Machine learning atomic-scale stiffness in metallic glass

Abstract Due to lack of either translational or rotational symmetries at atomic-scale, predicting properties of amorphous materials from static structure is a challenging task. To circumvent the dilemma, a supervised machine-learning strategy via neural network is proposed to predict the atomic stiffness of metallic glass from discretized radial distribution function. The predicted stiffness and its spatial nature are calibrated with molecular dynamics simulations. After which, the origin of atomic constraint is interpreted via the learning structural input. Inadequacy of the model is discussed in terms of incompleteness in both machine-learning configurational space and structural descriptor.