Data-driven prediction of thin-film metallic glass forming ability via Bayesian classification and experimental verification in the Cu-Zr-Al system.

Identifying alloy compositions suitable for metallic glass formation from the vast and multi-dimensional space of elemental combinations remains a significant challenge. In the present study, we develop-using literature data-a Bayesian machine learning model based on Gaussian process classification for predicting the glass-forming ability (GFA) of metal-alloy compositions. The model incorporates a wide variety of descriptors computed based on the alloy compositions and physical properties of the constituent elements. With the optimal descriptor set the model achieves a prediction accuracy of 87% on an independent test set. The model shows predictive power that encompasses multiple alloy systems, and in this work we demonstrate this power by focusing on three ternary systems for which we compute GFA for multiple compositions that are not in the test set. Additionally, archetypal Cu-Zr-Al thin-film alloys of selected compositions are synthesized using magnetron sputtering and compared with the model predictions. By characterizing the as-deposited sample structures and their evolution under heat treatment, we verify that Cu-Zr-Al system tends to have better GFA when the Al content is less than 20%. The overall results presented herein show that the model captures the composition-GFA relationships well, and further provides valuable guidance for the design of a wide range of amorphous alloys.