Generative Modeling for the Design of Polycrystalline Materials

Polycrystalline materials represent a large class of materials encompassing most metals, ceramics, and some polymers. Their performance is of key interest in a wide range industries/applications such as aerospace, automotive, biomedical, and others. Despite their importance, computational tools for the design of polycrystalline materials have remained lacking due to several fundamental barriers: 1) The expense and consequential scarcity of obtaining materials data. 2) The high dimensionality of microstructure information (a materials microscopic structure), and 3) the complex local states which characterize polycrystals (the pointwise orientation of their atomic lattice). This talk will cover the use of deep generative models for design under these constraints. Topics will include the use of conditional normalizing flows within active learning protocols, training denoising diffusion models on physical systems with inherent symmetries, and techniques to minimize the need for experimental data when training generative models