Inverse design of crystal structures for multicomponent systems

We developed an inverse design framework, constrained crystal deep convolutional generative adversarial networks (CCDCGAN), enabling automated generation of stable multicomponent structures. Their formation energy can be optimized in the latent space based on reversible images with continuous representation. After training by 52,615 structures from Materials Project, CCDCGAN model is able to generate new structure. It demonstrated that when 9160 are generated CCDCGAN, 8310 them have not been reported set. Detailed analysis 15 selected materials systems reveals unreported below convex hull discovered 6 them. Moreover, efficiency further improved considering extra hypothetical training. This paves way perform possible multi-objective optimization.