Huang, YuhangKanai, TakashiZordan, Victor2024-08-202024-08-202024978-3-03868-263-9https://doi.org/10.2312/sca.20241163https://diglib.eg.org/handle/10.2312/sca20241163We propose a novel learning-based approach for predicting fractured shapes based on collision dynamics at run-time and seamlessly integrating realistic brittle fracture animations with rigid-body simulations. Our method utilizes BEM brittle fracture simulations to create training data. We introduce generative geometric segmentation, distinct from instance and semantic segmentation, to represent 3D fracture shapes. We adopt the concept of a neural discrete representation learning framework to optimize multiple discrete fractured patterns with a continuous latent code. Additionally, we propose a novel SDF-based cagecutting method to create fragments by cutting the original shape with the predicted fracture pattern. Our experimental results demonstrate that our approach can generate significantly more detailed brittle fractures compared to existing techniques, while reducing computational time typically when compared to traditional simulation methods at comparable resolutions.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Animation; Neural networks; Learning latent representationsComputing methodologies → AnimationNeural networksLearning latent representations10.2312/sca.202411632 pages