Rotational Data Augmentation for DeepSDF-based 3D Generative AI for Industrial Structures
Please login to view abstract download link
Generative AI has revolutionized 3D modeling, yet its industrial application remains limited by two factors: the neglect of mechanical performance in visual-centric models, and the "small data problem" caused by the high computational cost of generating large-scale structural analysis datasets. This study proposes a mechanics-aware 3D generative model based on DeepSDF [1, 2], designed to learn robustly from a limited dataset of 100 topology-optimized aircraft engine brackets. To address the difficulty of reconstructing thin-walled industrial structures, we first optimized the sampling strategy by implementing a mixed-variance approach, combining surface-concentrated sampling with domain-wide sampling. Furthermore, to overcome the data scarcity, we proposed a data augmentation method utilizing the rotational symmetry of the brackets. By rotating the 100 baseline cases around the Z-axis, we expanded the dataset to 3,600 cases. Quantitative evaluation using Chamfer Distance demonstrated that the proposed method significantly improved generalization performance compared to the baseline model without augmentation. The model successfully generated smooth, defect-free shapes that satisfy specified mechanical constraints, proving that high-quality, mechanics-aware generative AI can be constructed even from small datasets.
