Data-driven crack assessment based on surface measurements

Advanced data analysis is increasingly popular with materials engineering. There are many interesting applications, e.g., to identify links between material properties and structural behavior. Most of these applications also entail challenges like compliance with safety requirements for parts and components. These challenges often are specific to the engineering domain, which sets them apart from many other disciplines where data-science already is established. To successfully approach materials science problems with machine learning, one has to identify and address these specifics. In this paper, we pursue this question for crack assessment. More specific, we study whether the prediction of critical stress states is feasible only based on simulated surface measurements of a three-point bending structure. To this end, we pursue several approaches to gain insights into crack initiation and material behavior. We compare different data sets and machine-learning methods to identify variables, such as specific surface locations, that are relevant for high prediction accuracy. Based on our analyses, we discuss the applicability of machine learning for component assessment. Finally, we summarize general principles for machine learning with materials and structural engineering.