Predicting failure progressions of structural materials via deep learning based on void topology

Abstract

Despite considerable mechanics modeling-based efforts, accurate predictions of failure progressions of structural materials remain challenging in real-world environments primarily due to complex damage factors and defect evolutions. Here, we report a novel deep learning-based method for predicting failure properties based on defect state evolutions, which enables the full reflection of the damage accumulated in a material until the time of its examination. The method uniquely combines nondestructive X-ray computed tomography (X-CT), persistent homology (PH), and deep learning. It exploits the PH-encoded features from 3D X-CT images as its only input, and outputs failure-related properties. Using two fracture datasets based on low-alloy ferritic steel as a representative structural material, the method was demonstrated to reliably classify or predict the local strain (tensile dataset) and fracture progress (fatigue dataset). The excellent deep learning performances are attributed to both PH analysis and multimodal learning, where key topological features of internal voids, such as their size, density, and distributions, are precisely quantified. The proposed method enables accurate prediction of failure-related properties at the time of material examination based on void topology progressions, and can be extended to various nondestructive failure tests for practical use.