Quantitative fractography for brittle fracture via multilayer perceptron neural network

Abstract

Quantitative fractography is instrumental in the failure analysis of brittle materials, yet the methodology is currently only applied to a handful of materials and loading scenarios as the generalization of the methodology is severely hampered by the unknown roles of many physical factors. With the development of computational tools, artificial neural networks (ANNs) are widely used in engineering and could provide a reliable connection between inputs and outputs. In this work, ANN models were used to analyze the fracture strength of various glasses and ceramics fractured in both flexure and tension. A large set of data consisting of over 4,500 experimental fracture surfaces obtained from 41 types of glasses and ceramics were collected from 82 references for training the models, and the best ANN models for analyzing each scenario were selected. The trained models were further validated by experimental fracture tests conducted in this study. This study showed that the developed ANN could often outperform accepted empirical relations in the literature when predicting the fracture strength and suggested that the developed ANN could be reliably extended to estimate the fracture strength of a broader set of brittle materials. However, it was concluded that for more accurate predictions in the case of ‘non-glass’ ceramics, additional relevant physical factors should be considered in the analysis.