Physics-Informed Data-Driven Models for Predicting Time- and Temperature-Dependent Viscoelastic Material Behaviors of Optical Glasses

Abstract

In glass transition regime, glass exhibits a viscoelastic property. This is a time-dependent property inherited from a nonequilibrium and non-crystalline state of materials. The property, furthermore, varies significantly with temperature. Understanding the complex time- and temperature-dependent viscoelastic material behaviors of glass is highly essential in numerous glass processing such as hot forming of glass. In conventional studies, characterization of the viscoelastic responses is prerequisite, requiring intensive experiments conducted over the entire temperature range of the glass transition. Instead, this work introduces an alternative data-driven approach using machine learning, indicating that the time- and temperature-dependent viscoelasticity in the temperature range from glass transition to softening temperatures can be predicted. Different supervised machine learning algorithms were implemented to understand the complexity of the intrinsically non-linear relationships of the existing viscoelastic data, while the most robust and reliable machine learning model was eventually determined by comparing the prediction performance of the selected learning algorithms. Furthermore, the models were developed by employing physics-informed descriptors using the Mauro-Yue-Ellison-Gupta-Allan (MYEGA) viscosity model. We demonstrate that the physics-informed descriptors are beneficial to enhance the extrapolation capability of the predictive model for two extreme temperature ranges of glass viscosity e.g., at the softening temperature and in the vicinity of glass transition temperature, while the dimensionality of the model is reduced. The promising result of this study allows glass manufacturers to bypass the costly experimental characterization required for understanding the thermo-viscoelastic material behaviors in glass processing.