A scalable crystal representation for reverse engineering of novel inorganic materials using deep generative models

The efficient search for crystals with targeted properties is a significant challenge in materials discovery. The rapidly growing field of materials informatics has so far primarily focused on the application of AI/ML models to predict the properties of known crystals from their fundamental and derived properties as descriptors. In the last few years, deep learning-based approaches have spawned a slew of innovative data-driven materials research applications. Materials scientists have used these techniques for the reverse engineering of crystal structures for target applications. However, one of the challenges has been the representation of the crystal structures in the machine readable format. Proposed representations in the literature lack in generality and scalability. In this paper, we train a conditional variational autoencoder with a scalable and invertible representation along with the elemental properties of the constituents as descriptors to inverse-design new crystal structures with specified attributes. When targeting formation energy, we show that our model predicts structures that are not in the complete OQMD database. Finally, we use first-principles density functional theory calculations to validate our findings and show that the developed model is able to generate novel crystal structures for targeted property, i.e. formation energy in this case.