Accurate fractographic strength estimates in single-crystal silicon via neural network approach

Abstract

Unstable cracks in single-crystal silicon generate smooth fracture surfaces at low crack velocity while forming rough surface features when deflecting onto more energy-favorable cleavage planes at higher propagating speeds. In this study, the surface features induced by cracks propagating on the (110) plane due to a flexural stress field were studied. An iteration approach was used to predict the locations where the crack tip branched onto the {111} plane. The numerical scheme successfully predicted the deflection boundaries for specimens within a broad range of strengths. Characteristic fractographic dimensions discernable on the fracture surfaces produced by 3-point bending were first defined and then utilized to train a neural network model. The proposed model significantly improved the strength estimations in single-crystal silicon compared to traditional fractographic methods. Notably, unlike traditional fractography, the neural networks model displayed comparable strength prediction accuracy when analyzing asymmetric fractographic fracture surfaces formed in the presence of secondary surface damage or lateral cracks.