Efficient and accurate simulation of vitrification in multicomponent metallic liquids with neural network potentials

Constructing an accurate interatomic potential and overcoming the exponential growth of structural equilibration time are challenges for atomistic investigations of the composition-dependent structure and dynamics during the vitrification process of deeply supercooled multicomponent metallic liquids. In this work, we describe a state-of-the-art strategy to address these challenges simultaneously. In the case of the representative Zr–Cu–Al system, in combination with a general algorithm for effectively and accurately generating the neural network potentials (NNPs) of multicomponent metallic glasses, we propose a highly efficient atom-swapping hybrid Monte Carlo (SHMC) algorithm for accelerating the thermodynamic equilibration of deeply supercooled liquids. Extensive calculations demonstrate that the newly developed NNP faithfully reproduces the phase stabilities and structural characteristics obtained from ab initio calculations and experiments. In the combined NNP-SHMC algorithm, the structure equilibration time at deeply supercooled temperatures is accelerated by at least five orders of magnitude, and the quenched glassy samples exhibit comparable stability to those prepared in the laboratory. Our results pave the way for next-generation studies of the vitrification process and, thereby, the composition-dependent glass-forming ability and physical properties of multicomponent metallic glasses.