{"id":"arxiv:2604.01017","kind":"article","title":"Parameter-Efficient Fine-Tuning of Machine-Learning Interatomic Potentials for Phonon and Thermal Properties","authors":["Jonas Grandel","Philipp Benner","Janine George"],"author_count":3,"year":2026,"published":"2026-04-01","venue":"npj Computational Materials","cited_by":0,"doi":"10.1038/s41524-026-02338-w","arxiv_id":"2604.01017","nct_id":null,"source_url":"https://arxiv.org/abs/2604.01017","archive_url":"https://space-kitty.com/papers/arxiv:2604.01017","updated":"2026-04-10","pmid":null,"pmcid":null,"openalex_id":"W7148347846","pdf_url":"https://arxiv.org/pdf/2604.01017","abstract":"Machine-learning interatomic potentials are widely used as computationally efficient surrogates for density functional theory in atomistic simulations, enabling large-scale, long-time modeling of materials systems. We investigate how different fine-tuning strategies influence the prediction of harmonic phonon band structures, thermal properties, and the potential energy surface along imaginary phonon modes. We achieve substantial accuracy improvements with minimal additional data, with as few as 10 additional training structures already yielding significant gains. In addition to existing approaches, we introduce Equitrain, a finetuning framework that implements LoRA-based adaptation. Across 53 materials systems, we show that fine-tuned models consistently outperform both the underlying pretrained model and models trained from scratch. Equitrain achieves the best overall performance, and our results demonstrate that fine-tuning enables accurate phonon predictions.","abstract_license":"CC0-1.0 (arXiv descriptive metadata)","license":null,"is_open_access":false,"retracted":false,"peer_reviewed":null,"categories":["cond-mat.mtrl-sci","physics.comp-ph"],"details":null,"provenance":"arxiv","attribution":"Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.","topics":[{"slug":"ai-for-science","name":"AI for Science","domain_slug":"ai"}],"cited_in_dossiers":[]}