Regenerated realistic sand shapes

Reliability analysis and optimized design in geotechnical and structural engineering stimulates the requirement for realistic virtual sand grain datasets with rich irregular morphologies. However, the significant limitations of the present regeneration methods remain, such as ideal distribution assumptions, unverifiable alterations of particle representations, lack of diversity, and high demand for databases. To bypass these issues, a hybrid generation method, combining the spherical harmonic (SH) analysis and vector quantisation techniques was presented. Three-dimensional particle shapes with rich morphology were first compressed as SH coefficients. The SH coefficients were then encoded as tokens in low- and high-frequency domains using vector-quantised autoencoders. Integrated generative diffusion models were subsequently trained to produce new token samples. Incorporating with the decoding process of autoencoders, dataset enrichment of realistic sand grain shape was finally achieved. The reliability and reproducibility of the present method were confirmed by validating the SH coefficients and shape parameter distribution of three kinds of synthesized sand grains. On this basis, regenerated particles were utilized for granular column collapse simulation, showcasing the applicability of particle-related tasks. The present hybrid method is promising to underpin the comprehensive and reproductive study of mechanical and dynamic properties in sand grain materials.