Design Topological Materials by Reinforcement Fine-Tuned Generative Model

Topological insulators and topological crystalline insulators are characterized by robust surface states and insulating bulk behavior, rendering them highly valuable for quantum computing, spintronics, and other emerging technologies. However, the discovery of such materials—particularly those with a full band gap—remains challenging, primarily due to the limitations of conventional approaches that rely on the screening of known materials. Here, we employ reinforcement fine-tuning on a pre-trained generative model to facilitate the discovery of topological materials. This approach enables targeted material generation while preserving chemical validity and structural stability. Our fine-tuned model significantly improves the likelihood of generating topological insulators and topological crystalline insulators, leading to the identification of numerous new candidate materials. Among them, Ge2Bi2O6 emerges as a strong topological insulator with a full band gap of 0.26 eV—one of the largest reported to date. These findings highlight the power of reinforcement-based generative design for discovering materials with targeted properties. The authors develop a reinforcement fine-tuned generative model to design topological materials, increasing the yield of topological insulators while preserving validity. It identifies new candidates, including a large-gap topological insulator Ge2Bi2O6.