Conditional Generative Adversarial Networks for Inorganic Chemical Compositions

In recent years, inverse material design using machine learning techniques has attracted attention for material development. Almost all studies have used crystal structures of materials, although material engineers rarely store the crystal information and they only save chemical compositions and target properties for high-throughput materials discovery. Thus, we propose a method to generate chemical compositions for desired target properties by using conditional generative adversarial networks (CondGAN) and a post-processing method to balance the oxidation numbers. Numerical experimental results demonstrate that our CondGAN generates chemical compositions holding the desired properties.