Dataset merging strategy for efficient construction of a machine-learning force field: A case of CeO2 surfaces with an oxygen vacancy

The accuracy and transferability of machine-learning force fields (MLFFs) critically depend on the quality of the training database. Here, we propose a dataset merging strategy for efficiently constructing a transferable MLFF for pristine and oxygen-defective CeO2 surfaces. The proposed strategy is based on three key elements: (1) initiating on-the-fly training from small, less stable slab models to accelerate early stage convergence by sampling diverse high-energy configurations; (2) sampling diverse local structures by gradually increasing the temperature during molecular dynamics simulations while avoiding excessively rapid atomic motions and unphysically short contacts; and (3) merging independently generated datasets from slab models with different thicknesses and defect conditions to improve transferability. The resulting MLFF is validated on larger slab models containing 144 atoms and applied to structural optimization of pristine and oxygen-defective surfaces as well as lattice oxygen diffusion pathways on defective surfaces. The predicted pathways and energy barriers are in good agreement with density functional theory results, particularly for pathways involving relatively large structural and energetic changes. Furthermore, the merged MLFF was applied to ideal CeO2(111) slabs containing up to 5183 atoms, yielding only weak size dependence of the predicted vacancy-migration barrier. These results highlight that systematic dataset merging across model sizes and configurations provides a practical route to constructing transferable MLFFs for complex oxide surfaces.