Application study on conditional generative adversarial network (cGAN) to generate ballast particles for discrete element method simulation

ABSTRACT Understanding ballast particle morphology is important for evaluating the load-bearing capacity of ballasted track foundations using numerical simulations. Although several methods, such as digital imaging, laser scanning, and computed tomography scans, are widely used to capture ballast morphology, real ballast layers consist of tens of thousands to millions of particles of varying shapes and sizes, making these methods complex. Therefore, an efficient solution needs to be found that can generate large-scale ballast datasets for the simulations. This study aims to develop a conditional generative adversarial network (cGAN) to generate ballast particles classified as: angular, subangular, subrounded, or rounded. The cGAN model consists of a generator and a discriminator network, where the generator aims to produce generative data based on the distinction from real data estimated by the discriminator. To find the optimal network architecture for the cGAN, the energy distance metric was investigated by varying the learning rate and number of neurons in hidden layers. The ballast particles generated using the optimal cGAN model were assessed using the receiver operating characteristic area under the curve (ROC AUC). The average ROC AUC was 0.9827, indicating a high classification performance. In addition, roundness coefficients were computed, showing that the morphology of ballast particles generated aligned well with the predefined ballast classes. The cGAN model was therefore shown to be effective at creating realistic ballast particles for further numerical simulations.