PaperlayerView exact revision on arXiv

Source-linked AI summary

A Symmetry-orientated Divide-and-Conquer Method for Crystal Structure Prediction

Xuecheng Shao, Jian Lv, Peng Liu, Sen Shao, Pengyue Gao, Hanyu Liu, Yanchao Wang, Yanming Ma

arXiv:2109.13395v2cond-mat.mtrl-sciphysics.comp-ph

TL;DR

Crystal structure prediction requires global minimization over a rugged, high-dimensional energy landscape. The paper partitions the search by symmetry and uses an AI-based strategy to explore selected subspaces while excluding P1. It reports successful predictions for complex binary Lennard-Jones mixtures and high-pressure ice, including recovery of experimental ice structure with fewer optimized structures than LPSO required.

  • Problem

    Crystal structure prediction seeks global minima on rugged, high-dimensional energy landscapes with an enormous number of minima that grows exponentially with system size.

    Source: §I · p. 1

  • Method

    The method uses a hierarchical symmetry tree graph to partition structure space and a population-based, symmetry-constrained search that excludes P1 and assigns scout, onlooker, and employee agents.

    Source: §II · p. 2

  • Results

    208 optimized structures, versus LPSO's >750, identified experimental ice I41amd; the method also predicted BLJM structures more energetically favorable than those found by MH and BH.

    Source: §III · p. 5

  • Takeaways & Limitations

    The authors expect the method to be applicable to compositionally and structurally complex structures with large unit sizes, potentially thousands of atoms, as computational resources grow.

    Source: §IV · p. 5

Abstract

from arXiv · show

Crystal structure prediction has been a subject of topical interest, but remains a substantial challenge, especially for complex structures as it deals with the global minimization of the extremely rugged high-dimensional potential energy surface. In this manuscript, a symmetry-orientated divide-and-conquer scheme was proposed to construct a symmetry tree graph, where the entire search space is decomposed into a finite number of symmetry-dependent subspaces. An artificial intelligence-based symmetry selection strategy was subsequently devised to select the low-lying subspaces with high symmetries for global exploration and in-depth exploitation. Our approach can significantly simplify the problem of crystal structure prediction by avoiding exploration of the most complex P1 subspace on the entire search space and have the advantage of preserving the crystal symmetry during structure evolution, making it well suitable for predicting the complex crystal structures. The effectiveness of the method has been validated by successful prediction of the candidate structures of binary Lennard-Jones mixtures and high-pressure phase of ice, containing more than one hundred atoms in the simulation cell. The work, therefore, opens up an opportunity towards achieving the long-sought goal for crystal structure prediction of complex systems.

I. INTRODUCTION

Crystal structure prediction seeks energetically favorable structures, but global minimization is difficult because the energy landscape contains an enormous number of minima. The paper proposes symmetry-based decomposition and selection to reduce the search burden.

I. INTRODUCTION

  • CSP seeks structures that are energetically favorable and synthesizable, with energy corresponding to a global minimum on a high-dimensional potential-energy landscape.Source: §I · p. 1
  • The number of energy minima can be astronomical and grows exponentially with the number of atoms, making global minimization an NP-hard challenge.Source: §I · p. 1
  • Common CSP methods sample the entire potential-energy surface, but limited simulation budgets make this direct strategy insufficient, especially for unit cells with more than 50 atoms.Source: §I · p. 1
  • The proposed method constructs a symmetry tree graph to divide structure space into symmetry-dependent subspaces, exclude the complex P1 subspace, and use AI-based selection to explore promising subspaces.Source: §I · pp. 1–2

II. METHODS

The method organizes crystal structures into symmetry-dependent subspaces and uses a symmetry-preserving population search to explore them. Symmetry reduces structural degrees of freedom, while the agent strategy allocates exploration and exploitation across the hierarchy.

II. METHODS

  • For 100 atoms, the structural degrees of freedom range from 303 in triclinic symmetry to 23 in cubic symmetry; the estimated minima for 64 atoms are about 10^85 in P1 versus 10^6 in P2_13.Source: §II · p. 2, Figure 1 · p. 3
  • Random MgAl2O4 structures were 85.85% P1 and 14.15% other symmetries, illustrating the low-symmetry overrepresentation that motivates symmetry-directed search.Source: §II · p. 2, Figure 1 · p. 3
  • Earlier symmetry-based generation methods did not preserve symmetry during evolution, while some symmetry-adapted crossover methods required predefined cell parameters and applied only to specific systems.Source: §II · p. 2
  • The symmetry tree graph hierarchically partitions structures by space group, site-symmetry groups, and crystallographic orbits, with each level defining increasingly specific subspaces.Source: §II · p. 2
  • The population search initializes symmetry-constrained structures, excludes P1, ranks relaxed candidates by energy, and assigns scouts, onlookers, and employees to high-, middle-, and low-energy structures.Source: §II · p. 2
  • Scouts explore the 229 non-P1 space groups while avoiding repeats until all have been examined; onlookers explore combinations of Wyckoff positions within a space group.Source: §II · pp. 2, 4
  • Onlooker selection probabilities weight Wyckoff-position choices by energy-based fitness, while scout and employee roles can change as candidate energies evolve.Source: §II · p. 4
  • In the implementation, structures are generated under symmetry and distance constraints, duplicates are rejected by fingerprints, and variable-cell relaxation and ranking precede SPAI generation until termination.Source: §II · p. 4, Figure 3 · p. 4

III. RESULTS AND DISCUSSION

The method is reported to reproduce known structures efficiently and to identify candidate structures for compositionally complex materials, binary Lennard-Jones mixtures, and ice.

Known-structure benchmarks

  • For MgAl2O4 and SrTiO3, SPAI and LPSO both reproduce experimental structures with 100% success, while SPAI requires fewer structures on average.Table I summarizes 50 runs per system without cell-parameter constraints.Source: §III · p. 4, Table I · p. 5
  • For Mg3Al2Si3O12, SPAI achieves 100% success without cell-parameter constraints, averaging 393 structures; the earlier approach fails without biased experimental cell parameters and has 20% success.The cited comparison describes the earlier result for LPSO.Source: §III · p. 4

Compositionally complex materials

  • The method predicts the plausible ordered Ba2Ca2YFe5O13 structure with 92 atoms without prior experimental knowledge.The starting composition is Ba1.6Ca2.3Y1.1Fe5O13.Source: §III · p. 4

Binary Lennard-Jones mixtures

  • For BLJMs containing 60, 80, and 256 atoms, SPAI predicts structures with energies more favorable than those found by minima hopping and basin-hopping.The predicted structures include layered features, and the BLJM-256 unit cell exceeds 97 angstroms.Source: §III · p. 5, Table II · p. 5

Ice

  • At 10 GPa, the method reproduces the observed I41amd ice structure using 208 optimized structures, versus more than 750 for LPSO, and predicts a 144-atom Fddd structure favored at 600 K.At 0 K, Fddd is 2.0 meV/atom higher in static energy than I41amd.Source: §III · p. 5

Broader applicability

  • The authors expect the symmetry tree graph to be applicable beyond 3D crystals, including 0D, 1D, and 2D structures.For isolated 0D structures, they suggest using point groups C2, Cs, and C2v.Source: §III · p. 5

IV. CONCLUSION

The conclusion presents a crystal-structure-prediction method combining a hierarchical symmetry tree graph with a biased symmetry-adapted AI algorithm, demonstrated on complex systems with hundreds of atoms per cell.

IV. CONCLUSION

  • The method combines a hierarchical symmetry tree graph with a biased symmetry-adapted artificial intelligence algorithm.It directs search toward promising subsets for identification and in-depth exploitation.Source: §IV · p. 5

AUTHOR’S CONTRIBUTION

The available passages include an equal-contribution statement and fragments describing Figure 4.

AUTHOR’S CONTRIBUTION

  • X.S and J.L contributed equally to this work.Source: AUTHOR’S CONTRIBUTION · p. 5
  • Figure 4(a) compares a predicted 144-atom Fddd structure with the stable I41amd structure.Source: Figure 4 · p. 6
  • Figure 4(b) shows the Gibbs free-energy difference between Fddd and I41amd as a function of temperature up to 1,000 K.The caption specifies a harmonic approximation level.Source: AUTHOR’S CONTRIBUTION · p. 6

AUTHOR DECLARATIONS

The authors report no conflicts to disclose.

AUTHOR DECLARATIONS

Loading 2109.13395v2…