Predicting dynamic heterogeneity at the glass transition temperature using machine learning

Structural relaxation is strongly heterogeneous in glassy liquids. Some regions actively rearrange while other are frozen, a phenomenon called dynamic heterogeneity, which might be key to better understand the glass transition itself. An important open question is how dynamic heterogeneities can be connected to the amorphous, microscopic glass structure [1].

In this talk I will present a physics-inspired supervised machine-learning technique, called GlassMLP, to systematically search for structure-dynamics relationships and predict dynamic heterogeneity in glassy liquids [2]. I will show that the trained GlassMLP is able to predict dynamic correlation lengths over a wide range of temperatures. Using transferability in system-size geometric changes of the rearranging regions can be uncovered which strongly impact growing dynamic heterogeneity. Finally, I demonstrate that GlassMLP can be transferred to temperatures beyond the range it has been trained on but still features high predictability. Utilizing this transferability dynamic heterogeneity is analyzed at the glass transition temperature inaccessible with standard computer simulations [3].