IADR Abstract Archives
Predicting Biaxial Flexural Strength of Ceramics by Materials Informatics Approach
Objectives: High flexural strength of glass ceramics is required for further extending clinical applicability, and ISO 6872:2015 recommends testing of biaxial flexural strength. However, for fragile materials, the conventional in vitro approach with repetitive testing is time-consuming due to the difficulty of specimen preparation. Application of computational methodology involving machine learning to the process and interpret dataset concerning materials, so called Materials informatics (MI), is known as a useful approach in engineering field. The aim of this study was to predict the biaxial flexural strength of ceramics by MI approach.
Methods: The scanning electron microscopic (SEM) image and in vitro biaxial flexural strength of 28 commercially available/experimental glass ceramics were collected. The total of 344 SEM images were prepared as the input data. The topological features underlying SEM images were extracted by the persistent homology analysis and compressed by principal component analysis. Support vector regression was employed to develop a machine learning model to predict biaxial flexural strength from the topological features. Bayesian optimization was conducted with 50 repetitions to explore appropriate principal components.
Results: The topological features were compressed into 12 principal components at the 42nd reputations by Bayesian optimization. The machine learning model developed using 12 principal components showed the greatest regression accuracy (training score: 97.7%, validation score: 86.8%, and the coefficient of determination: 0.91). With the machine learning model, the biaxial flexural strengths were predicted with the test score of 91.3%.
Conclusions: The MI approach established in this study successfully predicted the biaxial flexural strength of glass ceramics from the topological features underlying SEM images with acceptable test score. The MI approach promises to make dental materials research more time-efficient than the conventional in vitro approach.
Methods: The scanning electron microscopic (SEM) image and in vitro biaxial flexural strength of 28 commercially available/experimental glass ceramics were collected. The total of 344 SEM images were prepared as the input data. The topological features underlying SEM images were extracted by the persistent homology analysis and compressed by principal component analysis. Support vector regression was employed to develop a machine learning model to predict biaxial flexural strength from the topological features. Bayesian optimization was conducted with 50 repetitions to explore appropriate principal components.
Results: The topological features were compressed into 12 principal components at the 42nd reputations by Bayesian optimization. The machine learning model developed using 12 principal components showed the greatest regression accuracy (training score: 97.7%, validation score: 86.8%, and the coefficient of determination: 0.91). With the machine learning model, the biaxial flexural strengths were predicted with the test score of 91.3%.
Conclusions: The MI approach established in this study successfully predicted the biaxial flexural strength of glass ceramics from the topological features underlying SEM images with acceptable test score. The MI approach promises to make dental materials research more time-efficient than the conventional in vitro approach.