Prediction of Glass-Ceramics Color Parameters by Machine Learning Models

Objectives: Color and translucency are important properties of dental ceramic materials. These properties which must match the shade of natural teeth, are designed by experimental trial-and-error approach. In comparison, materials informatics (MI) can be a favorable alternative scientific approach. The objective of this study was to utilize the MI approach to predict color parameters of dental glass-ceramics.
Methods: A total of 364 samples of experimental lithium disilicate (LDS) glass-ceramics were collected, and dataset was constructed. The dataset contained 32 variables including composition, process conditions, L*a*b* color coordinates on white and black backgrounds, Total Transmittance(T.T.), and Transparency Parameter(TP). Two machine learning (ML) models, Gaussian Mixture Regression (GMR) and Student’s-t Mixture Regression (SMR), which can handle multi-target variables and direct inverse analysis were employed to perform regression analysis. Hyperparameters of these ML models were optimized by Bayesian Optimization. The performance of regression analysis of the two models was evaluated by Root-Mean-Squared-Error(RMSE) which shows the difference between observed values and predicted values. Thus, a lower value indicates better model performance. Additionally, Principal Component Analysis was carried out for data visualization of the color parameter space of LDS glass-ceramics.
Results: Model performance at each target variable on the test dataset is shown in the Table. SMR showed lower RMSE scores than that of GMR, except for b*(White) and TP.
Conclusions: ML models were utilized to predict color properties of dental LDS glass-ceramics as MI approach. MI provides a cost-effective approach for material design and better knowledge of the material space. In this study, results suggest that SMR and GMR can recognize the pattern between the design of dental LDS glass-ceramics and color parameters, and support prototyping these materials.