A Machine Learning Based Proxy Model for the Rapid Prediction of Hydraulic Fractures

Abstract

ABSTRACT Optimizing the net present value of an unconventional asset requires understanding of an optimal number and placement of horizontal wells placed in the target formations - the stacking and spacing of horizontal wells. Such optimization relies on accurately estimating the created conductive fracture dimensions, which depend on several parameters (e.g. rock properties, stresses, pumping rate, proppant loading), for which there exists a large uncertainty. Commercially available physics-based hydraulic fracture simulators model the growth of the hydraulic fractures, providing insights about the conductive fracture dimensions. A full field-scale hydraulic fracture simulation usually takes between several minutes to hours. This prevents running the thousands of simulations needed to explore the uncertainty space by, e.g., a Monte Carlo type workflow, which would be required to consider this uncertainty while making business decisions. This work aims at replacing these time-consuming fracture simulations by a proxy model, running in only a few seconds, into the Monte Carlo workflow. The proxy model is developed by training a Machine Learning (ML) based algorithm with a dataset made out of high fidelity and physics-based fracture simulations. The advantage of such an approach is that these physics-based fracture simulations can be run at liberty during the down-time of field operations and used to build a repository of dataset for the ML model. Each input required by the physics-based simulator is given a range obtained by geologic interpretations and observations from field data spanning any possible values for different applications. A Latin-hypercube is then utilized to generate instances for the input dataset. This allows us running the different simulations in an offline mode (i.e. the fracture simulations are run way before business decisions need to be made). The trained machine learning algorithm provides results that are reasonably accurate in comparison to the physics-based simulators but with a runtime that allows for probabilistic (Monte Carlo) workflow. INTRODUCTION Since all the fracturing in oil and gas (or geothermal) wells is performed in rock formations which are usually thousands of feet below the earth's surface, it has not been possible to observe the impact of a fracturing treatment directly. Most of our understanding of the fracture behavior of the subsurface rock during the treatment is based on pressure measurements, certain assumptions, analog laboratory experiments and solid and fluid mechanics based principles involving the fracturing fluid, the rock formation, and the proppant. In addition to the reservoir characteristics like the elastic properties, stress, permeability, there are many design variables associated with a planned fracturing treatment: • Injected fluid volume and the pumping rate • Injected fluid type and its properties • Type, size and concentration of the proppant