Predicting glass transition temperatures using neural networks

The glass transition temperature (T-g) is a kinetic property of major importance for both fundamental and applied glass science. In this study, we designed and trained an artificial neural network to induce a model that can predict the T-g of multicomponent oxide glasses. To do this, we used a dataset containing more than 55,000 inorganic glass compositions and their respective experimental values of T-g. These compositions contain from 3 to 21 of the 45 chemical elements studied here. We implemented an optimization procedure to find artificial neural network hyperparameter values that were able to induce a model with high predictive performance. The resulting neural network model can correctly predict, with 95% accuracy, the published T-g value within less than +/- 9% error, whereas 90% of the data are predicted with a relative deviation lower than +/- 6%. This level of uncertainty is equivalent to the level present in the original dataset and allows a very satisfactory description of the T-g for multicomponent oxide glasses containing combinations of the 45 studied chemical elements. The prediction uncertainty does not depend on the number of elements in the glass composition. However, it is larger for glasses having very high T-g (above 1250 K). The most important aspect is the algorithm's ability to predict the T-g of glasses that are not included in the experimental dataset used for training, thus showing a high generalization ability. Besides, the procedure used here is general and can be easily extended to predict several other properties as a function of the glass composition. This handy feature will most probably help to develop new multicomponent glass compositions having remarkable properties. (C) 2018 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.