Only one of twelve bolts held: Researcher wants artificial intelligence to secure Norwegian rock cuttings

Mapping with drones and synthetic training data

Traditionally, geologists map rock faces by measuring fractures manually with a geological compass – an instrument used to measure the inclination and direction of a geological surface. Chiu instead uses remote-sensing technologies such as laser scanning and drone photogrammetry to collect data. In this way, she builds a point cloud of the rock cutting: a detailed representation of the surface consisting of millions of georeferenced points.

“It is common to use drones on construction sites. Still, not a great deal has been done when it comes to actually exploiting the high-resolution point cloud for analysis and as a basis for rock anchoring,” Chiu observes.

Once the point cloud model has been created, the machines must learn to recognise the underlying fractures. This normally requires thousands of manually labelled images. Because manual labelling is time-consuming and often produces varying results across geologists, Chiu has developed a method that generates virtual rock faces and simulates fracture networks. In this perfect, virtual simulator, the machine-learning algorithm trains itself to understand the fundamental geometry of the rock mass. The algorithm is then fine-tuned by mixing in a small proportion of real images from the actual cutting, so that the machine learns to recognise fractures in the real, complex world.

“The synthetic data we create is not realistic. But even though they look so unrealistic, they gain value when applied to actual rock cuttings,” she explains.

Probability calculations and block movements

Once the system has mapped the fractures, it analyses their intersections. It is when a minimum of three fractures and the surface meet that they carve out and form a three-dimensional network of loose rock blocks. After the puzzle has been mapped, the system calculates the rock-cutting stability. Traditionally, engineers have used deterministic methods with fixed safety margins to compensate for rock uncertainty. Chiu uses probabilistic methods, in which thousands of data simulations calculate the actual probability of rock failure.

“This method prevents over-design of the anchoring and provides a more precise picture of safety,” she says.

Chiu also extends established geological methods to calculate exactly how the rock blocks interact mechanically. In a rock cutting, the blocks function like pieces in a vertical puzzle. If a single key block at the bottom begins to slide, the support and foundation for the blocks above disappear. The result is a dangerous chain reaction in which thrust forces are transferred from block to block until a larger slide is triggered.

To anticipate and prevent this domino effect, the system simulates the rock mass under different levels of weathering – from fresh rock to fully weathered rock, where moisture and wear in the fractures act as a lubricant – and calculates which key blocks will trigger slides over time. In this way, the system finds the exact positions where bolts must be placed to stop the entire chain reaction, rather than bolting randomly across the face.

“The main goal here is to convert complex raw data into a finished and precise bolt design. The consultant can then give the contractor direct, precise instructions on exactly where each bolt is to be installed in the rock face, the researcher elaborates.”

Optimal positioning

Anchoring rock cuttings is about finding the right balance: too many bolts means unnecessary steel consumption; too few can, in the worst case, cost lives.

“It is a classic trade-off between safety, material costs, and installation time,” says Chiu.

Here she uses evolution-inspired algorithms – artificial intelligence that systematically searches for what mathematicians call Pareto-optimal solutions: a golden mean in which maximum safety, minimum steel consumption, and the fastest possible installation are weighed against one another. The algorithms test different combinations of bolt lengths and angles and evaluate hundreds of alternatives to find the best positions. Using artificial intelligence, this takes less than ten minutes – a task that might otherwise take an engineer more than two hours.

“We found that the solution closely resembled the way a human would have placed the bolts. But we also achieved lower material consumption and a more robust solution,” says Chiu.

The rock's digital twin preserves the documentation

The final element of the research project concerns long-term information storage. Chiu uses building information modelling to establish a digital twin of the rock cutting – in practice, a virtual copy in which each bolt is recorded with its length, type, and position in an open file format.

“Some information is fixed, but some must be dynamic and updated when new measurements are obtained,” says Chiu.

The idea is straightforward: if an engineer is to inspect the rock cutting in thirty years, all the necessary information must be available in the digital twin. By combining artificial intelligence and three-dimensional geometric models, the framework replaces manual assumptions with precise mathematical calculations. The result is reduced material use and greater predictability – both during construction and throughout the entire maintenance cycle.

“The goal is for this to be used in real projects to build more safely, more efficiently, and more economically,” Jessica Ka Yi Chiu concludes.