Predicting and following T1 events in dry foams from geometric features

Machine learning techniques have been recently applied in predicting deformation in amorphous materials. In this study, we extract structural features around liquid film vertices from images of flowing 2D foam and apply a multilayer perceptron to predict local yielding. We evaluate their importance in the description of the T1 events and show that a high level of predictability may be achieved using well-chosen combinations of features as the prediction data. The most relevant features are extracted by performing the predictions separately for isolated sets of features, and these findings are verified using principal component analysis. Using this approach, we determine which properties of the images are most important with regard to the physics of the processes. Our findings indicate that film lengths and angles between the liquid films joining at the vertex are the most important features that predict the local yield events. These two features describe 83% of the yield events. As an application, we extract the statistics of event waiting times from the experiment.