PNNL

Abstract

Boron can exist as either 3- or 4-fold coordination ([3]B or [4]B) in oxide glasses and their partition strongly affect the properties of borate and borosilicate glasses. Prediction of boron coordination in multicomponent glasses is critical in glass science and technology in order to understand the composition-structure-property relationships and design of new glass compositions with desirable properties. In this work, we have collected experimental data of boron coordination for a total of 657 borate and borosilicate glasses from literature along with their sample preparation details (e.g., thermal history) as which affect the coordination state of boron. We have developed models with analytical functional forms based on the well accepted Dell, Xiao and Bray model and applied to the whole dataset and a sub-dataset with only borosilicate glasses. Good prediction of boron coordination with a R2 higher than 0.8 have been achieved. However, we also observed large variance of boron coordination experimental data, which was originated from variations of sample preparations and characterization approaches that led to difficulties in obtaining models with a better prediction performance. Machine learning (ML) approaches have been applied to evaluate the ability of different ML algorithms (e.g., K-nearest neighbor, artificial neural network and Gaussian process regression) to predict boron coordination in multicomponent glasses. ML algorithms have the advantage of a slightly better prediction performance as compared to the models using analytical functional forms; however, interpretation of relationships between composition and boron coordination is less straight forward. This study developed and compared various models, providing insights on future improvements on better prediction of boron coordination in multicomponent glasses.