CAJ | 학술논문

以86个离子化合物的正、负离子的有效核电荷Z~(*+)、Z~(*-)、离子半径r~+、r_-以及正离子的荷径比Z~(*+)/r~+5种结构参数作为自变量,以晶格能U作为因变量,采用BP神经网络建立关于无机离子晶体晶格能的结构-性质关系(QSPR)模型.该模型由输入层、隐含层和输出层构成3层BP神经网络,86个离子化合物样本则按文献分别划分为训练集和验证集.研究表明,当隐含层神经元个数为5时模犁效果最佳:该模型对训练集拟合结果的决定系数R~2=0.996 5,平均相对误差MRE=1.63%:对验证集预测结果的R~2=0.995 2,MRE=1.85%.
이86개리자화합물적정、부리자적유효핵전하Z~(*+)、Z~(*-)、리자반경r~+、r_-이급정리자적하경비Z~(*+)/r~+5충결구삼수작위자변량,이정격능U작위인변량,채용BP신경망락건립관우무궤리자정체정격능적결구-성질관계(QSPR)모형.해모형유수입층、은함층화수출층구성3층BP신경망락,86개리자화합물양본칙안문헌분별화분위훈련집화험증집.연구표명,당은함층신경원개수위5시모리효과최가:해모형대훈련집의합결과적결정계수R~2=0.996 5,평균상대오차MRE=1.63%:대험증집예측결과적R~2=0.995 2,MRE=1.85%.
QSPR model of the lattice energy of mineral crystal was built by BPNN consisted of three layers, with the descriptors of Z~(*+) (effective nuclear charges on cations), Z~(*-) (effective nuclear charges on anions), r~+ (Goldshmidt radius of cations), r_- (Goldshmidt radius of anions) and Z~(*+)/r~+.86 samples of the mineral crystal were divided into training and test set as the literature. The coefficient of determination R~2 for the training set is 0.996 5 and for the test set is 0.995 2, which can be considered very satisfactory. In addition, the mean relative error MRE was within 1.63% and 1.85% for the training set and the test set respectively. The study showed that the descriptors in the model were of clear physical and chemical meaning and the predicting results were superior to those in literature evidently.