Glass Viscosity Prediction

Neural network for predicting the temperature-dependency of viscosity

Introduction

The viscosity of a glass has a substantial effect on other physical properties, such as melting, softening, and crystallization characteristics, as well as the pressure and temperature ranges within which the glass can be worked. Therefore, In the context of oxide glass-forming liquids, the temperature-dependence of viscosity is used to adjust process variables for glass making

Problem Statement

The viscosity of a glass has a substantial effect on other physical properties, such as melting, softening, and crystallization characteristics, as well as the pressure and temperature ranges within which the glasses can be worked.These predictive models are expected to increase the speed and reduce the cost of developing new materials. In this context, the most used machine learning technique by far is neural networks (NN) , which are particularly good at finding patterns and modeling non-linear dependencies between a set of features (input) and targets (output).We use the MYEGA equation to understand the relation between the temperature and viscosity relation.

Oxide Properties

In the MYEGA equation, η is the viscosity, T is the absolute temperature, the realtion used is given below:

log10 (η (T, η∞, Tg, m)) = log10 (η∞)

Pre - Processing Technique

Having data segregated as X and Y where the data relating to the oxides were stored in X whereas the data regarding the temperature variation was stored in Y.

So example we have an oxide such as Al203 has input 0.35,0.67...etc at an absolute temperature 2180,3000 (in celsius). The Neural Network predicts the relation between the two data and predicts the viscosity at different temperatures which is further validated with a separate portion from the dataset . Thereby we had 2 arrays that would be combined as the input to the model to predict the desired output i.e, viscosity at lower losses.

Interference

The first tabular column shows a comparison between the predicted viscosity(from the model) and actual viscosity (from the data set ) and as observed that the predicted viscosity comes very close to that of the actual values .

The second tabular column shows that the predicted values are a little over a scale of 1 and only 246 elements are greater than 1 in a 17000+ dataset.