3D particle shape generation for Hybrid particulate process modeling: GANs and Diffusion Transformers
Please login to view abstract download link
Traditional 3D reconstruction methods for powders and particles, such as X-ray Computed Tomography (XRCT), are time-consuming and repetitive. Dynamic image analysis techniques like QicPic and Camsizer are faster but limited by the small number of 2D images generated. To overcome these challenges, we present two independent AI-driven approaches leveraging Generative Adversarial Networks (GANs) and Diffusion Transformers (DiT) to reconstruct high-fidelity 3D particle shapes using only 2D images. The GAN-based approach generates 3D models from Gaussian noise, projecting them into three orthogonal 2D views for evaluation by a discriminator alongside experimental images. A conditional GAN variant extends this to reconstruct 3D models from a single 2D image, replicating complex particle shapes without needing 3D ground truth data. Separately, DiT iteratively refines noisy embeddings into precise 3D representations using attention-based denoising, capturing intricate global relationships and long-range dependencies. These methods operate independently, providing flexible, efficient alternatives to traditional techniques. Furthermore, the generated 3D particle models can enhance Discrete Element Method (DEM) simulations, improving accuracy in representing particle shapes for granular material analyses. By addressing the limitations of conventional methods, this framework enables scalable and realistic 3D particle modeling for scientific and industrial applications.
