Machine learning predictions of Knoop hardness in lithium disilicate glass-ceramics

Predicting the effects of ceramic microstructures on macroscopic properties, such as the Knoop hardness, has long been a difficult task. This is particularly true in glass-ceramics, where multiple unique crystalline phases can overlap with a background glassy phase. The combination of crystalline and glassy phases makes it difficult to quantify the percent crystallinity and to predict properties that are the result of the chemical composition and microstructure. To overcome this difficulty and take the first step to build a system for characterizing glass-ceramics, we predict the Knoop hardness based on scanning electron microscopy images using two computational techniques. The first technique is a computer vision algorithm that allows for physical insights into the system because the features used in a predictive model are extracted from the images. The second technique is machine learning with convolutional neural networks that are trained through transfer learning, allowing for more accurate predictions than the first method but with the downside of being a black box. Discussion of the relative merits of the models is included.

Automated summary

Predicting the effects of ceramic microstructures on macroscopic properties, such as Knoop hardness, has been challenging, especially in glass-ceramics with overlapping crystalline and glassy phases. To address this, two computational techniques, computer vision algorithm and machine learning with convolutional neural networks, were employed to predict Knoop hardness based on scanning electron microscopy images. The computer vision algorithm provides physical insights by extracting features from images, while the machine learning method allows for more accurate predictions but lacks transparency. The relative merits of the models are discussed.