The science of fractography revolves around the correlation between topographic characteristics of the fracture surface and the mechanisms and external conditions leading to their creation. While being a topic of investigation for centuries, it has remained mostly qualitative to date. A quantitative analysis of fracture surfaces is of prime interest for both the scientific community and the industrial sector, bearing the potential for improved understanding on the mechanisms controlling the fracture process and at the same time assessing the reliability of computational models currently being used for material design. With new advances in the field of image analysis, and specifically with machine learning tools becoming more accessible and reliable, it is now feasible to automate the process of extracting meaningful information from fracture surface images. Here, we propose a method of identifying and quantifying the relative appearance of intergranular and transgranular fracture events from scanning electron microscope images. The newly proposed method is based on a convolutional neural network algorithm for semantic segmentation. The proposed method is extensively tested and evaluated against two ceramic material systems (Al 2 O 3 ,M gAl 2 O 4 ) and shows high prediction accuracy, despite being trained on only one material system (M gAl 2 O 4 ). While here attention is focused on brittle fracture characteristics, the method can be easily extended to account for other fracture morphologies, such as dimples, fatigue striations, etc.The fracture process of materials is governed by both extrinsic (e.g. imposed loading, environmental conditions) and intrinsic (microstructure) characteristics. One may thus expect, that the fracture surface will contain evidence regarding the influence of both the intrinsic and extrinsic characteristics of the fracture process. Fractography is a powerful tool employed to study fracture

Modern computer vision and machine learning techniques, when applied in Fractography bare the potential to automate much of the failure analysis process and remove human induced ambiguity or bias. Given the complex interaction between intrinsic (e.g. microstructure) and extrinsic (e.g. environment, loading history) factors leading to failure, deep learning methods, which exhibit very high efficiency in establishing complex interconnections between the input data, may end up revealing new correlations and information that is encoded onto the complex geometries of fracture surfaces and remained hidden from us so far. In this work, we examine the potential use of an unsupervised learning pipeline to classify fracture surfaces of five tungsten heavy alloys following their chemical content (i.e. Tungsten percentage). Encouraged by the success of the algorithms, we move on and analyze the features on the fracture surfaces which are governing the decision process of the algorithms. The fractographic interpretation of these features shows that the extent of plasticity on the fracture surface serves as a measure for the classification process. The examined pipeline can be used to identify failures originating from erroneous manufacturing processes, leading to locally varying Tungsten concentrations and ultimately premature failure.

No abstract

We introduce a novel machine learning computational framework that aims to compute the material toughness, after subjected to a short training process on a limited meso-scale experimental dataset. The three part computational framework relies on the ability of a graph neural network to perform high accuracy predictions of the micro-scale material toughness, utilizing a limited size dataset that can be obtained from meso-scale fracture experiments. We analyze the functionality of the different components of the framework, but the focus is on the capabilities of the neural network. The minimum size of the dataset required for the network training is investigated. The results demonstrate the high efficiency of the algorithm in predicting the crack growth resistance in micro-scale level, using a crack path trajectory limited to 200-300 grains for the network training. The merit of the proposed framework arises from the capacity to enhance its performance in different material systems with a limited additional training on data obtained from experiments that do not require complex or cumbersome measurements. The main objective is the development of an efficient computational tool that enables the study of a wide range of material microstructure properties and the investigation of their influence on the material toughness.

No abstract