Utilizing Drilling Data and Machine Learning in Real-Time Prediction of Poisson's Ratio

Abstract

Rock elastic properties influence drilling performance, estimation of in-situ stresses, and hydraulic fracturing design. Therefore, having complete and accurate information on rock properties is essential. While those properties are conventionally measured experimentally or using well logs, this work proposes to estimate the Poisson's ratio (PR) from parameters available while drilling. Various machine learning techniques were employed, such as artificial neural network (ANN), support vector machine (SVM), and random forest (RF). The dataset utilized contains more than 5800 data points, each of them has a value of PR and six drilling parameters such as rate of penetration (ROP), rotary speed (RPM), and weight on bit (WOB). The dataset was divided into three parts, two were fed to the algorithms for training and testing the models, while the last group (around half of the dataset) was hidden to be used to validate the models later. The models had a good fit with the actual PR values with correlation coefficients as high as 0.99 and errors as low as 1%. Among the used algorithms, ANN and RF yielded the best accuracy in all datasets with no significant difference between the training and the validation performance which indicate good generalization without an overfitting problem. Using drilling data to predict rock mechanical parameters allows building a complete geomechanical model at an early time. It also saves the time and cost associated with laboratory tests.