Machine learning prediction of the liquidus temperature for oxide glass-formers

Reference	Presenter	Authors (Institution)	Abstract
10-041	Graziela Pentean Bessa	Bessa, G.P.(Federal University of São Carlos); Cassar, D.R.(Federal University of São Carlos); Zanotto, E.D.(Federal University of São Carlos (UFSCar));	The liquidus temperature (TL) is the highest temperature where one or more crystal phases can co-exist with the liquid phase in thermodynamic equilibrium. TL is a very important glass-making parameter because it determines the minimum temperature of the melting process. Undesired crystallization, due to insufficient melting, can cause visual defects and even lead to product failure. Also, the higher the melting temperature, the higher the production cost, hence there has been an intensive quest to optimize and speed up the measurement of TL for complex, multicomponent glass-forming compositions. It is possible to obtain the TL from equilibrium phase diagrams, but this method is limited to simpler compositions containing up to three oxides. The TL for complex compositions, containing several different chemical elements, is available for selected compositions. The objective of this work is to train an artificial neural network that is capable of predicting the TL for multicomponent oxide compositions. We collected a dataset of 36019 compositions (containing 40 chemical elements, each one present in at least 1% of all data) and their respective experimental values of TL. First we analyzed the quality of the dataset by comparing the reported values of TL with the values of the equilibrium diagrams of B2O3-Na2O and SiO2-Na2O and observed that 77% and 86%, respectively, of the reported data lie within the expected value with a maximum error of ± 5%. After filtering for the outliers, the predictive power of our best network was much improved and will be fully discussed at the conference.
<< Back