High-Precision identification and classification of alloy fatigue microcracks through deep learning and in-situ SEM

Abstract

The formation and propagation of microcracks are critical stages leading to fatigue failure. Traditional methods for microcrack analysis essentially rely on manual detection, which poses challenges in accuracy and efficiency. The present investigation deals with a novel and highly effective strategy for detecting, classifying, and analyzing fatigue microcracks in Ti-6Al-4 V (TC4) titanium alloy with artificial intelligence (AI) in the context of deep learning. By integrating in-situ scanning electron microscope (SEM) images and convolutional neural network (CNN) algorithm, we propose a PGI-CrackNet model that is able to detect microcracks of length around 15 μm and thereby effectively outperform the detection capabilities of traditional models. Based on fracture mechanics models, the proposed model is capable of automatically identifying the main stages (i.e., initial crack, type I crack, type II crack, and break) in the formation of microcracks. Simultaneously, the proposed model bridges the gap between the AI-based image analysis and the physical crack propagation models, enabling the extraction of key information such as microcrack length and width, and further supporting the analysis of fatigue crack growth rates associated with various microcrack stages. The model could discover that in the initiation stage, the crack of TC4 titanium alloy grows at a fairly slow rate (∼3.6 μm/cycle) and occupies most of the crack life cycle. After the initiation stage, the crack first propagates as the type I cracks with a significantly faster crack growth rate (∼50 μm/cycle). Then, the type II crack occurs with a substantially reduced growth rate (∼25 μm/cycle). In the final stage, as the microcrack reaches a critical size, the growth rate increases sharply, leading to break. In summary, this improved PGI-CrackNet-based model enables more accurate tracking of crack growth over the fatigue life of materials and better classification of crack types based on their propagation mechanisms, making it highly suitable for early warning applications of material failure.