Development of Neural Network System for Prediction of Catalytic Performance

本研究旨在探讨神经网络系统预测最佳催化剂的可能性，从各种实验结果中提取催化剂设计所需的知识，并获得以下结果。以19种促进氧化锡催化剂上乙苯氧化脱氢反应为例，应用神经网络对催化活性和选择性进行了预测。五种产物的预测活性和选择性与实验结果吻合较好，实验误差在合理范围内。以一系列镧系氧化物上烷烃的氧化为例，考察了外推预测的可能性，其中催化性能已知是镧系离子第四电离势的函数。在丁烷氧化反应中，神经网络能很好地预测出关联度两端以及中间的活性。神经网络预测了甲烷氧化过程中C_2烃、CO和CO_2活性和选择性的变化，甚至火山类型的变化。这些结果表明，神经网络既可以用于催化剂优化的内插预测，也可以用于催化剂改进的外推预测。在预测丁烷氧化过程中镧系氧化物的催化活性时，发现第四电离势是一个关键的控制因素，只有当输入数据中包含电离势时，预测精度才会高。这些结果表明，输入数据的留一检验对于控制因素的选择是有效的：每个输入数据的重要性可以通过在训练集中留下一个输入数据来估计。在此基础上，对神经网络预测控制催化性能的关键因素和利用神经网络开发催化剂进行了评述。(3)选择性还原氮氧化物的催化实验以新型催化反应为例，采用离子交换沸石催化剂和氧化物催化剂对丙烯和甲烷还原氮氧化物进行了实验研究，并对神经网络系统在氮氧化物还原催化剂设计中的应用进行了初步试验。少

The present research aims at examining the possibility of neural network system to predict optimum catalyst extracting knowledge required for catalyst design from various experimental results, and the following resuts were obtained.(1) Prediction of Catalytic Performance by Neural NetworkNeural network was applied for the prediction of catalytic activity and selectivity by taking as an example the oxidative dehydrogenation of ethylbenzene on 19 promoted tin oxide catalysts. The predicted activity and selectivities of five products were in good agreement with those measured experimentally within reasonable experimental error.The possibility of extrapolative prediction was examined by taking as an example the oxidation of alkanes on a series of lanthanide oxides, in which catalytic performance is known to be a function of fourth ionization potential of lanthanide ions. In butane oxidation, where a monotonous correlation had been empirically established between the catalytic activity and … More the fourth ionization potential, the neural network well predicted the activities in both ends of the correlation as well as those in between. The neural network predicted, in methane oxidation, even volcano type changes of the activity and the selectivities of C_2 hydrocarbons, CO and CO_2. These results indicate that the neural network can be applied to both of the interpolative prediction for optimization of catalysts and the extrapolative prediction, at least for improvement of catalysts.(2) Estimation of Factors Controlling Catalytic Performance by Neural NetworkIn the prediction of catalytic activities of lanthanide oxides in butane oxidation, it was found that the fourth ionization potential is a key controlling factor, because the prediction accuracy was high only when input data include ionization potential. These results suggest that a leave-one-out test of input data would be effective for the selection of controlling factors : The importance of each input data could be estimated by leaving one of the input data out of the training set.On the basis of these results, some remarks were given on the estimation of the key factor controlling catalytic performance by neural network and on the catalyst development by using neural network.(3) Catalytic Experiments in Selective Reduction of Nitrogen OxideAs an example of novel catalytic reaction, the reduction of nitrogen oxide with propylene and methane was conducted by using ion-exchanged zeolite catalysts and oxide catalysts, and the preliminary test was conducted to apply the neural network system for the catalyst design of reduction of nitrogen oxide. Less