AI-Based Particle Size Distribution Prediction from 3D Packed-Particle Tomographic Volumes: A Comparative Study with classical Watershed Segmentation

Abstract Determining the Particle Size Distribution (PSD) of granular media and powder compacts is vital across condensed matter physics, geological sciences, and materials engineering. While X-ray computed tomography (µCT) captures granular microstructure non-destructively, traditional quantitative PSD estimation relies on classical morphological segmentation (binarization, distance transform, and watershed flooding), which is computationally intensive and sensitive to contact-neck ambiguities. In this work, we investigate whether a compact 3D Convolutional Neural Network (Compact3DCNN, 300,900 parameters) can predict normalized 20-bin PSD histograms directly from unsegmented 3D grayscale volumes without intermediate particle instance extraction. Using a synthetic dataset of 60 independent non-overlapping polydisperse packed-sphere volumes (483 voxels) with exact analytic PSD ground truth, we benchmark the CNN against a classical 3D Euclidean distance-transform watershed baseline on an untouched, held-out test split (N = 10, 40 train / 10 validation / 10 test). On this low-to-moderate packing regime (15%–25%), the classical watershed baseline achieved superior accuracy, attaining a mean Earth Mover’s Distance (EMD) of 0.0974 ± 0.0261 voxels compared to 0.3428 ± 0.0949 voxels for the CNN (paired 95% bootstrap CI: [+0.1837, +0.3156] voxels). Measured runtime was 1.19 ± 0.27 ms per volume for the CNN on an NVIDIA RTX 3050 Laptop GPU versus 33.71 ± 7.50 ms for the watershed pipeline on an Intel CPU; however, we emphasize that this reflects disparate computing architectures rather than a controlled hardware comparison. Furthermore, comparison against an empirical training-set-mean baseline (0.6206 ± 0.2440 vx EMD on validation) confirms that the CNN learns volume-specific morphological features (0.3970 ± 0.1335 vx, a 36.0% error reduction). A controlled validation experiment evaluating direct 1D Wasserstein-1 loss training yielded substantially higher error (1.1382 ± 0.4907 vx EMD) than the primary Kullback-Leibler (KL) divergence configuration, illustrating optimization challenges in cumulative L1 histogram penalties. Our findings characterize the accuracy-throughput profile in direct volumetric PSD regression and delineate key open challenges for dense granular systems.