Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Accelerating crystal structure prediction by machine-learning interatomic potentials with active learning

Evgeny V. Podryabinkin, Evgeny V. Tikhonov, Alexander V. Shapeev, and Artem R. Oganov
Phys. Rev. B 99, 064114 – Published 27 February 2019
PDFHTMLExport Citation
Abstract
Authors
Article Text
  • INTRODUCTION
  • MACHINE LEARNING INTERATOMIC POTENTIALS
  • LEARNING ON-THE-FLY
  • CRYSTAL STRUCTURE PREDICTION WITH OUR…
  • CONCLUSION
  • ACKNOWLEDGMENTS
    • Supplemental Material
    References

    Abstract

    We propose a methodology for crystal structure prediction that is based on the evolutionary algorithm USPEX and the machine-learning interatomic potentials actively learning on-the-fly. Our methodology allows for an automated construction of an interatomic interaction model from scratch, replacing the expensive density functional theory (DFT) and giving a speedup of several orders of magnitude. Predicted low-energy structures are then tested on DFT, ensuring that our machine-learning model does not introduce any prediction error. We tested our methodology on prediction of crystal structures of carbon, high-pressure phases of sodium, and boron allotropes, including those that have more than 100 atoms in the primitive cell. All the the main allotropes have been reproduced, and a hitherto unknown 54-atom structure of boron has been predicted with very modest computational effort.

    • Figure
    • Figure
    • Figure
    • Figure
    • Figure
    • Received 22 February 2018
    • Revised 29 November 2018

    DOI:https://doi.org/10.1103/PhysRevB.99.064114

    ©2019 American Physical Society

    Physics Subject Headings (PhySH)

    1. Research Areas
    Atomic & molecular structureCrystal structurePotential energy surfaces
    Condensed Matter, Materials & Applied PhysicsAtomic, Molecular & Optical

    Authors & Affiliations

    Evgeny V. Podryabinkin1,*, Evgeny V. Tikhonov1,2,3, Alexander V. Shapeev1, and Artem R. Oganov1,4

    • 1Skolkovo Institute of Science and Technology, Skolkovo Innovation Center, Nobel St. 3, Moscow 143026, Russia
    • 2Sino-Russian Joint Center for Computational Materials Discovery, State Key Laboratory of Solidification Processing, School of Material Science and Engineering, Northwestern Polytechnical University, Xi'an, 710072, China
    • 3International Center for Materials Discovery, School of Material Science and Engineering, Northwestern Polytechnical University, Xi'an, 710072, China
    • 4Moscow Institute of Physics and Technology, 9 Institutskiy per., Dolgoprudny, Moscow Region 141701, Russia

    • *E.Podryabinkin@skoltech.ru

    Article Text (Subscription Required)

    Click to Expand

    Supplemental Material (Subscription Required)

    Click to Expand

    References (Subscription Required)

    Click to Expand
    Issue

    Vol. 99, Iss. 6 — 1 February 2019

    Reuse & Permissions
    Access Options
    CHORUS

    Article Available via CHORUS

    Download Accepted Manuscript
    Author publication services for translation and copyediting assistance advertisement

    Authorization Required

    Log In
    Other Options
    ×

    Images

    • Figure 1
      Figure 1

      The scheme of learning on-the-fly. An active selection algorithm estimates the degree of extrapolation for each configuration sampled. If it is high, then the configuration is learned. After this, the energy, forces, and stresses are calculated by MTP and returned to the relaxation process.

      Reuse & Permissions
    • Figure 2
      Figure 2

      Comparison of learning curves for pretrained MTP and for learning from scratch.

      Reuse & Permissions
    • Figure 3
      Figure 3

      The two-stage relaxation scheme.

      Reuse & Permissions
    • Figure 4
      Figure 4

      The best (lowest-energy) structures found with our method. The estimate for the time required to find these structures with DFT is based on the number of configurations treated by MTP in our search and the time required for vasp to process all these structures on a single core. The actual time spent with DFT can be up to 10 times less than indicated, because at early stages of structure relaxation, cheaper computational settings are usually applied.

      Reuse & Permissions
    • Figure 5
      Figure 5

      Comparison of the found 106-atom structure (upper figure) and the structure from [37] (lower figure).

      Reuse & Permissions
    ×

    Sign up to receive regular email alerts from Physical Review B

    Log In

    Cancel
    ×

    Search

    Article Lookup

    Paste a citation or DOI
    Enter a citation
    ×