A Novel Deep Learning Based Method for Automated Multi-scale Discrete Fracture Network Generation

Rahul Prabhakaran and Giovanni Bertotti and David Smeulders. ( 2019 )
in: 2019 Ring Meeting, ASGA

Abstract

Modeling of naturally fractured reservoirs (NFRs) using Discrete Fracture Network (DFN) approaches, in which spatial organization of fractures is represented explicitly within the porous continuum, have become increasingly popular and relevant in NFR modeling. Fracture patterns within the subsurface are largely unknown owing to their sub-seismic size and the relative scarcity of quantitative data. Current industry practices heavily rely upon the use of averaged fracture statistics derived from borehole data or outcrop scanline measurements that are extrapolated using Poisson process statistical models to stationary DFNs that are distributed over geo-cellular models. Fractures traced from Digital Outcrop Models (DOMs) based on UAV photogrammetry are an alternative source of data that includes additional information on fracture architecture such as length scaling relationships, orientation, topology, and intensity. We present a novel method for DFN extrapolation that is based on Graphical Convolutional Networks (GCNs) for the generation of DFNs from deterministic outcrop DFNs. Convolutional Neural Networks (CNNs) have been instrumental in the development of deep learning algorithms that surpass human level efficiency in various image processing tasks such as object classification, image segmentation, and object detection. One of the limitations of CNN is that they can only be trained using Euclidean data such as images. On the other hand, GCNs are a class of deep learning techniques that can learn from non-Euclidean data such as graph representations and point clouds. Fracture patterns that are accurately traced from outcrop photogrammetry are in the form of complex graph structures with nodes and edges and hence are well suited to train using GCNs. We showcase examples of a GCN trained on different types of outcrop fracture pattern graphs in multiple lithologies and varying tectonic settings to highlight the applicability of GCN based DFN generation to model reservoir scale, non-stationary fracture networks. The GCN generated fracture networks are compared with commercially available stochastic DFN generators with respect to fracture statistical parameters such as fracture intensity, topology and dynamic parameters such as uid pressure response. The proposed new technique is well poised to take advantage of the multitude of graphical fracture datasets that are easily generated from UAV photogrammetry and LIDAR point clouds.

Download / Links

    BibTeX Reference

    @INPROCEEDINGS{PrabhakaranRM2019,
        author = { Prabhakaran, Rahul and Bertotti, Giovanni and Smeulders, David },
         title = { A Novel Deep Learning Based Method for Automated Multi-scale Discrete Fracture Network Generation },
     booktitle = { 2019 Ring Meeting },
          year = { 2019 },
     publisher = { ASGA },
      abstract = { Modeling of naturally fractured reservoirs (NFRs) using Discrete Fracture Network (DFN) approaches, in which spatial organization of fractures is represented explicitly within the porous continuum, have become increasingly popular and relevant in NFR modeling. Fracture patterns within the subsurface are largely unknown owing to their sub-seismic size and the relative scarcity of quantitative data. Current industry practices heavily rely upon the use of averaged fracture statistics derived from borehole data or outcrop scanline measurements that are extrapolated using Poisson process statistical models to stationary DFNs that are distributed over geo-cellular models. Fractures traced from Digital Outcrop Models (DOMs) based on UAV photogrammetry are an alternative source of data that includes additional information on fracture architecture such as length scaling relationships, orientation, topology, and intensity. We present a novel method for DFN extrapolation that is based on Graphical Convolutional Networks (GCNs) for the generation of DFNs from deterministic outcrop DFNs. Convolutional Neural Networks (CNNs) have been instrumental in the development of deep learning algorithms that surpass human level efficiency in various image processing tasks such as object classification, image segmentation, and object detection. One of the limitations of CNN is that they can only be trained using Euclidean data such as images. On the other hand, GCNs are a class of deep learning techniques that can learn from non-Euclidean data such as graph representations and point clouds. Fracture patterns that are accurately traced from outcrop photogrammetry are in the form of complex graph structures with nodes and edges and hence are well suited to train using GCNs. We showcase examples of a GCN trained on different types of outcrop fracture pattern graphs in multiple lithologies and varying tectonic settings to highlight the applicability of GCN based DFN generation to model reservoir scale, non-stationary fracture networks. The GCN generated fracture networks are compared with commercially available stochastic DFN generators with respect to fracture statistical parameters such as fracture intensity, topology and dynamic parameters such as uid pressure response. The proposed new technique is well poised to take advantage of the multitude of graphical fracture datasets that are easily generated from UAV photogrammetry and LIDAR point clouds. }
    }