Atomic structure generation from reconstructing structural fingerprints

Abstract Data-driven machine learning methods have the potential to dramatically accelerate rate of materials design over conventional human-guided approaches. These would help identify or, in case generative models, even create novel crystal structures with a set specified functional properties then be synthesized or isolated laboratory. For structure generation, key bottleneck lies developing suitable atomic fingerprints representations for model, analogous graph-based SMILES used molecular generation. However, finding data-efficient that are invariant translations, rotations, and permutations, while remaining invertible Cartesian coordinates remains an ongoing challenge. Here, we propose alternative approach this problem by taking existing non-invertible desired invariances algorithm reconstruct through gradient-based optimization using automatic differentiation. This can coupled model which generates new within representation space, rather than data-inefficient space. In work, implement end-to-end generation atom-centered symmetry functions as conditional variational autoencoders model. We able successfully generate valid sub-nanometer Pt nanoparticles proof concept. Furthermore, method readily extended any structural representation, thereby providing powerful, generalizable framework towards structure-based