Fracture toughness evaluation of silicon nitride from microstructures via convolutional neural network

Abstract

The fracture toughness of silicon nitride (Si3N4) ceramics was evaluated directly from their microstructures via deep learning using convolutional neural network models. Totally 156 data sets containing microstructural images and relative densities were prepared from 45 types of Si3N4 samples as input feature quantities (IFQs) and were correlated to the fracture toughness as an objective variable. The data sets were divided into two groups. One was used for training, resulting in the creation of regression models for two kinds of IFQs: the microstructures only and a combination of the microstructures and the relative densities. The other group was used for testing the validity of the created models. As a result, the determination coefficient was approximately 0.8 even when using only the microstructures as the IFQs and was further improved when adding the relative densities. It was revealed that the fracture toughness of Si3N4 ceramics was well evaluated from their microstructures.

Artificial Intelligence‐based determination of fracture toughness and bending strength of silicon nitride ceramics

“…It is also well known that accuracy of AI‐determination using CNN models is largely affected by the size of data sets used for training the models. In fact, the R 2 value for K IC from the testing results was approximately 0.85 in this study using the 330 data sets, which is somewhat higher than that of the previous study, 0.80, with the 156 data sets, 22 although the precise comparison is difficult due to the different types of sintering additives. The accuracy, however, is also strongly dependent on the quality of the data sets, that is, their range and distribution.…”

“…On the contrary, the narrow range of the data sets leads to poor prediction of the properties, even with large data volume. The relatively high accuracies of the AI‐determinations of the properties in this study as well as in the previous one 22 are largely because the data sets are widely distributed.…”

“…So far, predictions of mechanical properties using neural network models have been reported in metals and alloys, 10–17 ceramics, 18 and metal–ceramic composites 19–21 . The authors also successfully predicted fracture toughness directly from microstructure images of Si 3 N 4 ceramics via convolutional neural network (CNN) models 22 . This result suggests that a CNN model can extract microstructural information containing complicated interactions that influence the target mechanical properties.…”

“…However, Grad-CAM gave us an important indication of the AI recognition of microstructure; the parts the CNN models focused on for determining both the K IC and σ in this study are essentially the same as those closely observed for the K IC only in the previous work. 22 This suggests that the CNN models determine different properties by focusing on the same parts of the image (in other words, by keeping the prediction modalities), while, for example, a human being is likely to change the focal point depending on the target property.…”

“…[19][20][21] The authors also successfully predicted fracture toughness directly from microstructure images of Si 3 N 4 ceramics via convolutional neural network (CNN) models. 22 This result suggests that a CNN model can extract microstructural information containing complicated interactions that influence the target mechanical properties. However, it is desirable if multiple properties can be generated in a single CNN model analysis because the target properties to be obtained are often multiple.…”