Imagine that materials is subjected to some extreme condition, such as an explosion or impact, that causes it to fracture and fail in a few microseconds. In order to improve the ability of the material to withstand these conditions we need to understand how it fails, which is much easier if we can observe the failure process. But how can we do that? High-speed video is one option, but it only allows us to see the surface and has limited spatial resolution. X-ray techniques such as computed tomography allow us to peer inside and see the 3D structure, but tomography is too slow for many dynamic events. One alternative is x-ray phase-contrast imaging (XPCI) which can be fast, with frame times of 200 ns or better, but which provides only a 2D projected view of the structure as it evolves.

In this project we seek to establish quantifiable links between the evolving 3D microstructure of a material and our measured 2D XPCI images. To do so, we first build physics-based models that allow us to calculate the XPCI image that would be formed from a known (but hypothetical) microstructure. Such models give us several ways to approach characterization of unknown microstructures. First, we can use them to quantify certain aspect of the microstructure (such as porosity) direction from measured images. Second, we can fit the model to the data to determine various parameters (such as the mean defect size) using an optimization scheme (such as Bayesian optimization). Finally, we can use our model to generate synthetic data sets that can used to train machine learning models to extract microstructural and geometric parameters from measured XPCI images.