SHARE
Research Highlight

Crystal Structure Generation by Reconstructing Representations

Topic:

Brief: Researchers have developed a novel architecture-agnostic method for the generation of atomic structures for use in materials discovery.

Accomplishment: A primary prerequisite for developing generative models for atomic structures is an invertible input representation. Currently, there are no suitable invertible representations for crystal structures like the Simplified molecular-input line-entry system (SMILES) representation used in molecular generation, which significantly hinders progress in this area.  Here, researchers developed a computational approach to reconstruct non-invertible representations back into the Cartesian atomic coordinates which define a material exactly.  This is done by using a gradient-based global optimization algorithm to solve for the optimal atomic coordinates, using gradients obtained from automatic differentiation.  This can then be coupled to a generative machine learning model which generates new materials within the representation space, rather than the data-inefficient Cartesian space.  This two-step process is model-agnostic and does not require a specialized machine learning architecture for the generation of atomic structures, which allows one to harness effective existing generative algorithms such as invertible neural networks.  This approach thereby provides a promising solution for the problem of materials design at the atomic level. 

Crystal Structure Generation by Reconstructing Representations
The problem of atomic structure generation is framed as a two-step process, whereby a machine learning model such as an invertible neural network generates materials within the atomic representation space. Here, we represent the crystal structure with symmetry functions that incorporate geometric invariances for efficient learning. In the second step, we use our newly developed algorithm which maps from the symmetry function representation into the Cartesian coordinates, to then obtain the atomic-level structure of a material. This reconstruction process is illustrated in the bottom panel, whereby a gradient-based global optimization algorithm is used to update the positions of atoms until the minimum solution is found, which corresponds to the ground truth atomic structure.

Acknowledgement: This research is sponsored by the Artificial Intelligence Initiative as part of the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the US Department of Energy under contract DE-AC05-00OR22725.

Publications resulting from this work:  
Fung, V., Zhang, J., Yin, J., Ganesh, P., Atomic structure generation from reconstructing structural fingerprints. Under preparation.

Contact: Victor Fung (fungv@ornl.gov)

Team: Victor Fung, Jiaxin Zhang, Junqi Yin, Panchapakesan Ganesh

References: 

  1. Noh J., Gu G. H., Kim S., Jung Y. Machine-enabled inverse design of inorganic solid materials: promises and challenges. Chem. Sci. 11, 4871-4881 (2020).
  2. Fung V., Zhang J., Hu G., Ganesh P., Sumpter B. G. Inverse design of two-dimensional materials with invertible neural networks. npj Comput. Mater. 7, 200 (2021).