The Artificial Neural Network Method Is Applied Research In Analytical Chemistry

Artificial neural networks (ANN) are abstract mathematical models based on the research findings of Modern Neurosciences.They are composed of a large number of highly interconnected artificial neurons working in parallel to solve a specific problem. It's just a brainlike system in the simple, abstract and similar way the human brain does, not a real neuron system. Some of the properties of the neural networks are adative-learning, self-organization, good fault tolerance and nonlinear simulation, etc. So, artificial neural networks have been considered as the important methods of Chemometrics recently.we developed these methods to solve the complex problems in chemistry, such as the chemical pattern recognition, quantitative structure-property relationship study, protein secondary structure prediction and the multivariate calibration.The outlines of the main work of the thesis are:1 Some neural network models are briefly described, the properties of the neural networks are compared.2 The feedforward back propagation neural network (BPNN) has been applied to the spectrometric simultaneous analysis of multicomponent-multisystems. The results show that the BPNN method is better than the orthogonal decomposition method in the spectrometric analysis of multicomponent multisystems.3 Two typical artificial neural networks in pattern recognition: [1] the correlation model between main chemical compositions and sensory qualities of fluecured tobaccos was constructed by generalized regression neural network (GRNN). The results showed that the prediction accuracy is good. So, the GRNN method may be the potential method for the internal quality evaluation of fluecured tobaccos. [2] Based on probability statistic theory and Bayes Classification Rule, probabilistic neural network (PNN) model was proposed. PNN was applied to classify the producing areas of the fluecured tobaccos.The inputs of PNN were the main chemical compositions of fluecured tobaccos, and the outputs were the codes of the producing areas of thefluecured tobaccos--Guizhou Province, Hunan Province, Shandong Province,Henen Province. The results show that PNN can classify the producing area exactly.4 Five neural network models, back-propagation neural network(BPNN), radial basis neural network (RBFNN) , generalized regression neural network(GRNN), cascade forward backpropagation neural network(CFNN) and Elman backpropagation neural network(ELMNN), are evaluated in predicting protein secondary structures. The prediction accuracy of GRNN is better than the others. In addition, some affecting factors (the training sets and the parameters of network) are also discussed.5 Artificial neural networks in quantitative structure-property relationship (QSPR): BP-ANN is applied to predict the boiling point of the benzene derivatives. The prediction is examined by a self-consistency test and an independent-dataset. The self-consistency test and the independent-dataset test obtain good results when using the physical property parameters about structure as the inputs. This method is proved to be effective on predicting other important properties of organics. We developed a quantitative structure-property relatinship model to predict the antioxidative strength of antioxygen. The result is satisfactory.