A Wyckoff-Aware Crystal Generative Flow Network for Materials Discovery

In this presentation, we will discuss an interpretable and controllable generative flow network to accelerate targeted design of inorganic crystals. Prior crystal generative models parametrize probability distributions over unit cells. In contrast, our model learns distributions over Wyckoff positions in the asymmetric unit, which generates the crystal upon applying the space group symmetry operations. This approach equips the model with probability densities that are invariant under space group actions without architectural restrictions. To accelerate discovery of materials with desired properties and convex hull construction for guided synthesis, our generation process optionally enables hard-constrained sampling from user-specified space groups or composition spaces at inference time. We will discuss the generative flow network’s potential to accelerate materials design through online and offline learning tasks with reward- and likelihood-based training objectives.