Predicting Interface Structure using the Minima Hopping Method with a Machine Learning Interatomic Potential

TL;DR

This work develops a robust hybrid framework that combines the Minima Hopping Method with the Allegro machine-learning interatomic potential to predict atomic-scale SrTiO$_3$ interfacial structures. A defect-representative training strategy and random structure generation enable MLIP extrapolation without interface-specific retraining, and predictions are validated against experimental observations of SrTiO$_3$ grain boundaries. The method consistently identifies lowest-energy interfacial configurations across multiple stoichiometries for Σ3$(112)[110]$ and captures experimental-like structures for Σ5$(310)[001]$, demonstrating strong potential to bridge ab initio predictions with real interfaces. Limitations include the current MLIP’s inability to capture long-range electrostatics, motivating future work to incorporate explicit electrostatics for mesoscale interfacial phenomena.

Abstract

Predicting atomic-scale interfacial structures remains a central challenge in materials science due to their structural complexity and the difficulty of direct comparison between computational and experimental results. In this study, we present an efficient approach for interface structure prediction that integrates the Minima Hopping Method (MHM) with the state-of-the-art machine learning interatomic potential (MLIP), Allegro. We demonstrate that the MHM-Allegro approach provides a robust and computationally efficient route for predicting interfacial structures in the benchmark system SrTiO3 Sigma 3 (112)[110] tilt grain boundaries (GBs), consistently identifying the lowest-energy configurations across different stoichiometries. Furthermore, we introduce a strategy for constructing defect-representative training datasets without explicitly including defective configurations, achieving excellent extrapolative performance in interface predictions. The predictive capability is further validated through direct comparison with experimental observations of the SrTiO3 Sigma 5 (310)[001] GB, where the predicted atomic configurations show strong agreement with experimental measurements. This work represents a significant step toward bridging the gap between ab initio predictions and experimentally observed interfacial structures.

Figures (6)

• Figure 1: Workflow of MLIP preparation and MHM-based interface structure search.
• Figure 2: The dataset for MLIP. (a) Initial randomly generated structures and final selected configurations; and (b)–(d) the energy, force, and stress components in the final DFT dataset after DIRECT sampling
• Figure 3: Interfacial energies of the $\Sigma$3(112)[110] SrTiO$_3$ grain boundaries as a function of the chemical potential of the TiO$_2$ component. The first term in the legend denotes the structure search algorithm, and the second term specifies the surrogate PES model employed during the search. All interfacial energies were evaluated at the DFT level. Wherever available, reference structures were recalculated using the same DFT settings as employed in this work to eliminate numerical inconsistencies. Because explicit atomic structures from the AGA study were not provided, we identified MHM-predicted structures that are as consistent as possible with the reported AGA results, based on the evidence presented in the Supporting Information. These structures were therefore assigned as AGA reference structures, and their interfacial energies were recalculated accordingly. In some cases, the lowest-energy structure identified by MHM coincides with the AGA reference structure, as indicated by overlapping markers and the MHM line style. The lowest-energy grain boundary at a given chemical potential is highlighted by a color-matched halo.
• Figure 4: Representative example of the interface-structure search workflow. Ninety independent MHM runs were performed in parallel for the SrTiO$_3$$\Sigma$3(112)[110] $\Gamma = 3$ grain boundaries. Interfacial energies were evaluated at $\mu_{\mathrm{TiO_2}} - g_{\mathrm{TiO_2}}^{0} = -0.7$ eV using a trained Allegro potential. Five representative MHM trajectories are highlighted and ultimately converge to the lowest-energy configuration for $\Gamma = 3$; four out of five originate from medium-energy starting points.
• Figure 5: Interfacial energies of the $\Sigma$5(310)[001] SrTiO$_3$ grain boundaries as a function of the chemical potential of the TiO$_2$ component. All interfacial energies are calculated at the DFT level. The lowest-energy grain boundary at a given chemical potential is highlighted by a color-matched halo.