Machine Learning Generation of Actinide Materials

Actinide materials play an essential role in nuclear energy production, and their unique electronic structures, stemming from 5f electrons, provide a rich source for harboring exotic magnetic orderings. These properties hold promising applications in advanced fields such as quantum computing. However, research on candidate materials has long suffered from a data scarcity problem, with discovered actinide compounds representing only a tiny fraction of all possible materials. Additionally, the safety handling requirements of actinide elements pose significant barriers to experimental exploration, emphasizing the need for computation-aided discovery of actinide compounds. In this MRS seminar, we present our latest work using a generative model to discover new actinide, rare-earth, and transition metal compounds with specific lattice type constraints. By augmenting the diffusion process in a diffusion-based generative model, we have successfully generated actinide materials with actinide elements positioned on square, triangular, kagome, and hexagonal lattices. After a multi-staged down selector, additional density-functional theory (DFT) performed on 26,000 generated materials show that half of them are stable to the DFT level. This generative model not only expands the range of potential actinide materials but also accelerates their discovery, providing a pathway to overcoming the limitations imposed by experimental challenges and data scarcity.