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.

Two mechanical properties, fracture toughness (K IC ) and bending strength (σ), of silicon nitride (Si 3 N 4 ) ceramics were determined from their microstructural images via convolutional neural network (CNN) models. The Si 3 N 4 samples used for database were fabricated using various kinds of sintering additives under different process conditions. In total, 330 data sets were prepared and used for building the CNN models for artificial intelligence-bassed determination of the two mechanical properties and testing the determination accuracy of the trained models. The determination coefficients (R 2 ), which were used as accuracy indices, were approximately 0.85 for K IC and 0.92 for σ. Although both the R 2 values were relatively high, the lower value for K IC suggests that it is influenced more by what is little obtained from the microstructural information, such as grain-boundary characteristics. Furthermore, gradient-weighted class activation mapping, which can visualize which parts of the image the CNN models focus on, showed that the trained models determined the two mechanical properties based on correct recognition of the microstructural difference among the images.