MagGen: A graph aided deep generative model for inverse design of stable, permanent magnets

A significant development towards inverse design of materials with well-defined target properties is reported. A deep generative model based on variational autoencoder (VAE), conditioned simultaneously by two target properties, is developed to inverse design stable magnetic materials. Structure of the physics informed, property embedded latent space of the model is analyzed using graph theory, based on the idea of similarity index. The graph idea is shown to be useful for generating new materials that are likely to satisfy target properties. An impressive ~96% of the generated materials is found to satisfy the target properties as per predictions from the target learning branches. This is a huge improvement over approaches that do not condition the VAE latent space by target properties, or do not consider connectivity of the parent materials perturbing which the new materials are generated. In such models, the fraction of materials satisfying targets can be as low as ~5%. This impressive feat is achieved using a simple real-space only representation called Invertible Real-space Crystallographic Representation (IRCR), that can be directly read from material cif files. Model predictions are finally validated by performing DFT calculations on a randomly chosen subset of materials. Performance of the present model using IRCR is comparable or superior to that of the models reported earlier. This model for magnetic material generation, MagGen, is applied to the problem of designing rare earth free permanent magnets with promising results.