CrystalGAN—Crystal Generative Adversarial Network for Periodic Materials Discovery

There are several online databases available that make available the periodic structure of a wide variety of crystal structures. These can play a crucial role in generative tasks to produce new, previously undiscovered periodic materials. Such a generative task can be achieved by using a Generative Adversarial Network(GAN) on ‘crystal images’, a representation built from other various voxel-based data that comprises of the 3D-fractional coordinates of each element in the crystal structure, that are stored in a separate ‘channel’, similar to the RGB representation of conventional images. We use a Wasserstein GAN to generate new materials, learning from Materials Project data. To take advantage of a larger subset of the MP dataset, we learn on all ternary materials and predict the lattice vector and the elemental identity of each channel of a given structure. We also preserve periodic, rotational, and translational invariance by also performing data augmentation. The trained GAN will allow us to reconstruct new materials from the different generative and prediction heads of the network.