Physical-informed deep learning prediction of solid and fluid mechanical properties of oxide glasses

Abstract

The deep learning technique is an efficient method to determine properties of unknown glass compositions. It is used to predict physical properties as density, Young’s modulus, Poisson’s ratio and three isokom temperatures of specific values of the dynamic viscosity. After a recall of models to determine the elasticity properties, the deep learning method is presented with the databases used to build data-sets. To predict density, the fitting is achieved on the molar volume with a large data-set. For the Young’s modulus and according to the Makishima–Mackenzie’s model, the fitting is done on the atomic packing fraction. The Poisson’s ratio is determined according to the Makishima–Mackenzie’s theory involving also the atomic packing fraction. For each prediction, a comparison with experimental data is provided. Finally, predictions are used to see which glass family is the more relevant to optimize the specific Young’s modulus as a function of the “melting” temperature.