Illustrative stock photograph of a natural quartz crystal cluster from Tibet, shown as an everyday example of a crystal.
Photo: JJ Harrison, via Wikimedia Commons, licence Creative Commons Attribution-ShareAlike 2.5 (https://creativecommons.org/licenses/by-sa/2.5/). File page: https://commons.wikimedia.org/wiki/File:Quartz,_Tibet.jpg
What it is
Many technologies depend on crystals: batteries, magnets, catalysts and materials that capture carbon dioxide. Finding a new crystal with a useful property has usually meant testing known materials one after another. MatterGen, built by Microsoft Research, works the other way round. It is a diffusion model, the same family of artificial intelligence used by image generators: it starts from a random arrangement of atoms and refines it step by step into a crystal. It learned from about 608,000 stable crystal structures taken from two open databases of materials, the Materials Project and Alexandria. After extra training, it can be asked for a crystal with a certain chemistry, symmetry or property, such as how strongly the material resists being squeezed. The code and training data are published openly.
Why it matters
Screening known materials can only find what has already been catalogued, and it runs out of candidates. In the Microsoft Research blog, a generative model kept finding new candidates for very hard-to-squeeze materials while a screening search ran out of known ones. The Nature paper reports that, compared with earlier generative models, MatterGen’s crystals are more than twice as likely to be new and stable, and they sit more than ten times closer to the nearest settled, lowest-energy arrangement of their atoms, a sign that they are closer to a stable form.
How to read it
The laboratory test is the number to look at, because computer results are only predictions. The team asked MatterGen for a material that resists squeezing with a strength of 200 gigapascals (a gigapascal is a unit of pressure). Thousands of computer designs were filtered down to 75, and the researchers chose four to make. One was made successfully: a compound of tantalum, chromium and oxygen. Testing the sample gave an estimated value of up to 169 gigapascals, which is within 20 percent of the target. The other results in the paper, such as stability, are computer calculations, so this laboratory test shows how a real sample behaves.
Limits
Only one of the four chosen designs could be made, and the real material had its tantalum and chromium atoms mixed at random, where the computer design had them in an orderly pattern. Most of the evaluation is calculation, and the authors write that real-world use needs more than these tests. The model also produces crystals with very low symmetry more often than the training data does, especially for larger crystals. The paper’s own summary calls the experiment a proof of concept.
