![[2504.19372] Composable and adaptive design of machine learning interatomic potentials guided by Fisher-information analysis](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2504.19372] Composable and adaptive design of machine learning interatomic potentials guided by Fisher-information analysis
Summary
This article presents a novel adaptive design strategy for machine learning interatomic potentials (MLIPs), leveraging Fisher-information analysis to optimize model configurations and performance metrics.
Why It Matters
The research addresses the need for flexible and efficient modeling in materials science, particularly in the context of machine learning applications. By combining model reconfiguration and evaluation, it enhances the accuracy and adaptability of MLIPs, which are crucial for simulating material behaviors at the atomic level. This could lead to significant advancements in computational materials science and engineering.
Key Takeaways
- Introduces an adaptive design strategy for MLIPs using Fisher-information analysis.
- Combines model reconfiguration with a unified training procedure for improved performance.
- Achieved a low force RMSE of 0.172 eV/Å and energy RMSE of 0.013 eV/atom in a niobium dataset case study.
- Balances flexibility and extensibility in model design, enhancing usability in various applications.
- Provides a framework that could streamline the development of interatomic potentials in materials science.
Condensed Matter > Materials Science arXiv:2504.19372 (cond-mat) [Submitted on 27 Apr 2025 (v1), last revised 26 Feb 2026 (this version, v2)] Title:Composable and adaptive design of machine learning interatomic potentials guided by Fisher-information analysis Authors:Weishi Wang, Mark K. Transtrum, Vincenzo Lordi, Vasily V. Bulatov, Amit Samanta View a PDF of the paper titled Composable and adaptive design of machine learning interatomic potentials guided by Fisher-information analysis, by Weishi Wang and 4 other authors View PDF HTML (experimental) Abstract:An adaptive physics-inspired model design strategy for machine-learning interatomic potentials (MLIPs) is proposed. This strategy relies on iterative reconfigurations of composite models from single-term models, followed by a unified training procedure. A model evaluation method based on the Fisher information matrix (FIM) and multiple-property error metrics is also proposed to guide the model reconfiguration and hyperparameter optimization. By combining the reconfiguration and the evaluation subroutines, we provide an adaptive MLIP design strategy that balances flexibility and extensibility. In a case study of designing models against a structurally diverse niobium dataset, we managed to obtain an optimal model configuration with 75 parameters generated by our framework that achieved a force RMSE of 0.172 eV/Å and an energy RMSE of 0.013 eV/atom. Comments: Subjects: Materials Science (cond-mat.mtrl-sci); Machine Learn...