Structural rationale for boson peak in metallic glass informed by an interpretable neural network model

In amorphous solids, there exists a universal phenomenon of vibrational anomaly, i.e., an excess of vibrational modes over the Debye level in the low-terahertz-frequency regime, which is termed a boson peak. Although the origin of the boson peak has been studied extensively for decades, quantitative prediction of its intensity remains elusive. In this paper, we try to quantify the intensity of the boson peak via an interpretable machine-learning strategy based on a dataset consisting of >1600 glass samples prepared with cooling rates spanning four orders of magnitude. We have attempted to extract information from a pure static structure and vibrational localization through four different feature inputs, among which the partial pair distribution function (PDF) yields the best predictive performance. It is found that the first several neighbor shells corresponding to a characteristic subnanometer length scale are important for capturing the structural genes of the boson peak in amorphous solids, regardless of the size of the glasses. Moreover, a higher boson peak relates to a more disordered atomic arrangement with lower peaks and shallower basins in the PDF. The obtained knowledge sheds light on rationalization of the boson peak in amorphous solids from pure structural information.

Automated summary

In this paper, an interpretable machine-learning strategy is used to quantify the intensity of the boson peak in amorphous solids. The study utilizes a dataset of over 1600 glass samples prepared with different cooling rates. The obtained knowledge provides insight into the rationalization of the boson peak in terms of pure structural information in amorphous solids.