Optimizing Diagnostic Fracture Injection Test (DFIT) interpretation using Machine Learning (ML) methods

Abstract

Diagnostic Fracture Injection Tests (DFIT), are commonly used to derive key parameters for hydraulic fracture design and modeling. Although this process can identify properties needed for well optimization, it is also time intensive, affected by interpretation bias, and incomplete data. In this thesis, I address these adversities by applying unsupervised clustering methods: K-Means, DB-Scan, Hierarchical modeling, and Gaussian mixture models to identify point density variation that correlates to key parameters on a DFIT pressure decline. Deep Neural networks (DNN) trained using labeled DFITs are further tested for event prediction. To test these methods a variety of platforms are tested such as R-Studio Shiny Web App® to create user-friendly testing platforms and Python® for its computational ability when faced with supervised learning methods. Collectively unsupervised and supervised learning methods show significant promise in the DFIT interpretation realm.