Machine-learning inspired density-fluctuation model of local structural instability in metallic glasses

Although significant progress has been made in predicting the local structural instability of various disordered systems through the development of supervised machine learning (ML) models, the generalization and interpretability of these models remain to be addressed. Based on the systematic analysis, we find that there is a significant correlation between the weighting function of the ML prediction model and the radial distribution function (RDF) of the predicted system. We further propose a density-fluctuation model with the radial symmetry function of each atom as the local structural descriptor and the modified global RDF as the weighting function. Unlike the recent ML prediction models, which need to utilize dynamic information as a monitoring signal, this model is only based on static structure information and has a good generalization ability. The model can provide a more reliable prediction ability of structural instability for different MG systems than other widely-investigated structural parameters. Our findings reveal the density-fluctuation features of the local structural instability and shed some light on the structure-property relations in disordered materials.