Deep Learning for Quantitative Dynamic Fragmentation Analysis

Abstract

We have developed an image-based convolutional neural network (CNN) that is applicable for quantitative time-resolved measurements of the fragmentation behavior of opaque brittle materials using ultra-high speed optical imaging. This model extends previous work on the U-net model, where we trained binary, 3 and 5 class models using supervised learning on experimentally measured dynamic fracture experiments on various opaque structural ceramic materials that were adhered on transparent polymer (polycarbonate or acrylic) backing materials. Full details of the experimental investigations are outside the scope of this manuscript but briefly, several different ceramics were loaded using spatially and time-varying mechanical loads to induce inelastic deformation and fracture processes that were recorded at frequencies as high as 5 MHz using high speed optical imaging. These experiments provided a rich and diverse dataset that includes many of the common fracture modes found in static and dynamic fracture including cone cracking, median cracking, comminution, and combined complex failure modes that involve effectively simultaneous activation and propagation of multiple fragmentation modes. While the training data presented here was obtained from dynamic fragmentation experiments, this study is applicable to static loading of these materials as the crack speeds typically higher a kilometer per second in these materials are on the order of 1-10 km/s regardless of the loading rate. We believe the methodologies presented here will be useful in quantifying the failure processes in structural materials for protection applications and can be used for direct validation of engineering models used in design.