Universal Interatomic Potentials as Configuration-Space Generators for One-Shot and Iterative Fine-Tuning of Ab Initio-Accurate Material-Specific Models.

Universal machine learning interatomic potentials (MLIPs) are rapidly becoming general-purpose tools for atomistic simulation, but their role in quantitative materials modeling when reactive events are involved remains unsettled. We compare five universal MLIPs across seven chemically diverse systems and find that strong performance on standard benchmarks does not guarantee accurate predictions of the target observables. In particular, zero-shot models do not reliably reproduce reactive, transport, or high-barrier processes, exemplified here in particular by the sulfur-vacancy jump in MoS2. We therefore benchmark a practical alternative against target observables: universal MLIPs are used to generate long molecular dynamics trajectories, the resulting configurations are subsampled and relabeled with DFT, and material-specific MLIPs are subsequently trained or fine-tuned on the resulting first-principles data sets. This workflow converts universal models into efficient configuration-space generators while retaining ab initio reference labels for training and turns their systematic softening of the potential energy surface from a liability into a sampling advantage. Across the tested systems, 2000 DFT-recalculated structures are often sufficient to obtain accurate fine-tuned or trained-from-scratch models. For the most challenging case, iterative self-training progressively refines the sampled configuration space and recovers the DFT MoS2 potential energy profile with only 600 first-principles calculations in total. For a 512-atom system on a single nvidia A100 GPU and four 32-core CPU nodes, the complete workflow─including training data generation, DFT labeling, and model creation─delivers 1,000,000 ab initio-quality MD steps in 3 days.