# Benchmarking phonon anharmonicity in machine learning interatomic potentials

**Authors:** Sasaank Bandi, Chao Jiang, Chris A. Marianetti  
**arXiv:** [2402.18891](https://arxiv.org/abs/2402.18891)  
**Date:** 2024-02-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.18891v1/

## Key points

- Machine learning approaches are powerful tools for investigating structure-property relationships in crystals and molecules
- Machine Learning Interatomic Potentials (MLIP) accurately reproduce first-principles data at a similar cost to traditional interatomic potential methods
- Lack of clear understanding of anharmonic terms encoded in MLIP despite extensive testing on various materials and molecules
- Study benchmarked popular MLIP using anharmonic vibrational Hamiltonian of ThO$_2$ in fluorite crystal structure constructed from Density Functional Theory (DFT)
- Three classes of MLIP were trained: Gaussian Approximation Potentials, Artificial Neural Networks (ANN), and Graph Neural Networks (GNN)
- Results evaluated by comparing phonons and their interactions, as well as phonon linewidths, lineshifts, and thermal conductivity
- Models trained on DFT molecular dynamics dataset with good agreement up to fifth-order for both ANN and GNN models
- Potential of MLIP in accurately characterizing anharmonicity in materials systems at a fraction of the cost compared to conventional first principles-based approaches
- Specific computational methods utilizing DFT calculations detailed in the study, providing valuable insights into understanding phonon anharmonicity within materials systems
- Research contributes to advancing understanding of anharmonicity through machine learning interatomic potentials, suggesting they can effectively capture complex structural properties with high accuracy and efficiency

## Summary

Machine learning approaches have become powerful tools for investigating structure-property relationships in crystals and molecules. One promising approach is the use of Machine Learning Interatomic Potentials (MLIP), which have shown to accurately reproduce first-principles data at a similar cost to traditional interatomic potential methods. However, there is a lack of clear understanding of the anharmonic terms encoded in MLIP despite extensive testing on various materials and molecules. In a recent study by Sasaank Bandi, Chao Jiang, and Chris A. Marianetti, popular MLIP were benchmarked using the anharmonic vibrational Hamiltonian of ThO$_2$ in the fluorite crystal structure. The Hamiltonian was constructed from Density Functional Theory (DFT) using highly accurate and efficient irreducible derivative methods to generate molecular dynamics trajectories. Three classes of MLIP were trained on this dataset: Gaussian Approximation Potentials, Artificial Neural Networks (ANN), and Graph Neural Networks (GNN). Results were evaluated by comparing phonons and their interactions, as well as phonon linewidths, lineshifts, and thermal conductivity. Additionally, the models were trained on a DFT molecular dynamics dataset with good agreement up to fifth-order for both ANN and GNN models. This analysis highlights the potential of MLIP in accurately characterizing anharmonicity in materials systems at a fraction of the cost compared to conventional first principles-based approaches. Furthermore, specific computational methods utilizing DFT calculations with parameters such as plane-wave basis cutoff energy and k-point mesh densities were detailed in the study. The computation of higher-order derivatives provided valuable insights into understanding phonon anharmonicity within materials systems. In conclusion, this research significantly contributes to advancing our understanding of anharmonicity through machine learning interatomic potentials and suggests that they can effectively capture complex structural properties with high accuracy and efficiency, paving the way for further advancements in material science research.