A new AI framework can predict how atoms arrange themselves in a crystal without borrowing structural templates from existing databases.
NextCrystal, described in a preprint from researchers working on crystal structure prediction (CSP), combines a large language model with a diffusion model to go from a material's chemical composition directly to a plausible 3D atomic arrangement. The LLM encodes chemical meaning and generates what are called Wyckoff site patterns — the positions atoms are allowed to occupy given a crystal's symmetry rules. A beam search algorithm then checks that those assignments stay algebraically consistent with how many of each atom are actually present. The diffusion model takes over from there, filling in exact atomic positions while staying within that symmetry-consistent skeleton.
CSP is notoriously hard: the number of possible arrangements grows combinatorially with composition, and most existing tools either cheat by pulling known structures from a database or struggle to enforce physical symmetry rules at scale. NextCrystal sidesteps the lookup problem entirely and claims state-of-the-art scores on the standard stability, uniqueness, and novelty benchmarks. As a proof of concept, the team used it to screen candidates for hafnium dioxide (HfO2) and turned up a previously unreported stable crystal phase sitting 0.056 eV/atom lower in energy than a known high-pressure form — a meaningful gap in materials terms.
HfO2 is already commercially relevant in semiconductor gate dielectrics, so a new stable phase is not a purely academic result. Still, a preprint benchmark win and one illustrative find are a long way from displacing the experimental validation that materials discovery actually requires.