PaperlayerView exact revision on arXiv

Source-linked AI summary

Crystal Structure Prediction via Particle Swarm Optimization

Yanchao Wang, Jian Lv, Li Zhu, Yanming Ma

arXiv:1008.3601v1cond-mat.mtrl-sci

TL;DR

Crystal structure prediction requires navigating a huge number of energy minima, making reliable searches difficult. This paper proposes particle swarm optimization for global free-energy minimization and reports successful application across known elemental, binary, and ternary systems with diverse bonding environments.

  • Problem

    Crystal structure prediction is difficult because it requires classifying a huge number of energy minima on the lattice energy surface.

  • Method

    The paper develops a particle-swarm-optimization methodology that globally searches free-energy surfaces for crystal structures from chemical composition under specified external conditions.

  • Results

    The method was successfully applied to known elemental, binary, and ternary structures spanning metallic, ionic, and covalent bonding environments.

  • Takeaways & Limitations

    The reported success rate demonstrates the methodology’s reliability and supports PSO as a promising technique for crystal structure determination.

Abstract

from arXiv · show

We have developed a powerful method for crystal structure prediction from "scratch" through particle swarm optimization (PSO) algorithm within the evolutionary scheme. PSO technique is dramatically different with the genetic algorithm and has apparently avoided the use of evolution operators (e.g., crossover and mutation). The approach is based on a highly efficient global minimization of free energy surfaces merging total-energy calculations via PSO technique and requires only chemical compositions for a given compound to predict stable or metastable structures at given external conditions (e.g., pressure). A particularly devised geometrical structure factor method which allows the elimination of similar structures during structure evolution was implemented to enhance the structure search efficiency. The application of designed variable unit cell size technique has greatly reduced the computational cost. Moreover, the symmetry constraint imposed in the structure generation enables the realization of diverse structures, leads to significantly reduced search space and optimization variables, and thus fastens the global structural convergence. The PSO algorithm has been successfully applied to the prediction of many known systems (e.g., elemental, binary and ternary compounds) with various chemical bonding environments (e.g., metallic, ionic, and covalent bonding). The remarkable success rate demonstrates the reliability of this methodology and illustrates the great promise of PSO as a major technique on crystal structure determination.

I. INTRODUCTION

Crystal structure is central to linking composition with material properties, yet structure prediction from composition alone remains difficult because it requires classifying many energy minima. The paper proposes particle swarm optimization (PSO) as a reliable, computationally efficient methodology for predicting diverse crystal structures.

  • Motivation: Crystal structure is central to materials science because solid-state properties are intimately tied to structure and basic composition.
  • Motivation: Experimental structure determination can fail when X-ray diffraction data are low quality, particularly under extreme conditions such as high pressure.
  • Challenge: Prediction from chemical composition alone is difficult because it involves classifying a huge number of energy minima on the lattice energy surface.
  • Prior methods: Existing approaches have limitations: some require starting structures near the global minimum, while data mining depends on extensive databases and cannot generate new structure types without similar-compound information.
  • Contribution: The paper proposes crystal structure prediction based on PSO within an evolutionary scheme, significantly reducing first-principles density functional calculation expense through efficient global optimization.
  • Results: Successful predictions for elemental, binary, and ternary compounds demonstrate the methodology’s reliability and promise as a major tool for crystal structure determination.

II. METHOD AND IMPLEMENTATION

CALYPSO applies particle swarm optimization to crystal structure prediction through a four-step cycle of symmetry-constrained generation, local optimization, uniqueness screening, and PSO-based iteration. Symmetry constraints, variable cell sizes, and geometrical structure factors reduce the search cost and improve convergence.

  • Workflow: CALYPSO’s crystal-structure prediction workflow generates symmetry-constrained random structures, locally optimizes them, identifies unique minima, and creates new structures through PSO iteration.The four steps are listed explicitly as structure generation, local structural optimization, geometrical-structure-factor post-processing, and PSO-based structure generation.
  • Symmetry-constrained generation: Symmetry-constrained generation within 230 space groups reduces the search space and optimization variables, accelerating global structural convergence.Symmetry checking also forbids identical symmetric structures, enabling diverse structures important for global minimization.
  • Variable cell size: Variable cell sizes are selected intelligently during structural evolution, significantly reducing the computational cost of searching for the global minimum.This avoids separately simulating all possible cell sizes and comparing their resulting structures.
  • Optimization and PSO evolution: Free energy serves as the fitness function; local optimization refines atomic coordinates and lattice parameters before PSO generates the next population.PSO updates particle positions using velocity, individual best location, and population global best location.
  • Structural uniqueness screening: The geometrical structure factor identifies similar structures from interatomic distances grouped by bond types, discarding equivalent candidates within preset tolerances.For binary systems, the evaluated bond types are A-A, A-B, and B-B; retained structures update comparison matrices after local optimization.

III. APPLICATION AND RESULTS

The methodology was benchmarked with DFT calculations on elemental, binary, and ternary systems of known structure. CALYPSO rapidly reproduced stable experimental structures and identified metastable phases using relatively few generated structures and local optimizations.

  • Benchmark setup: The benchmark covered elemental, binary, and ternary compounds with known structures using density functional theory and the all-electron projector augmented wave method in VASP.Basic CALYPSO parameters are reported in Table I, while benchmark systems appear in Tables II and III.
  • Elemental systems: Lithium structures were correctly predicted across complex pressure-induced phases with fewer than 300 generated structures; cI16 appeared in generation 6 after 210 structures.The cI16 search used a population size of 30, with each generation producing and locally optimizing structures.
  • Elemental systems: Simulations rapidly reproduced experimental structures of carbon, silicon, magnesium, and other elements, while predicting carbon metastable phases at 0 GPa and silicon bc8 at 2 GPa.These results indicate that the method can predict metastable structures.

IV. CONCLUSION

The study developed a CALYPSO methodology combining particle swarm optimization with ab initio structural optimization to search lattice free-energy space for ground-state and metastable structures. Its efficiency derives from symmetry constraints, geometrical structure-factor filtering, and variable cell-size selection, while applications demonstrate high efficiency and success across diverse compounds and bonding environments.

  • Methodology: The methodology combines PSO within an evolutionary scheme, ab initio structural optimization based on density functional theory, symmetry-constrained structure generation, and geometrical structure-factor elimination of similar structures.These components are implemented in the CALYPSO code.
  • Methodology: The approach efficiently searches lattice free-energy space and atomic configurations for ground-state and metastable structures in complex systems.The methodology targets structural prediction from the solid’s lattice geometry and atomic configuration.
  • Efficiency: Variable cell-size selection identifies appropriate cell sizes and significantly reduces computational cost.The technique enables intelligent selection of correct cell sizes during the search.
  • Results: The methodology successfully predicts known experimental structures in elemental, binary, and ternary compounds with metallic, ionic, and covalent bonding.The reported applications establish broad coverage across compositions and bonding environments.
  • Results and outlook: The method is reported to have high efficiency and a high success rate, with future development considered feasible for systems of ~100 atoms/cell or above.Such expansion is expected to support structure solutions for nanomaterials, surfaces or thin films, biomaterials, and materials design.

Table I

Table I reports the key PSO structure-generation and sampling parameters, including distance, generation proportion, k-point precision, and confidence factors.

  • 0.8 (Å): Minimal interatomic distances.
  • 0.6: Proportion of the structures generated by PSO.
  • 0.04-0.1: Precision of k-points sampling.
  • 2.0: Self confidence factor (c1).
  • 2.0: Swarm confidence factor (c2).

Table II

Table II reports the PSO structure-generation settings and outcomes for four elemental systems at specified pressures, using populations of 30 structures. The number of generated structures varies by system, from 1 to 30.

  • Table II: Li at 0 GPa produced 1 Bcca structure with a population size of 30.The table lists Li, pressure 0 GPa, structure Bcca, generation 1, and population size 30.
  • Table II: Si at 2 GPa produced 6 Bc8g structures with a population size of 30.The table lists Si, pressure 2 GPa, structure Bc8g, generation 6, and population size 30.
  • Table II: Mg at 0 GPa produced 6 Hcpn structures with a population size of 30.The table lists Mg, pressure 0 GPa, structure Hcpn, generation 6, and population size 30.

Table III · FIG. 2 · FIG. 3

The supplied passages list Table III entries for six systems, including pressure, reported structure, generation count, and population size. No substantive passages for FIG. 2 or FIG. 3 are provided.

  • Table III: 150 GPa is listed for the Rock salt structure, with 2 generations and a population size of 30.The entry is marked with superscript g.
  • Table III: TiH2 at 0 GPa is listed with I4/mmmi, 2 generations, and a population size of 20.The structure is marked with superscript i.
  • Table III: MoB2 at 0 GPa is listed with R-3mk, 1 generation, and a population size of 30.The structure is marked with superscript k.
  • Table III: TiB2 at 0 GPa is listed with AlB2-typel, 1 generation, and a population size of 30.The structure is marked with superscript l.
  • Table III: 120 GPa is listed for MgSiO3 with Cmcmm, 5 generations, and a population size of 20.The structure is marked with superscript m.
  • Table III: CaCO3 at 0 GPa is listed with Calciten, 13 generations, and a population size of 30.The structure is marked with superscript n.
Loading 1008.3601v1…