A deep learning-based surrogate model for trans-dimensional inversion of discrete fracture networks

Abstract

Fractures and their geometrical patterns are usually required to analyze the mechanical and flow properties of porous media in the subsurface. Fracture characterization is therefore regarded of crucial importance for optimizing production management or achieving maximum storage capacity. In this research, we propose to invert the fracture networks under the Bayesian framework for the uncertainty quantification. In particular, the number of fractures in the modelling system is treated as unknown, leading to a trans-dimensional inverse problem, and the reversible jump Markov chain Monte Carlo algorithm is applied to sample the model space with possible model moves proposed in the sampling process. A deep learning network is further applied as a surrogate model in the sampling process for increasing the computational efficiency, instead of using the physical forward simulator. We apply the proposed methodology to estimate the spatial distribution of fracture networks based on the head measurements from the steady-state flow simulation. The prior distributions of fracture parameters such as position, orientation and length are described using the discrete fracture networks approach that is deeply rooted in stochastic modelling. Due to the high non-uniqueness, the correct spatial distribution of fracture patterns cannot be successfully recovered in this case study, even a good match between observed and simulated head data is reached. More analysis could be performed in the future with the production historical data or more informative priors.