Scientists have been using X-ray crystallography for more than a century to determine the structure of crystalline materials such as metals, rocks, and ceramics. However, when dealing with powdered versions of materials, scientists face challenges in piecing together the overall structure. MIT chemists have developed a generative AI model that can simplify the process of determining the structures of powdered crystals. This model could assist researchers in characterizing materials for various applications, including batteries and magnets.
Crystalline materials are composed of lattices with repeating units, making up a distinct shape and size with atoms arranged precisely within. When X-rays interact with these lattices, they diffract off atoms, revealing information about their positions and bonds. Materials existing only as powdered crystals present difficulties in solving their structures due to the fragments lacking the full 3D structure of the original crystal. The MIT team trained a machine-learning model known as Crystalyze on data from the Materials Project database to predict structures based on X-ray diffraction patterns.
The Crystalyze model breaks down the process of predicting structures into several subtasks, from determining the size and shape of the lattice “box” to predicting the arrangement of atoms within it. This generative AI model generates multiple possible structures based on diffraction patterns, which can be tested by using a model to determine diffraction patterns for a given structure. The accuracy of the model was tested on both simulated and experimental diffraction patterns, with the model being correct about 67 percent of the time. They also successfully solved structures for previously unsolved diffraction patterns using this model.
By using their AI model, the researchers were able to identify structures for more than 100 previously unsolved diffraction patterns. They also discovered structures for three materials that their lab created by forcing non-reacting elements to form compounds under high pressure. This approach can lead to the development of new materials with different crystal structures and physical properties despite having the same chemical composition. The ability to determine the structures of powdered crystalline materials has implications for various fields, as it could aid researchers in designing new materials for applications such as permanent magnets.
This research, led by MIT chemists and computer scientists, was funded by the U.S. Department of Energy and the National Science Foundation. The researchers developed a web interface for their model, which is accessible at crystalyze.org. By utilizing generative AI technology, scientists can enhance their ability to determine crystal structures for a wide range of materials. This innovative approach has the potential to advance materials science research and accelerate the discovery of new materials for various applications.
