Chem. J. Chinese Universities ›› 2021, Vol. 42 ›› Issue (7): 2146.doi: 10.7503/cjcu20210294

• Article • Previous Articles     Next Articles

Analytical Gradient Method for Fundamental Invariant Neural Networks

SHANG Chenyao1,2, ZHANG Donghui1()   

  1. 1.State Key Laboratory of Molecular Reaction Dynamics,Dalian Institute of Chemical Physics,Chinese Academy of Sciences,Dalian 116023,China
    2.University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2021-04-29 Online:2021-07-10 Published:2021-06-17
  • Contact: ZHANG Donghui E-mail:zhangdh@dicp.ac.cn
  • Supported by:
    the National Natural Science Foundation of China(21688102);the Strategic Priority Research Program of the Chinese Academy of Sciences(XDB17010200);the Liao Ning Revitalization Talents Program, China(XLYC1808024)

Abstract:

It is very important for dynamic simulation to improve the running speed of the potential energy surface. Compared with the numerical gradient calculation, which is simple but time-consuming, the direct analytical formula to solve the gradient of the potential energy surface can greatly improve the running efficiency of the potential energy surface. In this work, a method for generating analytic gradients for fundamental invariant neural networks was developed. The code to calculate the analytic gradient can be generated automatically by program. After testing a large number of data points, it was found that this method can get the correct output of the gradient of potential energy surface. By measuring the calculation time of different potential energy surfaces, it was found that the analytical gradient method can bring more than a ten-fold improvement in performance. The larger the system, the more significant the performance improvement will be.

Key words: Potential energy surface, Neural network, Reaction dynamics

CLC Number: 

TrendMD: