Machine-Learning Work Flow Identifies Brittle, Fracable, Producible Rock Using Drilling Data

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 201456, “Machine-Learning Work Flow To Identify Brittle, Fracable, and Producible Rock in Horizontal Wells Using Surface Drilling Data,” by Ngoc Lam Tran, SPE, Ishank Gupta, SPE, and Deepak Devegowda, SPE, University of Oklahoma, et al., prepared for the 2020 SPE Annual Technical Conference and Exhibition, originally scheduled to be held in Denver, Colorado, 5–7 October. The paper has not been peer reviewed. The complete paper demonstrates the application of an interpretable machine-learning work flow using surface drilling data to identify fracable, brittle, and productive rock intervals along horizontal laterals in the Marcellus shale. The results are supported by a thorough model-agnostic interpretation of the input/output relationships to make the model explainable to users. The methodology described here can be generalized to real-time processing of surface drilling data for optimal landing of laterals, placing of fracture stages, optimizing production, and minimizing frac hits. Modeling Approach In a multiwell field development, the ability to improve recovery efficiency per rock volume depends on well spacing, stacking, completion strategy, and avoidance of frac hits. In this study, the authors use surface drilling data to predict mechanical rock facies (generated using Poisson’s ratio and Young’s modulus) using several supervised classification methods. The objective is to use the trained model to predict the mechanical facies in real time using surface drilling data in future wells to optimize the well trajectory and placement of fracture stages. The highlighted modeling approach is conducted as follows: - Data preparation. Preliminary univariate (histograms, box plots) and bivariate analyses (cross plots) are performed, outliers removed, and missing values handled. The clean data are scaled, and a few additional derived variables such as mechanical specific energy (MSE) are added (also called feature engineering) that may boost predictive efficiency. - Unsupervised clustering. Clusters are generated using the derived variables of Young’s modulus and Poisson’s ratio. The number of clusters is optimized by silhouette width and within the sum of squares. The clusters group points of similar brittleness/fracability and are referred to as a geomechanical facies in this paper. - Supervised classification. A classifier is built using K-nearest neighbors, gradient boosting, random forests, and neural networks to identify geomechanical facies from surface drilling data. Seventy-five percent of the data are used for training and 25% for testing. Tenfold cross validation is performed on the training data to prevent overfitting. In 10-fold cross validation, the training data is subdivided randomly into 10 parts. The model is trained on nine parts and then validated on the remaining part. This process is repeated multiple times for each machine-learning technique. Only those models are averaged to provide the final model that gives good results for the validation data.