Three-dimensional sand particle shape characterization by PointConv-Transformer deep learning

Abstract

Abstract Angularity and sphericity of sand particles significantly influence the shear strength, compressibility, void structure, and deformation behavior of soils. However, current computational geometry (CG) algorithms face challenges in simultaneously charactering angularity and sphericity, as well as handling defective granular three-dimensional(3D) mesh. To address these efficiency and robustness limitations, this paper introduces the PointConv-Transformer deep learning algorithm for characterization and classification of sand particles point cloud. The PointConv efficiently captures local features of angularity and sphericity. Subsequently, the Transformer integrates these local features into global features to form the judgement basis for classification. The 4800 particle point clouds are labeled with 12 angularity-sphericity classes. The PointConv-Transformer model, trained on the produced dataset, achieves an automatic classification accuracy of 96.65%. Furthermore, we explore the impact of normal vectors and point cloud size on the performance of the PointConv-Transformer model. Experimental results demonstrate that the optimal performance of the trained model is achieved when the point cloud size is 2000 and includes normal vectors. Finally, compared to traditional 3D CG, the classification results align closely in volume, surface area, and convex hull volume metrics. As the number of classified particles increases, the advantage in classification efficiency becomes more pronounced.