Towards Novel Metamaterial Discovery via Latent Space Regulation and Exploration

ICLR 2026 Conference Submission16113 Authors

Keywords: metamaterial design, voxel generation

TL;DR: We propose VOXPLORER, a voxel-based generative framework that regulates latent space and guides diffusion to overcome the quality–novelty trade-off, enabling discovery of high-quality and novel metamaterials.

Abstract: Metamaterials are artificially engineered structures whose unique mechanical and physical properties arise from geometry rather than composition, enabling applications in wave control, energy absorption, and soft robotics. To capture this structural programmability in a unified form, voxel representation provides a natural choice: it can express diverse classes of metamaterials including truss, shell, and porous metamaterials within a single cubic discretization. However, existing voxel-based generative models face severe limitations. The vast design space, combined with sparse and costly datasets, leads to a generalization dilemma: models tend either to memorize known designs, sacrificing novelty, or to produce invalid, low-quality structures. To address this, we propose VOXPLORER, a generative framework that couples voxel representation with latent space regulation and guided exploration. VOXPLORER introduces a repel-and-sink (RAS) mechanism to smooth and densify the latent distribution of valid structures, and a short-range repulsion (SRR) guidance during diffusion to promote exploration beyond memorized regions while preserving validity. We further contribute a systematic benchmark for voxel-based metamaterials and develop an evaluation module that jointly assess quality, novelty, and diversity. Extensive experiments show that VOXPLORER outperforms state-of-the-art baselines, achieving +8.9% in quality, +46.4% in novelty, and +128.6% in diversity on average across two datasets, establishing a principled pathway toward systematic discovery of next-generation metamaterials.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 16113