Optimizing Unconventional Hydraulic Fracturing Design Using Machine Learning and Artificial Intelligent

Abstract For optimizing the hydraulic fracture design in shales, it is challenging to understand the impact of several different parameters on fracture propagation and production, such as geomechanical properties and fracturing treatment parameters. Current frac simulators do not exhibit consideration of the anisotropy of rock elasticity in the shales. Additionally, using the fracture simulation linked with reservoir simulation for the parametric study is low efficient. Due to its lamination nature, shale has different geomechanical properties along with the directions vertically and horizontally. Anisotropic elastic properties and stresses lead to more complications for predicting the fracture. This study introduces a comprehensive workflow for fracturing design optimization by applying supervised machine learning. The research also aims to develop an algorithm that can help any shale reservoir optimize the pumping treatment design of hydraulic fracture. The workflow is divided into six steps. Firstly, acoustic and density logs for a research well in Marcellus shale are used to interpret Young's modulus, Poisson's ratio, and minimum horizontal stress magnitude by anisotropic VTI model. In step 2, the interpreted mechanical properties, including the current treatment design of the target well, are inserted into the frac simulator to obtain the conductivity distribution inside the fracture. The conductivity distribution converts to fracture permeability matrix. As for the third step, the fracture permeability matrix is consequently entered into the reservoir model for estimating the production. The output production is matched with the field history data. For the fourth step, a random sampling algorithm is applied to build a database with a rational sample size. In step 5, the generated database is employed to train and validate an artificial neural network model (ANN). Lastly, parametric studies are performed through the trained ANN model to analyze the multi-parameter effect on cumulative production. This workflow can predict the early and late production for a given fracture design based on multiple fracture treatment parameters such as initial fracture depth, cluster numbers of each stage, and proppant type. Besides, the study provides a capability for multivariable analysis to better understand the productivity behavior of the fractured well.