A framework for upscaling and modelling fluid flow for discrete fractures using conditional generative adversarial networks

Scaling up highly heterogeneous aperture distributions of fractures into equivalent permeability tensors enables a substantial reduction in the computational cost of simulating fluid flow in fractured porous media by allowing the employment of coarser grids while keeping the accuracy of an explicit model. This work proposes the adaptation and application of conditional generative adversarial networks (CGAN) for upscaling the permeability of single fractures. Three different types of aperture distributions are used as input in this work: layered media, Zinn & Harvey transformations and self-affine fractals. As output, the model predicts the pressure inside the fracture which is used for calculation of the equivalent permeability tensor. Our results show that the framework employing CGAN provides equivalent tensors that can capture accurately both the permeability angle and anisotropy of discrete fractures, with a substantial reduction of the computational time when compared to traditional frameworks that rely on the numerical simulations.

Automated summary

This study proposes the adaptation and application of conditional generative adversarial networks (CGAN) for upscaling the permeability of single fractures. The results show that the framework employing CGAN provides accurate capture of both the permeability angle and anisotropy of discrete fractures, with a substantial reduction in computational time.