ViscNet: Neural network for predicting the fragility index and the temperature-dependency of viscosity

TL;DR

This paper presents ViscNet, a physics-informed neural network that predicts the fragility index, glass transition temperature, and viscosity of oxide liquids across temperatures, enabling accurate extrapolation and practical applications in glass-making.

Contribution

The study introduces a novel neural network model that predicts key parameters of the MYEGA viscosity equation, improving viscosity and property predictions with strong extrapolation capabilities.

Findings

Achieved $R^2$ of 0.97 on test data for viscosity prediction.

Developed a model with good extrapolation for temperature-dependent viscosity.

Provided an open-source tool for predicting viscosity and related properties.

Abstract

Viscosity ($\eta$) is one of the most important properties of disordered matter. The temperature-dependence of viscosity is used to adjust process variables for glass-making, from melting to annealing. The aim of this work was to develop a physics-informed machine learning model capable of predicting $\eta(T)$ of oxide liquids. Instead of predicting the viscosity itself, the NN predicts the parameters of the MYEGA viscosity equation: the liquid's fragility index, the glass transition temperature, and the asymptotic viscosity. With these parameters, $\eta$ can be computed at any temperature of interest, with the advantage of good extrapolation capabilities inherent to the MYEGA equation. The dataset was collected from the SciGlass database; only oxide liquids with enough data points in the high and low viscosity regions were selected, resulting in a final dataset with 17,584 data points containing 847 different liquids. About 600 features were engineered from the liquids' chemical composition and 35 of these features were selected using a feature selection protocol. The hyperparameter (HP) tuning of the NN was performed in a set of experiments using both random search and Bayesian strategies, where a total of 700 HP sets were tested. The most successful HP sets were further tested using 10-fold cross-validation, and the one with the lowest average validation loss was selected as the best set. The final trained NN was tested with a test dataset of 85 liquids with different compositions than those used for training and validating the NN. The $R^2$ for the test dataset's prediction was 0.97. This work introduces three advantages: the model can predict viscosity as well as the liquids' glass transition temperature and fragility index; the model is designed and trained with a focus on extrapolation; finally, the model is available as free and open-source software licensed under the GPL3.