The onset temperature (Tg) of AsxSe1−x glasses transition prediction: A comparison of topological and regression analysis methods

Abstract

AsxSe1−x glasses are promising candidates as matrix for mid-infrared applications, but it is usually invasive, costly and time-consuming or even impossible to measure the onset temperature (Tg) of glass transition of each composition in the system for glass preparation and fiber processing by experimental methods. In this paper, topological and regression analysis (ridge regression, support vector regression and back-propagation neural network) methods are used to predict the Tg of AsxSe1−x glass system and compared with each other. The topological method predicts the Tg of AsxSe1−x glass system by composition dependence of quantitative structure, although its calculation range is limited in the composition range of 0≤x≤0.5 due to no enough knowledge of quantitative structures and their variation. In contrast, regression analysis methods can model the relationships between physical attributes and Tg without complex domain knowledge, thus extending the calculation range to x=0.6 and achieving much higher prediction accuracy. Among them, back-propagation neural network achieves the highest prediction accuracy with an RMSE of 1.21K (7.87K) and MAPE of 0.33% (1.96%) for training data (testing data). Significantly, a three-attribute correlation equation based on ridge regression is obtained, possessing much higher prediction accuracy than that of the topological method.