From ExtremeTech: Using artificial intelligence to create new things is all the rage right now. Whether you want text, computer code, or images, there are uncountable generative AI models that can oblige. Google DeepMind announces that it has developed an AI model to generate something else: crystal structures. Its new Graph Networks for Materials Exploration (GNoME) system has successfully generated 2.2 million new crystalline materials, and the team thinks 380,000 of them are stable enough to be helpful in advanced technology.
Many gadgets we use daily rely on crystal materials—batteries, solar panels, processors, and more technology would not exist without inorganic crystals. However, discovering new crystal structures in a laboratory setting is a painstaking process of trial and error. GNoME is not the first attempt to automate this process, but DeepMind claims it's dramatically faster and more accurate.
GNoME is what's known as a graph neural network (GNN) model. The input data is essentially a graph, in this case, describing the connections between atoms. That makes GNNs an ideal way to generate and assess new crystalline structures. There are two computing pipelines in GNoME: a computational model that randomly assembles possible crystals based on chemical formulas and a structural model that creates new crystals based on known structures.
View: Full Article
