Previsão da habilidade de formar vidros (gfa) via algoritmos de inteligência artificial

Abstract

Glasses are often formed through the supercooling of liquids. The ability to form glasses, known as glass-forming ability (GFA), is a fundamental property in the development of new glasses; however, obtaining it through experimental methods is costly. Therefore, there is significant scientific interest in the development of indirect methods for GFA acquisition. Previous studies have identified two fundamental parameters for predicting GFA: the liquidus temperature and the viscosity at the liquidus temperature. In this work, viscosity at Tliquidus was determined for 4950 oxide glass compositions by fitting the Vogel–Fulcher–Tammann–Hesse equation (VFTH, log10 η(T)=A+B/(T- T0), where η is in Pa.s and T is in K). This work proposes the integration of physical models and Machine Learning (ML) to predict the GFA of oxide glasses through two approaches. In the indirect pathway, an ML model is generated to predict the Tliquidus, and another model predicts the viscosity at Tliquidus calculated by the VFTH model. Subsequently, these two parameters are used to calculate GFA through the "Jezica"physical model, where GFA = η(Tl) / Tl 2 . The influence of each component in predicting viscosity was evaluated based on the roles of network-forming, intermediate, and modifier oxides, providing a new perspective in structural theories and their role in properties. On the other hand, in the direct pathway, Tliquidus and viscosity at Tliquidus are used to calculate GFA, and the ML model allows for the direct prediction of GFA. These approaches aim to enable the prediction of GFA solely based on the composition of oxide glass formers and to enhance the understanding of this property, thereby accelerating the development of new glasses.