Crystal structure prediction of materials with high symmetry using differential evolution

Wenhui Yang, Edirisuriya M. Dilanga Siriwardane, Rongzhi Dong, Yuxin Li, Jianjun HuarXiv 2021 · arXiv

Introduction

Crystal Structure Prediction (CSP) is often described as the "holy grail" of materials informatics. By determining the arrangement of atoms within a crystal, researchers can predict essential material properties—such as conductivity, hardness, or thermal stability—before conducting expensive laboratory experiments. A promising recent approach in CSP utilizes contact maps, which are 2D matrices representing the adjacency of atoms within a structure.

However, high-symmetry crystals (e.g., those in Cubic or Tetragonal systems) present a unique challenge. In these materials, atoms often occupy "special positions" where symmetry operations like rotation or reflection cause multiple potential sites to collapse into one. Standard optimization models often struggle with these constraints, leading to "dimension mismatch" where the predicted number of atoms does not align with the target. This paper introduces CMCrystalHS, a framework designed to overcome these hurdles using symmetry-aware templates and Differential Evolution.

Methodology

The CMCrystalHS (Contact Map based structure prediction of Crystals with High Symmetry) workflow is structured into three primary stages to ensure both symmetry preservation and optimization efficiency:

  1. Symmetric Template Generation: The process begins by using the PyXtal library to generate 50 random candidate structures. These templates are constrained to strictly obey the target chemical formula and space group symmetry from the outset.
  2. Smart Filtering: Candidates are screened to select the best starting template. The algorithm prioritizes structures with the highest initial contact map accuracy and the fewest Wyckoff positions, which corresponds to the highest degree of symmetry.
  3. Differential Evolution (DE) Optimization: The atomic coordinates are refined using a DE algorithm via the Nevergrad platform. A critical innovation here is the "locking" of special fractional coordinates (e.g., values like 0, 0.25, or 0.5 that define symmetry). By only allowing the optimizer to adjust "non-special" coordinates, the framework prevents the structure from losing its symmetry during the search process.
The CMCrystalHS workflow: from target crystal structure to predicted structure via template generation and differential evolution.

Results and Discussion

The researchers validated CMCrystalHS using 10 diverse test cases from the Materials Project database, specifically targeting high-symmetry space groups (ranging from 42 to 221). Notable examples include $Ce_4Se_8$ and the highly symmetric cubic structure $La_3Al_1N_1$.

For several targets, including $Fe_4O_4F_4$ and $Li_2Cr_2O_4$, the algorithm improved contact map accuracy from an initial random state (~66%) to a perfect 100% match. Furthermore, the Root Mean Square Distance (RMSD) error—measuring the deviation of predicted atoms from their actual positions—was reduced by up to 39.67%.

Comparison of predicted and real contact maps alongside Wyckoff positions for Ce2As2O6, illustrating the framework's ability to handle complex symmetry.

Key Insights

  • Symmetry as a Search-Space Reducer: While symmetry often complicates the mathematics of CSP, CMCrystalHS treats it as a beneficial constraint. By fixing special coordinates, the optimizer has significantly fewer variables to solve, accelerating convergence.
  • Eliminating Dimension Mismatch: By ensuring the initial templates are symmetry-consistent, the framework avoids the primary failure mode of previous contact-map-based methods on high-symmetry materials.
  • Optimization Efficiency: The process is computationally lean, typically requiring only 300 to 1000 seconds per structure to reach a solution.

Conclusion

CMCrystalHS demonstrates that integrating symmetry constraints directly into the optimization loop significantly enhances the reliability of crystal structure prediction for complex materials. While the current model assumes known space groups and fixed lattice constants, it serves as an excellent "front-end" for more computationally intensive methods. Future iterations may integrate lattice parameter optimization and direct coupling with Density Functional Theory (DFT) simulations to further verify the energetic stability of the predicted high-quality candidates.

Summary updated

BibTeX
@misc{yang2021crystal,
  author = {Wenhui Yang and Edirisuriya M. Dilanga Siriwardane and Rongzhi Dong and Yuxin Li and Jianjun Hu},
  title = {{Crystal structure prediction of materials with high symmetry using differential evolution}},
  year = {2021},
  eprint = {2104.09764},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.09764},
}
Figure