Design of new types of rockfall barriers: from sensitivity analysis to machine learning

Landslides, especially rockfalls, present a major threat to the safety of infrastructure and communities, making it crucial to develop effective protective measures. Hybrid rockfall barriers, known as attenuators, present an intriguing option in this context as they merge the capacity to absorb impact energy with the regulation of block trajectories. However, many factors, such as geomorphological, topographical and vegetation characteristics, as well as dynamic effects, such as earthquakes and other triggering factors, influence the design of such systems, making it complex. In order to simplify and generalize the design process, this thesis aims to identify the most important parameters for rockfall analysis. The block volume, velocity and direction of impact were selected through a literature analysis. Subsequently, combining these variables, a parametric analysis was conducted using Abaqus/CAE software. The results showed that some parameters have a greater impact than others, suggesting the need for a more sophisticated method to find correlations. For this reason, a machine learning algorithm was used, which enabled a global parametric analysis to be performed. In particular, the algorithm includes - detecting correlations between parameters using Pearson's coefficient and studying interactions at the moment of impact; - using the results extrapolated from Abaqus/CAE, modelling the output values of the simulations using a deep learning model based on LSTM networks. In this context, incorporating artificial intelligence represents a significant innovation as it enhances predictive abilities and streamlines design processes, decreasing the number of necessary simulations and yielding more dependable outcomes, ultimately facilitating the creation of more effective guidelines for landslide protection. Additional research efforts may concentrate on enhancing the machine learning algorithm, broadening the dataset with intricate simulations and experimental verifications, to further boost the dependability of predictive models.