Reproduce “Structural Constraint Integration in Generative Model for Discovery of Quantum Material Candidates” on OpenResearch

About this paper

Ryotaro Okabe, Mouyang Cheng, Abhijatmedhi Chotrattanapituk, Nguyen Tuan Hung, Xiang Fu, Bowen Han, Yao Wang, Weiwei Xie and 4 more

A framework called SCIGEN integrates specific geometric and symmetry constraints into pre-trained generative diffusion models for materials discovery without retraining. This enabled the generation of millions of novel material candidates, including stable Lieb-like lattice structures with confirmed flat electronic bands via DFT calculations, demonstrating targeted exploration of chemical space.

Key insights

• Explicitly integrating geometric and symmetry constraints into generative models allows for hypothesis-driven material design, targeting structures known to host specific quantum properties.
• A strategic masking and integration procedure within the denoising steps of a pre-trained diffusion model can achieve conditional sampling, guiding material generation towards specific constraints without requiring model retraining.
• Combining generative models with multi-staged machine learning pre-screening and high-throughput Density Functional Theory (DFT) validation provides an efficient pipeline for discovering novel, stable materials with desired structural motifs.

Summary by alphaXiv.