Machine Learning Interatomic Potentials

Introduction

For over decades, machine learning based interatomic potential () has been develop to fast the calculations of complex system in a high accuracy (Density Functional Theory level). To descript the atomic environment (atomic structures), several different descriptors have been developed. These descriptors are also connected to different machine learning methods. Although it is very hard to distinguish which method is better on predicting the properties of materials, the successful applications of

Descriptors

• Atomic Centered Symmetry Functions (ACSF)

ACSF was firtly introduced by Jörg Behler and Michele Parrinello in 2007 [1]. The main idea is to represent the total energy E of the system as a sum of atomic contributions Ei. In this formular, they define a cutoff function by a cosin function. such as

• Smooth Overlap Approximation Potential (SOAP)

• Moment Tensor Potential (MTP)

Based on the MTP method, our group develped a method to effectively sample the diffferent structure for the machine learning [3]. The detailed method and the could be found in the reference [4] and github [3], anyone is interested in our work could refer these.

Machine Learning

• Gaussian

• Neural Network

Codes

• DAS
• DeepMD
• n2p2: ACSF–NNP
• LASP: ACSF–NNP
• MTP: MTP
• GAP

References

[1] Jörg Behler and Michele Parrinello, Phys. Rev. Lett. 98, 146401 (2007)
[2] Jörg Behler, J. Chem. Phys. 134, 074106 (2011)