Conventional methods for crystal structure prediction in materials design require extensive structural sampling via computationally expensive energy minimization methods using force-field or quantum mechanical simulations. While emerging AI generative models have shown great promise in generating realistic crystal structures more quickly, most existing models fail to account for the inherent symmetry and periodicity of crystalline materials and are limited to structures containing only a few dozen atoms per unit cell. In this paper, we present LEGO-xtal (Local Environment Geometry-Oriented Crystal Generator), a symmetry-aware AI generative approach that overcomes these limitations. This method generates initial structures using AI models trained on a small, augmented dataset, and then optimizes them using a machine learning structure descriptor rather than conventional energy-based optimization. We demonstrate the effectiveness of LEGO-xtal by scaling the model to over 1,700 structures from 25 known low-energy sp2 carbon allotropes. All of these structures are within 0.5 eV/atom of the ground-state energy of graphite. This framework provides a generalizable strategy for the targeted design of materials with modular components, such as metal-organic frameworks and next-generation battery materials.
