Microstructural basis of AI predictions for material properties: A case study of silicon nitride ceramics using t-SNE

Abstract

Artificial intelligence (AI) models such as a convolutional neural network (CNN) are powerful tools for predicting the properties of materials from their microstructural images, etc. It is, however, critically essential to understand how the AI models use images and information to predict the target properties. In this study, we tried to gain insight into the inner workings of two AI models trained to predict bending strength (BS) and thermal conductivity (TC) of silicon nitride ceramics. Focusing on the intermediate feature representation of the microstructural images in the networks, the high-dimensional data points corresponding to sample images were mapped onto a two-dimensional plane using t-distributed stochastic neighbor embedding (t-SNE). The maps demonstrated that the AI models predicted BS and TC primarily based on the porosity and grain sizes of the samples. The result indicates that t-SNE is a useful technique for making the basis of models' predictions more understandable and well founded.