Discovery of a new criterion for predicting glass-forming ability based on symbolic regression and artificial neural network

Abstract

Metallic glasses (MGs) are widely used in various fields due to their superior physical properties. Glass-forming ability (GFA) represents the difficulty of forming MGs. Therefore, understanding and establishing the connection between materials characteristics and GFA is a great challenge in MGs research. In this work, to generate a new criterion to characterize GFA, symbolic regression and artificial neural network (ANN) were employed built on 7795 pieces of data. A completely new criterion was proposed and revealed the relationship between three characteristic temperatures (wherein Tg is the glass transition temperature, Tx is the onset crystallization temperature, and Tl is the liquidus temperature) and GFA. The new criterion not only exhibits a higher correlation to the critical casting diameter (Dmax) than the other 11 reported criteria but also illustrates the importance of high power (Tx − Tg)/(Tl − Tx) in characterizing GFA. Moreover, to test the criterion on unreported data, three models that can, respectively, perform GFA classification, predict Dmax, and three characteristic temperatures were built through artificial neural networks. Then, 439 new data generated by the ANN model were generated by models applied on Zr–Co–Al–X (X = W, Si, and Ni) alloys. On the testing data, the new criterion shows stronger generalization than other criteria, which proves its reliability and effectiveness.