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

Abstract

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 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 MoS$_2$. We therefore propose a practical alternative: universal MLIPs are used to generate long molecular dynamics trajectories, the resulting configurations are sub-sampled and relabeled with DFT, and material-specific MLIPs are subsequently trained or fine-tuned on the resulting first-principles datasets. This workflow converts universal models into efficient configuration-space generators while retaining ab initio reference labels for training. Across the tested systems, $2{,}000$ 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 MoS$_2$ potential energy profile with only $600$ first-principles calculations in total. The resulting workflow enables the generation of $1$ ns ab initio-quality trajectories - including training data generation and model creation - within three days.