Deep learning of simulated glassy dynamics 

Glass is physically defined as a solid with random structure (or as a liquid that exhibits extremely slow structural relaxation). Thus it has remained difficult to predict the dynamics from the initial random structure, but there have been many signs that indicate underlying correlations in between. Recently deep learning methodologies were revealed to offer unprecedented  accuracy in such prediction tasks, and therefore may open new paths for AI-assisted simulations and analyses.In my talk, I will introduce recent successful applications of deep learning techniques, including our GNN model, "BOnd TArgeting Network (BOTAN)", and show extensive and comparative "benchmarks" (conducted through an international joint research). Further, some ideas will be highlighted for facilitating development of new simulation techniques and understanding of the physics of glasses.