3D Crystal Size Distribution Analysis using Machine Learning for Image Segmentation: Application to the Bishop Tuff

Crystal size distributions (CSDs) preserved in volcanic rocks elucidate important pre and syn-eruptive parameters including magma residence time, viscosity (and thus eruptibility), and ascent rate. Despite their importance, quantifying 3D CSDs using X-Ray Tomography remains limited by resolution trade-offs, small N, and time intensive crystal classification following image acquisition. To address these limitations, we take ~30 sub-samples of one Bishop Tuff pumice clast and image ~10 sub-samples per resolution (1.24, 3.18, 5.72 µm/voxel) building on Pamukçu & Gualda (2010). Images were acquired at Argonne National Lab’s Advanced Photon Source (GSECARS), which is a synchrotron CT facility. This method captures ~ 10,000 crystals per resolution for a wide range of crystal sizes (Spherical Equivalent Diameter ~ 0.001-1 µm). Following image acquisition, we train a single 2D U-Net model per resolution using Dragonfly 2025.1 segmentation software. Our models successfully identify 6 phase groups: pore space, finely vesiculated glass, quartz/feldspar, pyroxene/biotite, and accessory minerals. From these classified image stacks, we extract ~90 crystal size distributions (1 per sub-sample). We find that distributions vary by sub-sample within a given resolution and phase group. This is most obvious for the feldspar/quartz group, wherein CSDs for some sub-samples fit a power law distribution, indicating fragmentation, and others fit multiple exponential distributions, indicating several episodes of continuous nucleation and growth. Fragmentation seems to be at least partly associated with melt inclusion decrepitation. These results indicate that intra-sample textural variability can be significant. As such, future work should utilize multiple sub-samples in tandem with machine learning for image segmentation, which can speed up lengthy post-processing from weeks to days.