Colloquia

Summary

Over the past years, the interest to understand avalanches and granular flows has increased. Many studies
have been done using Discrete Element Methods and experimental setups of particle beds in rotating drums
and tilted beds. The problem is that many simulations used for the dynamics of avalanches are slow and that
simulations are restricted in the number of simulated particles and time steps. The experimental approach of
monitoring avalanches created in rotating drums, takes up a lot of memory space and human knowledge to
process every individual data set. This study focuses on the implementation of machine learning models for
the prediction of the dynamic behavior of avalanches. The goal of this research is to determine if machine
learning methods can replace the DEM simulations and manual data processing.
This study starts with the detection of the avalanche slope angle of the particles in a rotating drum. From
this, the behavior of the slope angle can be established. The slope angle over time is used to train a model that
can predict the behavior of the avalanche slope angle. The model is trained using a Long Short-Term Memory
(LSTM) algorithm. LSTMs are good in learning from sequential data and are therefore a good tool for the
prediction of spacial and temporal dynamics. The results show that the LSTMs are suitable for the prediction
of the avalanche slope angle.
This research also focuses on the detection and tracking of individual particles in a rotating drum. The
detected particles can then be linked between frames over time, in order to create tracks on which the particles
move. The particle tracks can be used to learn how individual particles behave over time. The prediction of
particle locations is again done by implementation of an LSTM algorithm. The results show that LSTMs are
capable of learning what the next particle location will be, based on the previous locations of the particle.
Finally, a model reduction by proper orthogonal decomposition (POD) of the particle tracks is investigated.
A POD method proves useful to find correlations in the lower bulk of the flowing particle bed, but one has to be
careful not to reduce details in the non-linear avalanche flows at the top of the particle bed.
The results show that the detection of individual particles without labeled data is challenging. However, the
use of LSTMs for the prediction of particle locations can be feasible when a more accurate detection method
is available. Further research is required to better combine and improve the implementation of the detection
and prediction algorithms.