Speakers
Description
Molecular dynamics (MD) has been popularly utilized to understand the dynamical properties of materials such as thermal, electrical, and ionic conductivities. Ab initio MD provides universal, high-quality predictions for energy, forces, and stress of any material, but its usage is limited due to high computational costs. Recent machine-learned interatomic potentials (MLIPs), with their excellent size scalability and remarkable calculation efficiency, can address this issue. However, the reliability of MLIPs could not be guaranteed for out-of-domain configurations. Particularly, rare events, such as defect creation or phase transition precursors, are often missed in training data or have regularized away due to insufficient data during MLIP training. But, owing to their significant impact on dynamical properties [1], their behavior should be reproduced by MLIP. This study systematically investigates how an active learning (AL) scheme can deal with rare event training and accelerate the whole training process [2]. First, the configurational space is examined by MD using explorative MLIPs, such as NequIP [3] and SO3KRATES [4]. Second, all generated MD snapshots are evaluated based on the MLIP prediction uncertainty, which enables qualitative identification of unfamiliar data. Finally, an iterative loop is formed by incorporating unfamiliar data into training data to retrain MLIP models for the next round. Applying AL to 122 materials [1, 5] identifies two representative corrections for erroneous MLIP predictions: a loss of real rare events and a prediction of false events. In addition, under-(over-)estimation of phonon lifetimes in AgGaSe2(CuI) shows the impact of erroneous MLIP predictions in dynamical properties, while AL rectifies them. In the end, we make NequIP enable the direct evaluation of unfolded heat flux prediction via automatic differentiation [6], resulting in efficient heat conductivity evaluation using MLIPs. Finally, a whole AL process leads to a systematic MLIP approach for thermal conductivity predictions of thermal insulators.
[1] F. Knoop, T. A. R. Purcell, M. Scheffler, and C. Carbogno, Anharmonicity in thermal insulators: An analysis from first principles. Physical Review Letters, 130(23), 236301, 2023; https://doi.org/10.1103/PhysRevLett.130.236301
[2] K. Kang, T. A. R. Purcell, C. Carbogno, and M. Scheffler, Accelerating the Training and Improving the Reliability of Machine-Learned Interatomic Potentials for Strongly Anharmonic Materials through Active Learning. arXiv preprint arXiv:2409.11808, 2024; https://doi.org/10.48550/arXiv.2409.11808
[3] S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13(1), 2453, 2022; https://doi.org/10.1038/s41467-022-29939-5
[4] J. T. Frank, O. T. Unke, K.-R. Müller, and S. Chmiela, A Euclidean transformer for fast and stable machine learned force fields. Nature Communications, 15(1), 6539, 2024; https://doi.org/10.1038/s41467-024-50620-6
[5] NOMAD repository for aiMD data, DOI: 10.17172/NOMAD/2021.11.11-1
[6] M. F. Langer, F. Knoop, C. Carbogno, M. Scheffler, and M. Rupp, Heat flux for semilocal machine-learning potentials. Physical Review B, 108(10), L100302, 2023; https://doi.org/10.1103/PhysRevB.108.L100302
This project was supported by the NOMAD Center of Excellence (European Union's Horizon 2020 research and innovation program, Grant Agreement No. 951786) and the ERC Advanced Grant TEC1p (European Research Council, Grant Agreement No. 740233).
