Strontium titanate grain boundaries stop being guesswork: machine-learning-potential sampling yields ten stable structures and four electronic regimes
Synopsis
The atomic configuration and composition of perovskite grain boundaries such as strontium titanate have been hard to pin down, because a multi-element system must balance bonding energies, elemental chemical potentials and electrostatics. Combining grand canonical Monte Carlo with a universal machine-learning interatomic potential sampled the 5(310)[001] boundary and produced ten stable structures in four motif families: large Sr kite, small Sr kite, incomplete large Sr kite and Ti-rich. Density functional calculations show that composition and structure govern band-edge placement, giving localized electron-doped, delocalized electron-doped, insulating and hole-attracting regimes.
Interpretation
The atomic structure and composition of perovskite grain boundaries can now be searched systematically across chemical-potential space instead of assuming one configuration and computing its energy. Earlier ab initio grand canonical Monte Carlo evaluated energies step by step with density functional theory, which is too costly for the large configuration space of multi-element perovskite boundaries, and the approach had mainly been applied to surface growth prediction. On the 5(310)[001] strontium titanate boundary, the sampled stable structures reproduce experimentally observed kite-shaped cation motifs, which the authors take as support for applying the approach to complex oxide interfaces.
The search treats the boundary as an open system that gains and loses atoms at fixed chemical potential, volume and temperature, adjusting composition and geometry through six kinds of random moves. The six moves are atomic displacement, atom addition, atom removal, position swap between two atoms, direct replacement of one species by another, and rigid translation of whole grains; replacement and grain translation were added because removal and addition become unlikely once the boundary region fills, and because boundary width is generally unknown. Energies and structural relaxation come from the Equiformer V2 86M universal machine-learning interatomic potential; across 280 distinct-composition 5(310)[001] structures, its free energies differ from density functional results by a mean absolute error of 24.71 meV/Ų and its atomic forces by 0.05 eV/Å.
Boundary geometry shifts the competition between small and large kite motifs, so different boundary planes and inclination angles favor different dominant motifs. The 5(310)[001] boundary accommodates both motif sizes, whereas the phase diagrams of the other two boundaries are simpler and dominated by a single size. The 5(210)[001] boundary shares the misorientation angle and coincident-site-lattice index of 5(310)[001] but has a different boundary plane and a larger inclination angle, and its phase diagram is largely occupied by small kites; the 17(410)[001] boundary has a smaller misorientation angle and a smaller inclination angle, and much of its phase diagram is occupied by distorted large kites.
Boundary composition and atomic reconstruction together set the spatial distribution of band-edge states, interfacial band alignment and carrier localization. A boundary that is charged or neutral by formal valence does not automatically map onto one electronic behavior, and GB0, which contains the same TiOx motif, also shows localized electron doping. Layer-projected densities of states in 2 Å layers normal to the boundary, plus partial charge densities, for the ten 5(310)[001] boundaries give localized electron-doped, delocalized electron-doped, insulating and hole-attracting states; the electrostatic analysis shows a short-range field of about 1 nm at the reconstructed core superimposed on a longer-range field from space charge, and Ti-rich GB1 also develops tail-to-tail local polarization.
Perspective
The discussion covers symmetric tilt boundaries 5(310)[001], 5(210)[001] and 17(410)[001] of strontium titanate, a prototypical paraelectric perovskite, plus one asymmetric boundary; the electronic and electrostatic analysis addresses the ten structures of the 5(310)[001] phase diagram using periodic supercells containing two equivalent boundaries.
The authors note that the periodic supercell calculations suggest cation-rich boundaries can accumulate electrons near the core while retaining more delocalized carriers in the grain; in experiments with much larger grains, a lower average donor concentration may instead leave grain interiors weakly conducting or insulating alongside electron accumulation or trapping at the boundary, depending on self-consistent band bending and occupation of boundary-derived states. A quantitative description would require self-consistent coupling of density functional theory with a continuum model.
