Explainable Machine Learning Algorithms For Predicting Glass Transition Temperatures

Abstract

Modern technologies demand the development of new glasses with unusual properties. Most of the previous developments occurred by slow, expensive trial-and-error approaches, which have produced a considerable amount of data over the past 100 years. By finding patterns in such types of data, Machine Learning (ML) algorithms can extract useful knowledge, providing important insights into composition-property maps. A key step in glass composition design is to identify their physical-chemical properties, such as the glass transition temperature, Tg. In this paper, we investigate how different ML algorithms can be used to predict the Tg of glasses based on their chemical composition. For such, we used a dataset of 43,240 oxide glass compositions, each one with its assigned Tg. Besides, to assess the predictive performance obtained by ML algorithms, we investigated the possible gains by tuning the hyperparameters of these algorithms. The results show that the best ML algorithm for predicting Tg is the Random Forest (RF). One of the main challenges in this task is the prediction of extreme Tg values. To do this, we assessed the predictive performance of the investigated ML algorithms in three Tg intervals. For extreme Tg values ( ≤ 450 K and ≥ 1150 K), the top-performing algorithm was the k-Nearest Neighbours, closely followed by RF. The induced RF model predicted extreme values of Tg with a Relative Deviation (RD) of 3.5% for glasses with high Tg ( ≥ 1150 K), and RD of 7.5% for glasses with very low Tg ( ≤ 450 K). Finally, we propose a new visual approach to explain what our RF model learned, highlighting the importance of each chemical element to obtain glasses with extreme Tg. This study can be easily expanded to predict other composition–property combinations and can advantageously replace empirical approaches for developing novel glasses with relevant properties and applications.