Dataset for Accelerating In Situ X-ray Tomography Using Sparse Projections and Deep Learning

Abstract (from manuscript): In situ X-ray tomography experiments provide powerful insight into material deformation and damage evolution under mechanical load. However, laboratory-based tomography setups often suffer from long acquisition times that hamper real-time observation. In this work, we demonstrate a deep-learning-based reconstruction method capable of high-quality volume reconstructions from sparse (10X fewer) projection datasets. We apply this method to in situ tensile experiments on additively manufactured Inconel 718 dog-bone specimens, significantly increasing our sampling density from only a few tomography scans per experiment to over 20 scans at different load steps. We compare conventional FDK reconstructions using 1001 projections with both (a) FDK using only 101 projections, and (b) our trained deep-reconstruction (DR) model with 101 projections. We show that the deep-reconstruction images closely match those from the full-projection FDK, enabling effective 3D segmentation and quantification of evolving porosity in the test samples. The data here demonstration our approach to increasing data throughput and sampling density for in situ studies.