Material data are used to accelerate the development of materials. For the crystalline solid domain, however, informatics schemes are challenged by the encoding of crystal structures because embedding the invariances with respect to translation, rotation, and unit-cell choice is intractable by data augmentation schemes. We propose an efficient search space in which identical structures are reduced given such crystal operators. This crystal-invariant search space is created by morphing between the topologies of known crystal structures in accordance with the gradient of a smooth overlap of atomic positions descriptor formulated in reciprocal space as a metric. This crystal morphing is examined using an identification task of the structures from an x-ray diffraction spectrum. By applying Bayesian optimization on the search space, good initial structures for the Rietveld analysis are achieved via automatic determination without expert operations. Our method is applicable in the autonomous search of crystals, various structural analyses, and design of metamaterials.

• Received 22 June 2021
• Accepted 28 January 2022

DOI:https://doi.org/10.1103/PhysRevMaterials.6.023801