2020
DOI: 10.1021/acs.jpca.9b08723
|View full text |Cite
|
Sign up to set email alerts
|

Performance and Cost Assessment of Machine Learning Interatomic Potentials

Abstract: Machine learning of the quantitative relationship between local environment descriptors and the potential energy surface of a system of atoms has emerged as a new frontier in the development of interatomic potentials (IAPs). Here, we present a comprehensive evaluation of ML-IAPs based on four local environment descriptors -Behler-Parrinello symmetry functions, smooth overlap of atomic positions (SOAP), the Spectral Neighbor Analysis Potential (SNAP) bispectrum components, and moment tensors -using a diverse da… Show more

Help me understand this report
View preprint versions

Search citation statements

Order By: Relevance

Paper Sections

Select...
2

Citation Types

8
486
0

Year Published

2020
2020
2024
2024

Publication Types

Select...
8
1

Relationship

2
7

Authors

Journals

citations
Cited by 645 publications
(494 citation statements)
references
References 106 publications
8
486
0
Order By: Relevance
“…The ML interatomic potentials are regression models of the descriptors. Subsequently, many recent applications of ML interatomic potentials have achieved simulation lengths and timescales accessible to classical interatomic potentials, with near quantum mechanical accuracy 17,18 . Despite the progress, training the ML interatomic potentials remains a challenging task.…”
Section: Introductionmentioning
confidence: 99%
“…The ML interatomic potentials are regression models of the descriptors. Subsequently, many recent applications of ML interatomic potentials have achieved simulation lengths and timescales accessible to classical interatomic potentials, with near quantum mechanical accuracy 17,18 . Despite the progress, training the ML interatomic potentials remains a challenging task.…”
Section: Introductionmentioning
confidence: 99%
“…The ability to augment physics-constrained features with neural networks is causing a revolution in multi-scale modeling in various fields. 18,19 We believe that the ideas behind the universal battery modeling approach presented here will have a similar impact in modeling across length and time scales, capturing microscopic and macroscopic phenomena at unprecedented detail.…”
mentioning
confidence: 91%
“…1a. 16 The ability to break this kind of multi-scale modeling trade-off by physics-informed neural networks is now widely realized in atomistic simulations 18 and fluid mechanics. 19 The key insight that enables breaking this trade-off is to encode physical principles for data efficiency and extrapolation in conjunction with the representation power of neural networks.…”
mentioning
confidence: 99%
“…Four classes of methods are usually considered: (1) Gaussian approximation potentials [36,37], (2) Kernel Ridge Regression [38][39][40], (3) Sparse Linear Regression [41][42][43][44] and (4) Artifi-cial Neural Network (ANN) [45][46][47]. Recent comparisons between these different methods have also been carried out [48,49].…”
Section: Introductionmentioning
confidence: 99%