Application of Machine-Learning to Construct Equivalent Continuum Models from High-Resolution Discrete-Fracture Models

Xupeng He,Ryan Santoso,Hussein Hoteit
DOI: https://doi.org/10.2523/iptc-20040-ms
2020-01-13
Abstract:Modeling fluid flow in fractured media is of importance in many disciplines, including subsurface water management and petroleum reservoir engineering. Detailed geological characterization of a fractured reservoir is commonly described by a discrete-fracture model (DFM), in which the fractures and rock-matrix are explicitly represented by unstructured grid elements. Traditional static-based and flow-based upscaling methods used to generate equivalent-continuum models from DFM suffer from low accuracy and high computational cost, respectively. This work introduces a new deep-learning technique based on neural networks to accelerate upscaling of discrete-fracture models. The objective of this work is to automate the process of permeability upscaling from detailed discrete-fracture characterizations. We build an "image-to-value" model to map the nonlinear relationship between high-resolution DFM images, provided as input, and the equivalent-continuum model (output), which comprises the predicted equivalent permeability of each grid-block. The proposed upscaling workflow incorporates the generation of the training datasets, design of the neural network architecture, and the validation process. The implemented deep neural networks consist of 20 layers including 18 hidden layers. Good quality input data and suitable network structure are crucial to obtaining a successful trained model. Therefore, high-resolution simulations were used to generate the training set. A training-validating process is conducted to update of training data and to optimize the network architecture until the trained model reaches an accuracy exceeding 95%. We first verify the deep learning-based (DL-based) approach by applying it to a set of single rotating fractures with known analytical solutions. We then demonstrate it on a synthetic DFM including connected and disconnected fractures, and on a DFM from an actual outcrop. We compare the method performance to reference fully-resolved solutions and to an advanced flow-based upscaling method, referred to as a multi-boundary fracture upscaling method. In the tested cases, simulation results show that the DL-based method is computational more efficiency than the flow-based method by two orders of magnitude without significant loss of accuracy. This work demonstrates the potential of a physics-based DL approach for the upscaling of high-resolution images of fractured media. The proposed DL approach is more accurate than the static-based upscaling methods and more efficient than the flow-based upscaling methods, and therefore, has the potential to be used to improve the computational performance of CPU-intensive upscaling methods.
What problem does this paper attempt to address?