Inverse design of glass structure with deep graph neural networks

Directly manipulating the atomic structure to achieve a specific property is long pursuit in field of materials. However, hindered by disordered, non-prototypical glass and complex interplay between property, such inverse design dauntingly hard for glasses. Here, combining two cutting-edge techniques, graph neural networks swap Monte Carlo, we develop data-driven, property-oriented route that managed improve plastic resistance Cu-Zr metallic glasses controllable way. Swap as "sampler", effectively explores landscape, networks, with high regression accuracy predicting resistance, serves "decider" guide search configuration space. Via an unconventional strengthening mechanism, geometrically ultra-stable yet energetically meta-stable state unraveled, contrary common belief higher energy, lower resistance. This demonstrates vast space can be easily overlooked conventional atomistic simulations. The data-driven structural methods optimization algorithms consolidate form toolbox, paving new way glassy