Deep Learning for Quantitative Dynamic Fragmentation Analysis

We developed an image-based convolutional neural network (CNN) designed for quantitative, time-resolved measurement of fragmentation behavior in opaque brittle materials using ultra-high-speed optical imaging. Building on prior work with the U-net model, we trained binary, 3-class, and 5-class models via supervised learning, using data from dynamic fracture experiments conducted on various opaque structural ceramics adhered to transparent polymer backings (polycarbonate or acrylic). Different ceramics were subjected to spatially and temporally varied mechanical loads to induce inelastic deformation and fracture processes. These were captured at up to 5 MHz using high-speed optical imaging, producing a rich dataset with a range of fracture modes typical of static and dynamic fractures, including cone cracking, median cracking, comminution, and complex failure modes involving simultaneous activation of multiple fragmentation processes. We present alternative strategies to solve the problem at hand by introducing different segmentation tasks and discuss the overall performance where the highest recall obtained was 95.54%. While the training data was derived from dynamic experiments, the approach is equally applicable to static loading, as crack speeds in these materials remain in the range of 1–10 km/s irrespective of loading rate. We anticipate that the methodologies presented here will aid in quantifying failure processes in structural materials for protective applications and will support direct validation of engineering models used in design.