Speaker
Description
Dataset generation is a critical component for obtaining accurate machine-learned interatomic potentials (MLIPs). Molecular dynamics is widely used for reference data generation; however, it struggles to sample rare events such as transition states. This limitation is particularly problematic when fine-tuning foundation models, in which such configurations are typically underrepresented.
To address this challenge, we employ a recently developed algorithm, stochastic saddle point dynamics (SSPD) [1,2] to generate reference configurations without relying on predefined collective variables. This approach steers Langevin dynamics toward transition states, enabling efficient sampling not only at but also in the vicinity of these configurations. The use of SSPD has recently been demonstrated for identifying transition states in the decomposition of isopropanol and CO dissociation on Co surfaces [3].
Here, we apply SSPD to generate reference data near transition states associated with oxygen ion migration in perovskites. We demonstrate its capability to sample configurations close to transition states in SrTiO$_{3-\delta}$, enabling efficient fine-tuning of MACE foundation models to accurately predict migration barriers at the target level of theory. Furthermore, SSPD-driven active learning is used to perform larger-scale investigations of the more complex La$_{1-x}$Sr$_x$FeO$_{3-\delta}$ system across varying Sr concentrations. We observe substantial variations in migration barriers depending on the migration pathway, revealing a more intricate landscape than previously assumed for identifying efficient transport routes. Overall, SSPD provides a powerful addition to the dataset generation toolkit, addressing a key limitation of conventional approaches.
[1] S. Tănase-Nicola et al., 10.1103/PhysRevLett.91.188302
[2] T. Lelièvre et al., 10.1137/22M1541964
[3] M. Ketter et al., 10.26434/chemrxiv-2025-1psv7