Constructing Dual-Porosity Models from High-Resolution Discrete-Fracture Models Using Deep Neural Networks

Xupeng He, Ryan Santoso, Marwa Alsinan, Hyung Kwak, Hussein Hoteit

Research output: Chapter in Book/Report/Conference proceedingConference contribution

5 Scopus citations


Detailed geological description of fractured reservoirs is typically characterized by the discrete-fracture model (DFM), in which the rock matrix and fractures are explicitly represented in the form of unstructured grids. Its high computation cost makes it infeasible for field-scale applications. Traditional flow-based and static-based methods used to upscale detailed geological DFM to reservoir simulation model suffer from, to some extent, high computation cost and low accuracy, respectively. In this paper, we present a novel deep learning-based upscaling method as an alternative to traditional methods. This work aims to build an image-to-value model based on convolutional neural network to model the nonlinear mapping between the high-resolution image of detailed DFM as input and the upscaled reservoir simulation model as output. The reservoir simulation model (herein refers to the dual-porosity model) includes the predicted fracture-fracture transmissibility linking two adjacent grid blocks and fracture-matrix transmissibility within each coarse block. The proposed upscaling workflow comprises the train-validation samples generation, convolutional neural network training-validating process, and model evaluation. We apply a two-point flux approximation (TPFA) scheme based on embedded discrete-fracture model (EDFM) to generate the datasets. We perform trial-error analysis on the coupling training-validating process to update the ratio of train-validation samples, optimize the learning rate and the network architecture. This process is applied until the trained model obtains an accuracy above 90 % for both train-validation samples. We then demonstrate its performance with the two-phase reference solutions obtained from the fine model in terms of water saturation profile and oil recovery versus PVI. Results show that the DL-based approach provides a good match with the reference solutions for both water saturation distribution and oil recovery curve. This work manifests the value of the DL-based method for the upscaling of detailed DFM to the dual-porosity model and can be extended to construct generalized dual-porosity, dual-permeability models or include more complex physics, such as capillary and gravity effects.
Original languageEnglish (US)
Title of host publicationDay 1 Tue, October 26, 2021
StatePublished - Oct 19 2021

Bibliographical note

KAUST Repository Item: Exported on 2021-10-27
Acknowledgements: We would like to thank Saudi Aramco for funding this research. We would also like to thank King Abdullah University of Science and Technology (KAUST) for providing license for MATLAB.


Dive into the research topics of 'Constructing Dual-Porosity Models from High-Resolution Discrete-Fracture Models Using Deep Neural Networks'. Together they form a unique fingerprint.

Cite this