🤖 AI

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

This paper proposes a geology-driven machine learning framework that uses parametric discrete fracture network models to generate synthetic training data, enabling effective automated rock joint trace mapping through a combination of mixed training and fine-tuning on real-world datasets.

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

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a professional puzzle solver, but instead of a cardboard puzzle, someone hands you a massive, jagged mountain face and asks you to find every single crack and seam in the rock. This is what geotechnical engineers do to ensure that rock walls (like those in tunnels or highway cuts) don't collapse.

The problem? Doing this by hand is incredibly slow, exhausting, and prone to human error. You might miss a tiny crack, or your eyes might get tired and start seeing "ghost" cracks.

This paper describes a way to teach a computer to do this job using a clever "work smarter, not harder" strategy. Here is the breakdown:

1. The Problem: The "Rare Photo" Dilemma

To teach an AI (like a digital brain) to recognize cracks, you usually need thousands of photos of real cracks, all carefully labeled by experts.

Think of it like teaching a child to recognize a rare breed of bird. If you only have five photos of that bird, the child won't learn well. But in geology, getting those "perfect" photos is hard because real-world rock is messy, lighting is bad, and experts often disagree on where a crack actually starts or ends.

2. The Solution: The "Video Game" Training Ground

Instead of hunting for real photos, the researchers decided to build a digital universe.

They used mathematical models (called DFNs) to create "synthetic" rock images. Imagine playing a high-end video game like Minecraft or Skyrim, but instead of just making pretty mountains, the computer follows strict geological rules. It knows exactly where every crack is, how deep it goes, and how it connects to others.

Because the computer "built" this world, it has a "God Mode" cheat sheet. It doesn't need a human to label the cracks; it already knows exactly where every single pixel of every crack is located. This gives the AI a massive, perfect textbook to study from before it ever sees a real mountain.

3. The Strategy: The "Practice vs. Reality" Approach

The researchers tested two ways to move the AI from the "video game" world to the "real world":

The "Mixed Bag" Method (SimpleMixed): This is like studying a textbook and then immediately taking a test that has both textbook questions and real-world questions mixed together. This worked great for simple things, like looking at cracks in cardboard boxes.
The "Bootcamp then Field Trip" Method (Fine-tuning): This is like sending the AI to a rigorous training camp (the synthetic world) to learn the basics, and then sending it on a field trip to real mountains. The researchers found this was much better for complex, messy real-world rock slopes.

4. The "Eye Test": Why Math Isn't Everything

Usually, scientists judge AI by a math score (like "How many pixels did you get right?"). But the researchers discovered something fascinating: sometimes the math score lies.

Imagine an AI draws a crack that is slightly too thick or a little bit wiggly. The math score might say, "Fail! You missed the exact pixel!" But a human geologist looks at it and says, "Actually, that's a great job! That looks exactly like a real crack."

The researchers argued that for engineering, "looking right" (geological plausibility) is often more important than "being pixel-perfect." They created a new way to grade the AI based on whether the cracks it finds actually make sense to a human expert.

The Big Picture

By using "fake" data to teach "real" intelligence, the researchers have created a way to build reliable tools for engineers without needing to spend years manually labeling photos. It’s like teaching a pilot in a flight simulator so that when they finally sit in a real cockpit, they already know exactly what to do.

Technical Summary: Automated Rock Joint Trace Mapping Using a Supervised Learning Model Trained on Synthetic Data

1. Problem Statement

Digital rock joint mapping is essential for structural characterization and rock mass stability analysis. However, traditional manual mapping is labor-intensive, subjective, and prone to error. While machine learning (ML) offers a scalable alternative, its effectiveness is severely hindered by several "natural data challenges":

Data Scarcity: High-quality, large-scale, annotated datasets for rock joints do not exist in the public domain.
Annotation Difficulty: Manually labeling thousands of images is costly and subjective; ground truth is often "noisy" due to varying scales, lighting, and incomplete visibility.
Class Imbalance: Joint traces occupy a very small fraction of the total pixels in an image compared to the background.
Generalization Gap: Models trained on specific sites often fail when applied to different geological settings or lighting conditions.

2. Methodology

The authors propose a geology-driven machine learning framework that utilizes synthetic data as a "geological prior" to overcome the lack of real-world training data.

A. Synthetic Data Generation (The DFN-Parametric Workflow):
Instead of using generic computer graphics or generative AI (which lack physical logic), the authors use Discrete Fracture Network (DFN) modeling.

Parametric Modeling: They use a high-dimensional parametric space to generate 27 representative block shapes (parallelepipeds) that cover various geological classifications.
DFN Integration: These shapes define joint set orientations and intensities. The workflow incorporates joint chronology (sequential fracture formation), connectivity, and surface roughness (via Perlin noise).
Rendering: Using Rhino/Grasshopper, they render 2D images with "perfect" labels (pixel-accurate masks) by simulating camera distances and lens lengths, mimicking photogrammetry surveys.

B. ML Training Strategies:
The study compares two primary training strategies across two architectures (U-Net and DeepLabv3+):

SimpleMixed: Training the model simultaneously on a combination of synthetic and real data in a single stage.
Finetune: A two-stage approach where the model is first pre-trained on synthetic data (to learn structural logic) and then fine-tuned exclusively on a small amount of real data (to adapt to real-world textures/noise).

C. Datasets:
The researchers utilized four distinct datasets: Synthetic DFN (large-scale foundation), Synthetic Box (controlled testing), Real-world Box (near-perfect labels), and Real-world Rock Slope (noisy, complex field data from Norwegian road cuts).

3. Key Contributions

Geological Prior via DFN: The primary innovation is the use of DFN to embed structural logic (connectivity, persistence, and node-type distributions) into synthetic training data, rather than just visual realism.
Hybrid Training Framework: A systematic investigation of how synthetic pre-training and fine-tuning interact with varying proportions of real-world data.
Qualitative Evaluation Framework: Recognizing that standard metrics (like IoU/Dice) can be misleading when ground truth is imperfect, the authors introduced a structured qualitative scoring system (1–5 scale) based on geological recognisability, persistence, boundary localization, and noise.

4. Results

Domain-Dependent Performance:
- In the Box domain (cleaner, simpler geometry), the SimpleMixed strategy performed best, reaching peak performance with very little real data (0–10%).
- In the Slope domain (complex, noisy labels), the Finetune strategy was superior. Pre-training on synthetic data allowed the model to learn the "rules" of joints, which it then adapted to the noisy real-world labels.
The "Metric-Quality" Gap: The study found a strong correlation between quantitative Dice scores and qualitative quality, but noted that the highest Dice score does not always equal the most useful geological map. Often, intermediate training epochs produced more continuous and geologically plausible traces than the "best" epoch according to pixel-overlap metrics.
Generalization: While zero-shot transfer (synthetic $\rightarrow$ real) is limited by visual appearance (textures/lighting), the models successfully captured the underlying geometric patterns of the joints.

5. Significance

This research provides a blueprint for applying ML to specialized engineering domains where data is scarce. It demonstrates that synthetic data should not be a substitute for real data, but a foundation. By using DFN-based synthetic images, engineers can "teach" a model the physics and geometry of rock fractures before exposing it to the messy reality of field photography. This methodology reduces the need for massive manual annotation efforts and provides a path toward reliable, automated digital rock mass characterization in mining, tunneling, and civil engineering.