing bits. MSE can be used to compare previous bit runs but is limited in predicting ROP for future bit runs. Equipment and Processes Data governance consists of data collection, processing, cleaning, and compilation. Data are collected from American Standard Code for Information Interchange logs in real time and then are processed by removing missing values. The data sets are split by individual bits, with each bit characterized by data compiled from multiple wells in which the bit was run. Throughout this paper, ROP will refer to instantaneous on-bottom ROP. Given the success that ANNs have had, and continue to have, in predicting ROP in the literature, they are used for this work. Once a model is established based on statistical and physical performance, it is used as a base model to run optimization analysis for each bit. This paper examines two methods of optimization, the first being sensitivity analysis based on average ROP. The second optimization method used in this study is a Monte Carlo-style optimization. Once optimal parameters are generated from the sensitivity analysis and Monte Carlo models, they are plotted against the modeled values of offset ROP. This acts as a further verification of the optimal parameters to ensure they are within range of actual historical parameters. The optimal values are then implemented in future wells.