Rockfall Dynamics Prediction Using Data-Driven Approaches: A Lab-Scale Study

Predicting rockfall dynamics is essential for effective risk management and mitigation in mining and civil engineering, where uncontrolled rockfalls can have serious safety implications. This study explores machine learning (ML) approaches to model rockfall behavior, using experimentally derived data to predict key parameters: translational and angular velocity, coefficient of restitution (COR), and runout distance. Rockfall behavior is complex, influenced by factors such as rock shape and release angle, which create irregular, nonlinear patterns that challenge traditional modeling techniques. Three ML models—K-Nearest Neighbors (KNNs), Perceptron, and Deep Neural Networks (DNNs)—were initially tested for predictive accuracy. This study found that the Perceptron model could not capture the nonlinear intricacies of rockfall dynamics, while DNNs, though theoretically capable of handling complexity, faced issues with overfitting and interpretability due to limited data. KNNs emerged as the most effective model, offering a balance of accuracy and interpretability by using instance-based predictions to reflect localized patterns in rockfall behavior. Each parameter was modeled individually, leveraging KNNs’ strength in handling the dataset’s unique characteristics without excessive computational requirements or extensive preprocessing. The results demonstrate that KNNs effectively predicts rockfall trajectories across diverse shapes and release angles, enhancing its practical application for safety and preventive strategies. This study contributes to the understanding of rockfall mechanics by providing an interpretable, adaptable model that meets the challenges posed by small, high-dimensional datasets and complex physical interactions.