Gentle push: Fine-tuning foundation models for complex oxide surface reconstructions

Structure-prediction methods routinely sample highly distorted, far-from-equilibrium configurations. Conventional workflows based on machine learning interatomic potentials (MLIPs) therefore require the inclusion of unphysical local atomic environments in the training data to ensure stability during the search. We show that foundation MLIPs, fine-tuned on minimal datasets containing only physically reasonable structures, can serve as robust backbones for evolutionary structure prediction of complex surface reconstructions. We systematically benchmark fine-tuning strategies and identify that freezing all layers except the readouts as the most effective strategy in the low-data regime. Fine-tuning on as few as 23 structures is sufficient to perform evolutionary searches on reconstructed SrTiO 3 (110) surfaces. For BaTiO 3 (001), we further demonstrate that a lightweight active-learning approach, based on fine-tuning on as few as 20 structures, enables the identification of new low-energy (√5 × √5)R26.6° reconstructions.