Diagnostic pumping techniques are used routinely in proppant fracturing design. The pumping process can be time consuming; however, it yields technical confidence in treatment and productivity optimization. Recent developments in data analytics and machine learning can aid in shortening operational workflows and enhance project economics. Supervised learning was applied to an existing database to streamline the process and affect the design framework.
Five classification algorithms were used for this study. The database was constructed through heterogeneous reservoir plays from the injection/falloff outputs. The algorithms used were support vector machine, decision tree, random forest, multinomial, and XGBoost. The number of classes was sensitized to establish a balance between model accuracy and prediction granularity. Fifteen cases were developed for a comprehensive comparison. A complete machine learning framework was constructed to work through each case set along with hyperparameter tuning to maximize accuracy. After the model was finalized, an extensive field validation workflow was deployed.
The target outputs selected for the model were crosslinked fluid efficiency, total proppant mass, and maximum proppant concentration. The unsupervised clustering technique with t-SNE algorithm that was used first lacked accuracy. Supervised classification models showed better predictions. Cross-validation techniques showed an increasing trend of prediction accuracy. Feature selection was done using one-variable-at-a-time (OVAT) and a simple feature correlation study. Because the number of features and the dataset size were small, no features were eliminated from the final model building. Accuracy and F1 score calculations were used from the confusion matrix for evaluation, XGBoost showed excellent results with an accuracy of 74 to 95% for the output parameters. Fluid efficiency was categorized into three classes and yielded an accuracy of 96%. Proppant concentration and proppant mass predictions showed 77% and 86% accuracy, respectively, for the six-class case. The combination of high accuracy and fine granularity confirmed the potential application of machine learning models. The ratio of training to testing (holdout) across all cases ranged from 80:20 to 70:30. Model validations were done through an inverse problem of predicting and matching the fracture geometry and treatment pressures from the machine learning model design and the actual net pressure match. The simulations were conducted using advanced multiphysics simulations.
The advantages of this innovative design approach showed four areas of improvement: reduction in polymer consumption by 30%, reduction of the flowback time by 25%, reduction of water usage by 30%, and enhanced operational efficiency by 60 to 65%.
