催化剂定量构效关系建模方法应用进展

Application progress in QSAR modeling methods for catalysts

从近十多年的应用进展情况来看,常用的催化剂定量构效关系(QSAR)建模方法包括以多元线性回归(MLR)、主成分回归(PCR)和偏最小二乘法(PLS)为主的线性方法和以基于反向传播算法(BP)和径向基函数神经网络(RBFNN)为主的人工神经网络(ANN)非线性方法两种。最有效的线性方法是PLS法,其优点是模型机理明确,缺点是有时不如RBFNN模型的预测能力强;最有效的非线性方法是RBFNN法,其优点是模型的预测能力往往比PLS模型强,但缺点是机理不够明确。最成功的也是最具应用前景的方法是综合采用PLS法和RBFNN法同时建立某一具体催化剂的PLS线性模型和RBFNN非线性模型。用PLS模型指导新型高效催化剂的结构设计,而用RBFNN模型来预测所设计催化剂的性能,反过来修正所设计催化剂的结构,从而减少催化剂合成实验的工作量。

From the application progress in QSAR modeling methods for catalysts during the recent decade, it is noticed that the methods often used include linear ones, such as multiple linear regression（MLR）, principle component regression（PCR） and partial least square（PLS）, and non-linear ones, such as artificial neural networks（ANNs） based on back propogation（BP） or radial basis function neural network（RBFNN）. The most efficient linear method is PLS with transparent mechanism but sometimes poorer prediction ability than RBFNN, whereas the most efficient non-linear method is RBFNN with usually better prediction ability than PLS but opaque mechanism. The most successful and prospective strategy should be starting with a linear method such as PLS to build a linear moldel and then trying to build a non-linear moldel with a non-linear method such as RBFNN to check whether the pridiction ability is improved. While the PLS model is suitable for directing the design of new catalysts, the RBFNN one is for predicting the performance of the designed catalysts beforesynthesis experiments so as to correct the design and greatly reduce the amount of synthesis work.