Solving Crystals: A New Way to Predict Structures from Powder X-Rays

Powder X-ray diffraction, or PXRD, is a major tool used to study materials. While it’s commonly used, the part where humans have to manually analyze the data is tough, and automatic methods aren't great at fine details. Predicting exact crystal structures from PXRD has been a challenge. That's where XtalNet steps in. It's the first model of its kind that uses deep learning to predict crystal structures all the way from the PXRD data. Unlike older methods that only look at what the material is made of, XtalNet uses both the composition and the PXRD patterns. This helps it generate complex structures with up to 400 atoms in the unit cell. XtalNet works in two parts. First, it aligns the PXRD data with the actual crystal structure space using something called the CPCP module. Then, it generates possible crystal structures based on the PXRD patterns with the CCSG module. Tests on two sets of data (hMOF-100 and hMOF-400) showed XtalNet can predict the top 10 structures with high accuracy. It's like having a smart assistant that can predict crystal structures directly from experiments, without needing extra help or databases. This could speed up the discovery of new materials.