Exploring model complexity in machine learned potentials for simulated properties
Name
43578_2023_Article_1152.pdf
Size
3.72 MB
Format
Adobe PDF
Checksum (MD5)
7521b51e11a73d3c7534d3564b7424ed
Author(s)
Rohskopf, A.
Goff, J.
Sema, D.
Gordiz, K.
Nguyen, N. C.
Henry, A.
Thompson, A. P.
Wood, M. A.
Date Issued
September 18, 2023
Publisher
Springer International Publishing
Citation
Rohskopf, A., Goff, J., Sema, D., Gordiz, K., Nguyen, N. C. et al. 2023. "Exploring model complexity in machine learned potentials for simulated properties."
Version
Final published version
Abstract
Abstract
Machine learning (ML) enables the development of interatomic potentials with the accuracy of first principles methods while retaining the speed and parallel efficiency of empirical potentials. While ML potentials traditionally use atom-centered descriptors as inputs, different models such as linear regression and neural networks map descriptors to atomic energies and forces. This begs the question: what is the improvement in accuracy due to model complexity irrespective of descriptors? We curate three datasets to investigate this question in terms of ab initio energy and force errors: (1) solid and liquid silicon, (2) gallium nitride, and (3) the superionic conductor Li
$$_{10}$$
10
Ge(PS
$$_{6}$$
6
)
$$_{2}$$
2
(LGPS). We further investigate how these errors affect simulated properties and verify if the improvement in fitting errors corresponds to measurable improvement in property prediction. By assessing different models, we observe correlations between fitting quantity (e.g. atomic force) error and simulated property error with respect to ab initio values.
Graphical abstract
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
Creative Commons Attribution
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1557/s43578-023-01152-0
