Figure 2: Our x-ray transparent deformation apparatus, Mjölnir, in (a ) schematic (Butler et al., 2017), and (b ) installed at the PSICHE beamline. (c ) Stepped loading procedure. (d ) Reconstructed μCT image showing damage accumulation at one end of the sample (this occurred in both samples) – dashed white line shows the analyzed sub-volume.

X-ray imaging and image data pre-processing

X-ray μCT data was acquired using an sCMOS (scientific Complementary-Metal-Oxide Semiconductor) Hamamatsu ORCA Flash4.0 camera, with a Rodagon 50 lens, giving about 2.5x magnification (effective pixel size 2.7 µm), and a 250 µm thick LuAG:Ce scintillator. The white beam with an average detected energy of about 66 keV was filtered with 1 mm aluminium and 0.5 mm tungsten. During each scan, 1200 projections were acquired over 180°, with an exposure time per projection of 15-19 ms depending on the progressive darkening of the objective lens. A mix of absorption and phase contrast data was acquired, with a sample to detector distance of 125 mm.
Each µCT volume was reconstructed by filtered back projection. Reconstructions were performed at the PSICHE beamline, using both x-ray absorption and phase contrast modes as implemented in the PyHST2 software (Mirone et al., 2014), and yielded 3D volumes of 1700 x 1700 x 4102 equidimensional voxels, with a voxel edge length of 2.7 μm. These volumes were then processed to extract the fracture network from the reconstructed images. To deal efficiently with the huge size of each 3D volume (approx. 40 GB) and speed up the subsequent processing, we selected a sub-volume of interest – the region in the failed samples where the majority of damage had accumulated (Figure 2d, Table 1). Using the AvizoTM software package, this sub-volume was extracted from each of the full 3D volumes and de-noised with an anisotropic diffusion filter (stop value 0.4 over 4 iterations), which was chosen to emphasize the microcrack features as it preserves strong edges and enhances edge contrast. It was then down-sampled to 16-bit with a 32-bit threshold range of -0.3 to 0.8, yielding individual datasets of manageable size (approx. 3 GB).
Table 1: Dimensions of the whole sample and analyzed sub-volume, with uncertainties to two decimal places.