Generative Adversarial Networks (GANs) for inverse design of RuO 2 surfaces

Solving the longstanding puzzle of catalytically active RuO 2 structures in the oxidation of CO, has the potential to enable an efficient catalyst design for the conversion of CO exhausts in combustion processes. Reasons for the ongoing debate about catalytically active surface terminations of RuO 2 are that the experimental outcomes are strongly dependent on the catalyst pretreatment and the reaction conditions, as well as the limited computational resources in theoretical studies for the exploration of the RuO 2 potential energy surface. In recent theoretical studies, generative models have proven to be powerful tools for structure prediction of complex crystalline materials. To tackle the vast chemical space of possible surface terminations, we present a Generative Adversarial Network (GAN) that is capable of cheaply generating diverse structural guesses for novel RuO 2 surface structures. Two training sets, one with 28,903 and the other with 18,944 RuO 2 surface terminations, were created with a grand-canonical basin hopping method and a Gaussian Approximated Potential, respectively. The atomic positions of these structures were mapped as Gaussian densities to a three-dimensional grid for the GAN input. We demonstrate how two-dimensional images of RuO 2 structures with inferred lattice lengths and energy conditioning can be created in a two-dimensional Deep Convolutional Wasserstein-GAN (2D-DCWGAN) framework as a first step to realistic three-dimensional surface structures. The lattice lengths were predicted on-the-fly with two auxiliary networks in our GAN framework and the energy was ingrained in our latent space design. Additionally, the generation of realistic three-dimensional RuO 2 structures is incorporated in a three-dimensional Deep Convolutional Wasserstein-GAN (3D-DCWGAN) framework. These advances build the foundation which enables the implementation of realistic lattice lengths and an effective latent space design for the structure-energy-relationship in our 3D-DCWGAN framework in the future. ...