Classifying Synthetic 3D Particle Shapes Using Classical Descriptors and Point Cloud Networks

Abstract This paper evaluates geometric rules, Random Forest, PointNet and PointNet⁠+⁠+ in the classification of synthetic particles as either spherical, angular, elongated or platy. We created 8,000 labeled particles and a further 800 elongated/platy boundary set. Random Forest achieved 100% accuracy on the original test split while both deep 2,048-point network architectures reached 99.92%. On the boundary set, the validation tuned rule achieved 100% accuracy and Random Forest achieved 80.12%. Of the eight complete deep architectures tested, PointNet⁠+⁠+ with 2,048 points and without any augmented training rotations achieved the highest accuracy of 87.88%. A sensitivity analysis was carried out on factors including number of points, training augmentations, new test rotations, rule thresholds and class recall. The uncertainty estimates utilized bootstrapping within geometry groups. Interpretation is not straightforward: of 1,200 original test particles, 916 have exact geometric matching within training. Random Forest also achieved 100% accuracy on a further geometry-based test set, while the deep learning models were not tested on this split. Streamlit provides the inference of saved models.