Random generation of three-dimensional realistic ballast particles using generative adversarial networks

Natural ballast particles exhibit a wide array of shape characteristics and follow complex probability distributions, which are difficult to quantify using explicit formulas due to their implicit nature. Traditional methods for generating ballast particles are limited to shapes that adhere to specific shape indices. In this study, generative adversarial networks (GANs) were utilized to randomly generate realistic three-dimensional (3D) ballast particles. Representative ballast particle shapes were captured using 3D structured light scanning technology. The scanned shapes were then voxelized and augmented to create a comprehensive dataset representing various ballast particle shapes. This dataset served to train the generator and discriminator components of the GANs. The trained generator successfully produced 2,000 detailed 3D ballast particles. Further refinement of these particles was achieved using Gaussian blur, followed by the Laplace smoothing algorithms and marching cubes algorithm for surface reconstruction. The authenticity of the generated ballast particles was validated by comparing their shape indices with those of natural ballast particles, thus demonstrating the effectiveness of the trained GANs model. The generated ballast particles are suitable for use as templates in discrete element method (DEM) simulations, specifically for clumps and polyhedrons.