Machine Learning for Rockfall Analysis

Rockfalls represent a major hazard in open cut mining operations, threatening personnel, equipment, and critical infrastructure at the toe of highwalls, with potentially severe safety and financial consequences. Accurate assessment of rockfall hazard relies on predicting key indicators such as the energy and distance at first impact on the floor and the final run-out distance. However, these indicators depend on site-specific coefficients of restitution that are difficult to determine and highly variable in nature. Full-scale rockfall testing to estimate such parameters is expensive and time-consuming. This research aims to overcome these challenges by developing machine learning (ML) models capable of predicting rockfall hazard indicators directly from slope geometry and stratigraphy, thereby reducing the need for bespoke stochastic rockfall simulations.

Stage 1 of this project demonstrated the feasibility of using ML to provide fast and reliable predictions based on geometric features extracted from 3D photogrammetric models.

Stage 2, which builds on these findings, incorporates the influence of geological stratigraphy into the predictive framework. To achieve this, a custom variational autoencoder (VAE) architecture was developed to extract latent features (i.e. hidden features that cannot be observed) that capture both the geometry and material layering of slope profiles. These features, combined with those used in Stage 1, were then employed in state-of-the-art regression models to predict rockfall hazard indicators with substantially improved accuracy. The results show an increase in predictive performance compared to Stage 1. The trained models were rigorously validated through multiple cross-validation strategies and by comparison with field observations, consistently demonstrating strong generalisation capability. Finally, the approach was applied to generate site-specific maps of rockfall hazard indicators, enabling rapid and reliable hazard assessment at the base of highwalls.

The developed framework represents a significant step towards integrating data-driven intelligence with traditional rockfall modelling, providing an efficient and practical tool for mine operators and geotechnical engineers.