Data-driven prediction of the glass-forming ability of modeled alloys by supervised machine learning

The ability of a matter to fall into a glassy state upon cooling differs greatly among metallic alloys. It is conventionally measured by the critical cooling rate R-c , below which crystallization inevitably happens. There are a lot of factors involved in determining R(c )for an alloy, including both elemental features and alloy properties. However, the underlying physical mechanism is still far from being well understood. Therefore, the design of new metallic glasses is mainly by time- and labor-consuming trial-and-error experiments. This considerably slows down the development process of metallic glasses. Nowadays, large-scale computer simulations have been playing a significant role in understanding glass formation. Although the atomic-scale features can be well captured, the simulations themselves are constrained to a limited timescale. To overcome these issues, we propose to explore the glass-forming ability of the modeled alloys from computer simulations by supervised machine learning. We aim to gain insights into the key features determining R-c and found that the non-linear couplings of the geometrical and energetic factors are of great importance. An optimized machine learning model is then established to predict new glass formers with a timescale beyond the current simulation capability. This study will shed new light on both unveiling the glass formation mechanism and guiding new alloy design in practice.