A new criterion for predicting the glass-forming ability of alloys based on machine learning

Abstract In this paper, the dimensionless component weights of 37 characteristic temperatures are analyzed based on the neighborhood principal component analysis method of feature selection. Dimensionless parameters with higher weight coefficients than others are combined to form a new glass-forming ability (GFA) criterion. A criterion to represent correlation between characteristic temperature and GFA is derived by machine learning (ML) algorithmic routine as k = T g × T x × T l × ( T x - T g ) ( T l - T x ) 4 (wherein Tg is glass transition temperature, Tx is onset crystallization temperature and Tl is liquidus temperature), which exhibits correlation (coefficient of determination R 2 = 0.43 ) between criterion k with Dmax is better than other eleven criteria. The linear correlation between k and Dmax that can be expressed as: D max = ( 0.35432 ± 0.05664 ) + ( 0.16200 ± 0.00712 ) k . Finally, based on classical nucleation theory, the reliability of criterion has been analyzed, which proves the feasibility of ML in the research on GFA.