Researchers at the Massachusetts Institute of Technology (MIT) have developed a new generative AI model, capable of revealing the structures of crystalline materials from powdered samples. This innovation, detailed in a study published in the Journal of the American Chemical Society, could enhance the development of new materials, for applications ranging from batteries to magnets.
The model, named Crystalyze, addresses a long-standing challenge in X-ray crystallography. While this century-old technique works well with intact crystals, it has been notoriously difficult to determine structures from powdered samples, which contain random crystal fragments.
Danna Freedman, the Frederick George Keyes Professor of Chemistry at MIT and a senior author of the study, emphasises the importance of this breakthrough: "Structure is the first thing that you need to know for any material. It's important for superconductivity, it's important for magnets, it's important for knowing what photovoltaic you created."
The research team, which includes Jure Leskovec from Stanford University and lead authors Eric Riesel (MIT graduate student) and Tsach Mackey (Yale University undergraduate), trained their AI model on data from the Materials Project, a database containing over 150,000 materials.
Crystalyze operates by breaking down the structure prediction process into several steps. It first determines the size and shape of the crystal's lattice "box" and its atomic composition. Then, it predicts the arrangement of atoms within this box. For each diffraction pattern, the model generates multiple possible structures, which can be verified against known diffraction patterns.
The model's accuracy was tested on both simulated and experimental data. It achieved 67% accuracy on experimental diffraction patterns from the RRUFF database, which were not included in the training data. More impressively, the researchers used Crystalyze to propose structures for over 100 previously unsolved patterns from the Powder Diffraction File, a comprehensive database of solved and unsolved materials.
Freedman's lab also employed the model to discover structures for three new materials they created under high-pressure conditions. These bismuth-based compounds could have potential applications in the design of new permanent magnets.To facilitate widespread use, the research team has made a web interface for the model available at crystalyze.org.