Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automated rock joint trace mapping using a supervised learning model trained on synthetic data generated by parametric modelling

Jessica Ka Yi Chiu, Tom Frode Hansen, Eivind Magnus Paulsen, Ole Jakob Mengshoel

TL;DR

This paper tackles the scarcity and bias of real rock-joint labels by introducing a DFN-based synthetic data workflow to train supervised rock joint trace segmentation models. By parametricly generating field-relevant joint networks and textures, the approach embeds geological priors and enables pretraining, mixed training, and fine-tuning strategies to transfer to real slope and box imagery. The study demonstrates that synthetic data can support joint trace detection, with mixed-training excelling in well-controlled box-like domains and finetuning providing robustness in noisier slope-domain labels; zero-shot transfer remains limited, underscoring the importance of domain adaptation and small-site fine-tuning. Qualitative evaluation is emphasized alongside standard metrics to capture geological usefulness, and the results motivate production workflows that couple synthetic priors with limited real data to achieve reliable joint-mapping in engineering practice.

Abstract

This paper presents a geology-driven machine learning method for automated rock joint trace mapping from images. The approach combines geological modelling, synthetic data generation, and supervised image segmentation to address limited real data and class imbalance. First, discrete fracture network models are used to generate synthetic jointed rock images at field-relevant scales via parametric modelling, preserving joint persistence, connectivity, and node-type distributions. Second, segmentation models are trained using mixed training and pretraining followed by fine-tuning on real images. The method is tested in box and slope domains using several real datasets. The results show that synthetic data can support supervised joint trace detection when real data are scarce. Mixed training performs well when real labels are consistent (e.g. box-domain), while fine-tuning is more robust when labels are noisy (e.g. slope-domain where labels can be biased, incomplete, and inconsistent). Fully zero-shot prediction from synthetic model remains limited, but useful generalisation is achieved by fine-tuning with a small number of real data. Qualitative analysis shows clearer and more geologically meaningful joint traces than indicated by quantitative metrics alone. The proposed method supports reliable joint mapping and provides a basis for further work on domain adaptation and evaluation.

Automated rock joint trace mapping using a supervised learning model trained on synthetic data generated by parametric modelling

TL;DR

This paper tackles the scarcity and bias of real rock-joint labels by introducing a DFN-based synthetic data workflow to train supervised rock joint trace segmentation models. By parametricly generating field-relevant joint networks and textures, the approach embeds geological priors and enables pretraining, mixed training, and fine-tuning strategies to transfer to real slope and box imagery. The study demonstrates that synthetic data can support joint trace detection, with mixed-training excelling in well-controlled box-like domains and finetuning providing robustness in noisier slope-domain labels; zero-shot transfer remains limited, underscoring the importance of domain adaptation and small-site fine-tuning. Qualitative evaluation is emphasized alongside standard metrics to capture geological usefulness, and the results motivate production workflows that couple synthetic priors with limited real data to achieve reliable joint-mapping in engineering practice.

Abstract

This paper presents a geology-driven machine learning method for automated rock joint trace mapping from images. The approach combines geological modelling, synthetic data generation, and supervised image segmentation to address limited real data and class imbalance. First, discrete fracture network models are used to generate synthetic jointed rock images at field-relevant scales via parametric modelling, preserving joint persistence, connectivity, and node-type distributions. Second, segmentation models are trained using mixed training and pretraining followed by fine-tuning on real images. The method is tested in box and slope domains using several real datasets. The results show that synthetic data can support supervised joint trace detection when real data are scarce. Mixed training performs well when real labels are consistent (e.g. box-domain), while fine-tuning is more robust when labels are noisy (e.g. slope-domain where labels can be biased, incomplete, and inconsistent). Fully zero-shot prediction from synthetic model remains limited, but useful generalisation is achieved by fine-tuning with a small number of real data. Qualitative analysis shows clearer and more geologically meaningful joint traces than indicated by quantitative metrics alone. The proposed method supports reliable joint mapping and provides a basis for further work on domain adaptation and evaluation.
Paper Structure (43 sections, 6 equations, 12 figures, 2 tables)

This paper contains 43 sections, 6 equations, 12 figures, 2 tables.

Table of Contents

  1. Highlights
  2. Introduction
  3. Related works
  4. Methods of creation of artifacts
  5. Recent developments in automated rock joint trace mapping
  6. Applications of machine learning to image-based rock joint trace segmentation
  7. The role of synthetic data in machine learning–based image segmentation
  8. Discrete fracture networks in rock mass modelling
  9. Methodology
  10. Datasets and dataset preparation
  11. Synthetic discrete fracture network dataset
  12. Real-world rock slope dataset
  13. Synthetic box dataset
  14. Real-world box dataset
  15. Data preprocessing
  16. ...and 28 more sections

Figures (12)

  • Figure 1: Examples of realistic rock outcrop representations available online. (a) The Verdant Realms, a personal 3D environment created by patscheider_verdant_2024. (b) Results of a video tutorial example of sculpturing a realistic rock with 3D modelling tools, modified from olson_quick_2023. (c) Synthetic rock slope image generated using ChatGPT 5.2 with the prompt “generate an image of a rock slope in granite”.
  • Figure 2: Overview and example 2D images of ML datasets used in this study.
  • Figure 3: Calculated block shape parameters of the 8192 parallelepipeds generated using a parametric study (gray circles) and the selected 27 parallelepipeds (solids in random colours) that span a wide variety of block shape classes based on the classifications (a) after Palmstrom1995 and (b) after singh_modified_2022. The same parallelepiped is plotted with the same color in both plots.
  • Figure 4: Workflow for generating the synthetic DFN dataset. Block shapes are created via a parametric study in Grasshopper and used to define joint parameters for DFN generation in FracMan. Kinematic analysis in RocSlope3 simulates block removal, and Perlin noise adds surface roughness. Joint traces are extracted, given waviness and thickness, and overlaid on textured slope surfaces. Final images and perfect label masks are rendered for ML training.
  • Figure 5: Triangular plot of the proportion of I-, X- and Y-nodes for the joint trace networks in all the 27 DFN models (after manzocchi_connectivity_2002sanderson_use_2015). ${C_{L}}$ represents the average number of connections per line. Examples of DFN models in two clusters of joint network connectivities are shown, together with the edge boundary of the slope face.
  • ...and 7 more figures