Structure Generation in the Representation Space

Data-driven methods have the capability to greatly accelerate the rate of materials discovery and design over conventional human-guided approaches. Generative models are one such recent example which could potentially be used to generate completely novel materials with specified functional properties. When applying generative models for determining atomic structures, a key prerequisite lies in using suitable structural fingerprints or representations for the machine learning model, analogous to the graph-based or SMILES representations used in molecular generation. These representations would need to be invariant to translations, rotations, and permutations, while remaining invertible back to their Cartesian coordinates. The challenges associated with simultaneously meeting both invariance and invertibility requirements have prompted us to propose an alternative approach to this problem by developing methods for accurately reconstructing the structure using optimization-based techniques which can be applied towards non-invertible representations. Our recent findings show this approach can reliably reconstruct atomic structures with high accuracy, and when paired with a generative model, can produce diverse structures with very high data efficiencies.