MIND: Microstructure INverse Design with Generative Hybrid Neural Representation

• DOI: 10.48550/arxiv.2502.02607
• Publication Date: 2025-02-01
• Authors: Tianyang Xue; Haochen Li; Lu Liu; Paul Henderson; Pengbin Tang; Lin Lü; Jikai Liu; Haisen Zhao; Hao Peng; Bernd Bickel
• Keywords: Representation; Generative Design; Generative model

The inverse design of microstructures plays a pivotal role in optimizing metamaterials with specific, targeted physical properties. While traditional forward methods are constrained by their inability to explore the vast combinatorial space, offers compelling alternative directly generating structures that fulfill predefined performance criteria. However, achieving precise control over both geometry and material properties remains significant challenge due intricate interdependence. Existing approaches, which typically rely on voxel or parametric representations, often limit flexibility structural diversity. In this work, we present novel generative model integrates latent diffusion Holoplane, an advanced hybrid neural representation simultaneously encodes geometric This combination ensures superior alignment between Our approach generalizes across multiple microstructure classes, enabling generation diverse, tileable significantly improved property accuracy enhanced validity, surpassing existing methods. We introduce multi-class dataset encompassing variety morphologies, including truss, shell, tube, plate structures, train validate our model. Experimental results demonstrate model's ability generate meet target properties, maintain integrate seamlessly into complex assemblies. Additionally, potential framework through new microstructures, cross-class interpolation, infilling heterogeneous microstructures. source code will be open-sourced upon publication.