Due to the ever-growing demand for advanced materials with tailored properties, materials design remains a highly important and rapidly developing field. In this work, we present EnvXGen─a new method for the search of stable crystalline materials in cases where the positions of some building blocks (atoms or molecules), i.e., the environment, may be known from the experiment. To generate the initial set of structures, a random arrangement of structural elements at the nodes of a spatial grid, superimposed on the unit cell, is used, taking into account the interatomic distances. After crystal structure relaxation, clustering, and similarity analysis, the structure with the lowest energy and the highest similarity to the initial environment is considered as the best one. EnvXGen was verified on the known stable phases of atomic crystals, such as superconducting hydrides UH8, LaH10, and H3S, and energetic molecular cocrystals (CL-20 + CO2 and CL-20 + N2O), and showed high accuracy and good convergence. Using the developed method, we also found the positions of hydrogen atoms in the recently synthesized Y2H9 hydride. Also, we proposed an approach to study the divergence of the generated crystal structures using graph neural network embeddings, clustering, and dimensionality reduction techniques.
Propad et al. (Tue,) studied this question.