ML Potential for Phonon and Phase-Transition Prediction Shows Performance Differences Across Fine-Tuning Strategies

1/ AI 활용 방법: Material-specific fine-tuning (including transfer learning, multihead fine-tuning, LoRA, L²-SP)
2/ 적용 과제: Prediction of phonon band structure, thermal and elastic properties, and the potential-energy surface of unstable phonon modes for 53 materials
3/ 보고된 결과: Performance improved with only a small number of material-specific training structures, and some strategies improved thermal and elastic properties while reducing phonon error
4/ This paper applied machine learning interatomic potential to 53 materials to predict phonon band structure, thermal and elastic properties, and the potential-energy surface of unstable phonon modes, and systematically compared fine-tuning strategies such as targeted fine-tuning, transfer learning, multihead fine-tuning, LoRA, and L²-SP.
5/ According to the reported results, performance could improve with only about 10 material-specific training structures.
6/ The direct significance for materials research is that it practically shows which fine-tuning methods are advantageous for phonon and phase-stability-related predictions when a general-purpose MLIP is adapted to a new material family with a small amount of material-specific data.
7/ In particular, some strategies reduced phonon error while also improving thermal and elastic properties, and they showed differences in dynamic instability classification and reproduction of DFT benchmarks.
8/ There are also limitations.
9/ Even with similar phonon accuracy, dynamic instability and phase-transition prediction can differ greatly, and the reported performance depends on the specific dataset and evaluation setup.
10/ Therefore, rather than generalizing these results directly to other material systems, independent validation and inspection of the extrapolation range are needed in one's own domain. [1]