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

This repository contains the training data, model files, input/configuration files, and analysis scripts supporting the above study. Universal machine-learning interatomic potentials (MLIPs) are becoming general-purpose tools for atomistic simulation, but their reliability for quantitative materials modeling of reactive events 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. We propose a workflow in which universal MLIPs act as configuration-space: they run long molecular-dynamics trajectories, the resulting configurations are sub-sampled and relabeled with DFT, and material-specific MLIPs are then trained from scratch or fine-tuned on these first-principles datasets. Across the tested systems, 2,000DFT-recalculated structures are often sufficient for 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. The workflow enables generation of 1 ns ab initio-quality trajectories - including training-data generation and model creation - within three days. This deposit provides everything needed to reproduce the datasets, models, and analyses: the DFT-relabeled reference data, the trained and fine-tuned material-specific models, the training/fine-tuning and MD/evaluation scripts, and the complete iterative self-training records for the MoS₂ case. See README.md for a full description of the directory layout and file contents.