Graph Neural Networks for glassy materials

In fundamental physics, a crucial and unsolved problem is that of understanding the behavior of glassy liquids. In these materials the viscosity or any other characteristic time increases very quickly (about 13 orders of magnitude) when the temperature is varied by only a few tens of percent around a characteristic temperature T_g, without any obvious change in the geometrical structure of their elementary constituents. This raises the question: is structure important to glassy dynamics? Several studies have shown that a lot of information about the dynamical behaviour of these materials is contained into the static structure, but how to extract this information is still an open question. A recent branch of research focuses on applying machine learning (ML) methods to extract information from static structure. Here we pursue this direction by combining a powerful family of ML models, Graph Neural Networks, with expert knowledge in the field to reach better understanding of these materials and develop new ML tools.