Ghost Atoms: Flexible Crystal Generation With Voronoi Tessellations

Generative models for materials discovery fix the number of atoms per unit cell before generation and, as a result, the model never decides how populated its own crystal should be and has no way to add or remove atoms. We restore this degree of freedom by padding every cell to a fixed number of atoms with fictitious ghost atoms, placed at well-spaced interstitial sites. We propose a technique of iterative Voronoi maximin selection so that the padded training structures contain no unphysical atomic overlaps. Training the OMatG stochastic-interpolant model on an augmented MP-20 with Voronoi-placed ghosts, we demonstrate improvement in combined SUN and mSUN rates by 3.2% over both the base model and a matched control that places ghosts uniformly at random. The Voronoi-trained model learns to keep ghosts out of occupied space, placing far fewer near real atoms, so that converting a ghost into a real atom happens at a vacant site. Our best model reaches 25.27% combined SUN and mSUN rate on LeMat-GenBench, a new state of the art.