物理化学学报 >> 2010, Vol. 26 >> Issue (01): 188-192.doi: 10.3866/PKU.WHXB20100116

量子化学及计算化学 上一篇    下一篇

结合神经网络方法和扩大训练基组构建新B3LYP泛函

张家虎, 王秀军   

  1. 华南理工大学化学与化工学院应用化学系, 广州 510640
  • 收稿日期:2009-08-04 修回日期:2009-10-17 发布日期:2009-12-29
  • 通讯作者: 王秀军 E-mail:xjwangcn@scut.edu.cn

Neural Network Approach for a New B3LYP Functional with an Enlarged Training Set

ZHANG Jia-Hu, WANG Xiu-Jun   

  1. Department of Applied Chemistry, College of Chemistry and Chemical Engineering, South China University of Technology, Guangzhou 510640, P. R. China
  • Received:2009-08-04 Revised:2009-10-17 Published:2009-12-29
  • Contact: WANG Xiu-Jun E-mail:xjwangcn@scut.edu.cn

PDF (PC)

1693

赞

摘要：

神经网络方法成功地应用于修正密度泛函理论B3LYP方法中的三个参数(a0、ax和ac)以构建新B3LYP交换相关泛函. 本文采用包含输入层、隐藏层和输出层的三层式神经网络结构. 总电子数、多重度、偶极矩、动能、四极矩和零点能被选为物理描述符. 296个能量数据被随机地分成两组, 246个能量数据作为训练集以确定神经网络的最优结构和最优突触权重, 50个能量数据作为测试集以测试神经网络的预测能力. 修正后的三个参数a0、ax、ac从输出层处得到, 并用于计算体系的热化学性质如原子化能(AE)、电离势(IP)、质子亲合能(PA)、总原子能(TAE)和标准生成热(△fHΘ). 修正后的计算结果优于传统B3LYP/6-311+G(3df,2p)方法的计算结果. 经过神经网络修正后, 296个物种的总体均方根偏差从41.0 kJ·mol-1减少到14.2 kJ·mol-1.

关键词: B3LYP泛函, 神经网络, 描述符, 训练基组

Abstract:

A neural network approach was used to correct three parameters (a0, ax, and ac) in the B3LYP method for constructing a new B3LYP exchange correlation functional. A three-layer architecture which consisted of an input layer, a hidden layer, and an output layer, was adopted in the neural network. The total number of electrons, spin multiplicity, dipole moment, kinetic energy, quadrupole moment, and zero point energy were chosen as the most important physical descriptors. In this work, 296 energy values were randomly divided into two subsets: 246 energy values were used as the training set to determine the optimized structure of the neural network and the optimized synaptic weights; 50 energy values were used as a testing set to test the prediction capability of our neural network. Three modified parameters a0, ax, and ac that were obtained from the output layer were used to calculate thermochemical data such as the atomic energy (AE), ionization potential (IP), proton affinity (PA), total atomic energy (TAE), and standard heat of formation (△fHΘ). The newresults obtained, based on the neural network approach, are much better than the results calculated using the conventional B3LYP/6-311+G(3df,2p) method. Upon the neural network correction, the overall root-mean-square (RMS) error for the 296 species decreased from41.0 to 14.2 kJ·mol-1.

Key words: B3LYP functional, Neural network, Descriptor, Training set

MSC2000: 

  • O641