Topological features of lithium disilicate glass-ceramics uncovered through materials informatics.

Abstract

Objective: The aim of this study was to inversely predict the topological features underlying SEM images from arbitrary biaxial flexural strengths of glass-ceramics by Materials Informatics (MI) approach.

Methods: The scanning electron microscopic (SEM) image and in vitro biaxial flexural strength of 10 commercially available/experimental glass-ceramics were collected. The total of 200 SEM images were prepared as input data. Topological features underlying the SEM images were extracted using persistent homology analysis and compressed using principal component analysis. Gaussian mixture regression was employed to develop a machine learning model for predicting biaxial flexural strength based on the topological features. Arbitrary biaxial flexural strengths (390, 411, 442, 478, 515, 564, 597, 610, and 640 MPa) were defined, and an inverse analysis was conducted with the constructed machine learning model to overlay topological features onto SEM images.

Results: The topological features were compressed into 18 principal components. The machine learning model was selected and optimized based on the Bayesian Information Criterion. Using the constructed machine learning model, the biaxial flexural strengths were predicted with a test score of 72 % (Root Mean Squared Error: 53.5, Mean Absolute Error: 40.3). From the arbitrary biaxial flexural strengths, topological features were inversely predicted and overlaid onto SEM images.

Conclusion: The inverse analysis established in this study successfully predicted the topological features on SEM images of glass-ceramics from the biaxial flexural strengths. The MI approach with the inverse analysis promises to make the process to develop glassceramics more time-efficient than the conventional in vitro approach.