A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has developed a machine-learning framework that can significantly accelerate the process of discovering new materials, particularly for applications like carbon capture. The system, called DiffCSP, uses generative diffusion models to predict the atomic structure of stable materials from their chemical composition alone, a task traditionally requiring immense computational power and time. In tests, DiffCSP outperformed previous methods in both speed and accuracy, successfully generating known structures and proposing viable new ones. This advancement could dramatically shorten the years-long timeline for material discovery, opening faster pathways to developing better catalysts, battery components, and systems for removing carbon dioxide from the atmosphere. Read the full article at: https://technologyreview.com/2024/10/15/diffcsp-ai-material-discovery/