Composition Search in Multi-Component Alloys through Difussion Generative Models

Exploring the vast composition space of multi-component alloys present a challenging task for both ab initio (first principles) and experimental methods due to the multitude of possible compositions and the time-consuming procedures involved. These limitations ultimately impede the discovery of new stable materials with exceptional properties. Here, we utilize the Crystal Diffusion Variational Autoencoder (CDVAE) to characterize the stable compositions of NiFeCr alloys, including well-established compositions such as NiFe and CrNi2 from binaries, and then optimize the bulk elastic modulus across the entire compositional space. To this end, a computationally efficient framework is proposed which employs Molecular Dynamics (MD) simulations, equipped with high-quality interatomic potentials for inverse design of multi-component alloys in general. We also propose modifications to CDVAE that enhance its robustness when dealing with data characterized by large supercells and low variety of elements as well different material phases. The resulting workflow is valuable for the inverse design of other systems of interest involving various element types and varieties.