# Machine learning for graph-based representations of three-dimensional discrete fracture networks

**Authors:** Manuel Valera, Zhengyang Guo, Priscilla Kelly, Sean Matz, Vito Adrian Cantu, Allon G. Percus, Jeffrey D. Hyman, Gowri Srinivasan, Hari S. Viswanathan  
**arXiv:** [1705.09866](https://arxiv.org/abs/1705.09866)  
**Date:** 2017-05-27  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1705.09866v4/

## Key points

- Manuel Valera and colleagues study the role of structural and topological information in modeling flow and transport through fractured rock in the subsurface.
- They focus on Discrete Fracture Network (DFN) computational suites like dfnWorks for simulating flow and transport in porous media.
- A small backbone of fractures exists within these networks, where most flow and transport occur.
- By restricting the flowing fracture network to this backbone, a significant reduction in the network's effective size is achieved.
- Valera et al. develop an alternative network reduction approach combining graph theoretical principles with machine learning methods.
- Fractures are represented as nodes and their intersections as edges in a graph structure.
- Random forest and support vector machines are used to rapidly identify a subnetwork capturing flow patterns of the full DFN, primarily relying on node centrality features for predictions.
- Their supervised learning techniques trained on particle-tracking backbone paths by dfnWorks generate accurate results quickly compared to traditional simulations.
- The method can reduce the network to approximately 20% of its original size while producing consistent breakthrough curves with those from the full DFN model.
- This innovative approach streamlines computational processes and enhances understanding of transport phenomena in complex fractured rock systems.

## Summary

In the study "Machine learning for graph-based representations of three-dimensional discrete fracture networks," Manuel Valera and colleagues explore the role of structural and topological information in modeling flow and transport through fractured rock in the subsurface. They focus on Discrete Fracture Network (DFN) computational suites like dfnWorks, which are designed to simulate flow and transport in porous media. The researchers find that a small backbone of fractures exists within these networks, where most flow and transport occur. By restricting the flowing fracture network to this backbone, they achieve a significant reduction in the network's effective size. However, determining this reduction through particle tracking simulations is computationally intensive, making it impractical for large systems or robust uncertainty quantification of fracture networks. To address this challenge, Valera et al. develop an alternative network reduction approach that combines graph theoretical principles with machine learning methods. They represent fractures as nodes and their intersections as edges in a graph structure. Using random forest and support vector machines, the researchers rapidly identify a subnetwork that captures the flow patterns of the full DFN. Their approach primarily relies on node centrality features in the graph to make predictions. By training their supervised learning techniques on particle-tracking backbone paths identified by dfnWorks, they are able to generate accurate results in negligible time compared to traditional simulations. The team's findings demonstrate that their method can reduce the network to approximately 20% of its original size while still producing breakthrough curves consistent with those generated by the full DFN model. This innovative approach not only streamlines computational processes but also enhances understanding of transport phenomena in complex fractured rock systems. The research published in Computational Geosciences sheds light on how machine learning can be leveraged to optimize modeling strategies for studying subsurface flow and transport dynamics within fractured rock formations.