In an atomistic modeling study evaluating the capabilities of universal machine-learning interatomic potentials, researchers have determined that the pre-trained SevenNet potential establishes physical bounding thresholds rather than exact quantitative matches for ion transport in the cubic perovskite oxide strontium titanate. Computational investigations using molecular statics and molecular dynamics demonstrated that while the unrefined potential successfully captures fundamental transport mechanisms—such as the interstitial Grotthuss-type hopping of protons—it systematically underestimates the activation enthalpies of defect migration across multiple ionic species. Concurrently, the simulations yielded defect diffusion coefficients that notably exceed values established through experimental measurements, establishing that off-the-shelf implementations offer an upper limit for diffusivities and a lower threshold for migration barriers.
Atomistic simulations of solid-state diffusion traditionally face difficult computational compromises between accuracy and statistical reliability. Analytical pair potentials, such as those of the Buckingham form, easily accommodate the large supercells and long simulation times necessary to collect robust statistics for diffusion coefficients, yet they cannot model chemical bond breaking or formation and often fail to balance lattice reproduction with point-defect energetics. In contrast, quantum-mechanical density functional theory provides a high degree of physical accuracy and handles bond rearrangement, but high computational expense typically restricts such calculations to small supercells evaluated over very brief timeframes, leading to poor statistical resolution. Universal machine-learning potentials trained across extensive quantum-mechanical datasets have emerged to bridge this gap, aiming to combine quantum accuracy with the computational speed needed to evaluate extended solid-state ion transport.
To evaluate these capabilities, the researchers investigated three distinct defect transport processes in strontium titanate: the migration of interstitial protons, the diffusion of oxygen vacancies, and the hopping of strontium vacancies. Strontium titanate serves as an established benchmark system in defect chemistry and materials science, possessing a wide body of existing experimental and computational literature against which predicted activation barriers and absolute diffusion coefficients can be systematically compared. The research specifically assessed the graph neural network potential SevenNet, which was chosen due to its massively parallelized architecture capable of handling large-scale molecular dynamics in the LAMMPS code and its previously demonstrated benchmark accuracy in reproducing migration barriers in solid lithium-ion conductors.
Before exploring defect dynamics, the team evaluated the baseline structural and thermodynamic properties of the defect-free strontium titanate lattice modeled with SevenNet. At temperatures below 105 Kelvin, strontium titanate experimentally exhibits a tetragonal structure with I4/mcm symmetry before transitioning into its familiar cubic Pm-3m phase at higher temperatures. The potential successfully reflected this behavior: static energy minimization placed the tetragonal ground state 3.4 millielectronvolts per formula unit lower in energy than the cubic phase at zero Kelvin, and molecular dynamics confirmed the tetragonal structure was more stable at 100 Kelvin while accurately predicting the cubic phase at higher temperatures. Between 300 and 1000 Kelvin, the predicted linear thermal expansion coefficient was slightly higher than experimental findings, yielding 13.2 multiplied by 10 to the negative sixth per Kelvin compared to measured literature ranges between 9 and 12 multiplied by 10 to the negative sixth per Kelvin.
Calculations of the isobaric heat capacity, derived by differentiating a cubic spline fit of the internal energy across multiple simulation runs, showed close agreement with experimental and computational references up to 1700 Kelvin. Above this temperature, however, the simulation results exhibited pronounced positive deviations from linear expansion and an anomalous surge in heat capacity. Utilizing Wigner-Seitz defect analysis implemented via the OVITO software, the authors identified that spontaneous anti-Frenkel disorder emerged in the simulation cells between 1750 and 2250 Kelvin, generating pairs of oxygen vacancies and oxygen interstitials in the pristine lattice. A weighted linear regression analysis yielded an anti-Frenkel formation enthalpy of 7.2 electronvolts and an entropy of 18.4 Boltzmann constants. Comparing these numbers with broader ceramic literature indicated that the disorder entropy was significantly overestimated, leading to excessive spontaneous oxygen defect generation that produced computational artifacts at elevated temperatures.
The study then turned to cation transport, investigating the migration of strontium vacancies using nudged elastic band calculations implemented through the Atomic Simulation Environment. Because host cation diffusion in strontium titanate involves high activation energies and remains exceptionally slow over accessible nanosecond molecular dynamics timescales, direct jump observations were inaccessible without rare-event sampling. For isolated strontium vacancies, SevenNet predicted a direct linear migration path between initial and final atomic sites, matching geometric paths reported in earlier computational studies. However, the resulting migration activation energy of 3.2 electronvolts proved noticeably lower than previous experimental and computational values, which consistently span 3.5 to 4.1 electronvolts. For the vacancy associate pair composed of a bound oxygen vacancy and strontium vacancy, the potential predicted a curved trajectory and an activation barrier of 2.7 electronvolts, which also fell below prior simulation estimates of 2.9 and 3.41 electronvolts.
Oxygen transport, characterized by far higher mobile defect concentrations, was subsequently investigated via molecular dynamics in an eight-by-eight-by-eight supercell containing seven randomly introduced oxygen vacancies, corresponding to a dilute site fraction of 0.46 percent. Production runs were executed in the isothermal-isobaric ensemble across thirteen temperatures spanning 1350 to 2200 Kelvin. Oxygen tracer diffusion coefficients were derived from the linear mean squared displacement curves between 100 and 400 picoseconds using the Einstein relation. To prevent thermally generated anti-Frenkel defects from distorting the dilute concentration baseline, defect diffusion coefficients were evaluated strictly from data gathered below 1700 Kelvin, incorporating a theoretical tracer correlation coefficient of 0.69.
The resulting oxygen vacancy diffusion coefficients determined from the SevenNet simulations were consistently higher than experimental literature values, while the calculated migration activation enthalpy was 0.47 electronvolts, supported by a nudged elastic band result of 0.474 electronvolts. This activation enthalpy is substantially lower than the accepted experimental consensus range of 0.62 to 0.67 electronvolts. The authors emphasized that the migration barriers obtained via density functional theory vary widely depending on the chosen exchange-correlation functional and supercell size, citing published calculations where functional choices shifted barriers from 0.30 to 0.64 electronvolts and where odd-numbered supercell dimensions artificially restricted cooperative octahedral tilting. Consequently, relying directly on unvalidated quantum datasets to train machine-learning models risks perpetuating functional-dependent inaccuracies into downstream potential predictions.
Proton transport was modeled in a six-by-six-by-six supercell containing two interstitial hydrogen atoms, using 150 independent molecular dynamics trajectories initialized with distinct velocity distributions under the canonical ensemble at temperatures down to 800 Kelvin. SevenNet accurately stabilized the introduced hydrogen atoms at an equilibrium oxygen-hydrogen bond distance of approximately 1 angstrom, confirming that the potential inherently treats hydrogen as an interstitial proton without artificial bonding constraints. When hydrogen was initialized on an oxygen site, it spontaneously separated into an interstitial proton and an oxygen vacancy. The trajectory analysis showed that protons diffused via a classical Grotthuss mechanism, alternating between rotational reorientations around a host oxygen ion and transfer jumps across adjacent oxygen centers, yielding an activation enthalpy of 0.20 electronvolts and absolute diffusivities substantially exceeding published measurements.
These systematic findings delineate both the utility and the limitations of deploying pre-trained machine-learning interatomic potentials in materials discovery. Without system-specific fine-tuning, the SevenNet potential cannot deliver precise quantitative matches to experimental migration barriers or absolute defect diffusion rates in strontium titanate. However, because it consistently behaves as a physical bounding model—providing a lower limit for activation barriers and an upper limit for diffusion coefficients while naturally handling multi-species defect chemistry—it offers a viable computational screening tool. Researchers can rapidly screen and rank candidate perovskite oxide compositions for ionic transport prior to investing computational resources into functional-specific fine-tuning or experimental synthesis.
Subject of Research: Atomistic evaluation of the SevenNet machine-learning potential for ion defect transport in strontium titanate
Article Title: Simulating H\(^+\), O\(^{2-}\) and Sr\(^{2+}\) transport in SrTiO\(_3\) with a machine-learning potential
Article References: Perkampus, C., Adams, S., & De Souza, R. A. (2025). Simulating H$$^+$$, O$$^{2-}$$ and Sr$$^{2+}$$ transport in SrTiO$$_3$$ with a machine-learning potential. Journal of Materials Science: Materials in Energy, 1(1), Article 8. https://doi.org/10.1007/s44308-025-00011-0
Image Credits: AI Generated
DOI: 10.1007/s44308-025-00011-0
Keywords: strontium titanate, machine-learning potentials, SevenNet, molecular dynamics, nudged elastic band, oxygen vacancy diffusion, proton transport, strontium vacancy migration, defect activation enthalpy, anti-Frenkel disorder, perovskite oxides, computational materials science, LAMMPS, Grotthuss mechanism, density functional theory, diffusion coefficients, atomistic simulation
News Source: Teresa Odom. (November 21, 2026). SevenNet Machine Learning Potential Sets Defect Transport Boundaries in Strontium Titanate. Scienmag.



