Screening of Additives to a Co/SrCO₃ Catalyst by Artificial Neural Network for Preferential Oxidation of CO in Excess H₂

Abstract

Preferential oxidation (PROX) of CO in excess hydrogen was investigated over cobalt catalysts supported on SrCO3, which showed a high performance. From the results of X-ray diffraction, X-ray photoelectron spectra (XPS), and temperature-programmed desorption, it was concluded that the active species for the PROX of CO is not cobalt oxide but cobalt carbonate-like compound, which was formed in the catalyst preparation step from Co(NO3)2 precursor and SrCO3 support. On the basis of the multivariate analysis, characters by XPS analysis are the main factors to determine the CO conversion with Co/SrCO3 catalyst. The selectivity for CO oxidation was suppressed by the side reactions, such as H2 oxidation and reverse water-gas shift reaction. Therefore, new additive to the Co/SrCO3 catalyst for the retardation of the side reactions was investigated by using an artificial neural network (ANN). The activities of 17 mol % Co + 1.7 mol % X/SrCO3 (X = B, K, Sc, Mn, Zn, Nb, Ag, Nd, Re and Tl) and 16 physicochemical properties of those 10 elements were used as training data of ANN, and the most optimal additive among 53 elements in periodic table was predicted by ANN. From the result of the prediction and experimental verification, boron addition to the catalyst was effective to increase the activity for CO oxidation at the reaction temperature range from 200 to 240 °C. Actually, 17 mol % Co + 5.1 mol % B/SrCO3 catalyst showed 99% CO conversion with 52% selectivity at 200 °C with a feed composition of 0.69 vol % CO, 0.69 vol % O2, 4.5 vol % N2, 10 vol % H2 O, 17 vol % CO2 and H2 as a balance at a space velocity of 3 g·h/mol, and its activity was stable for 50 h.

Discovery and Optimization of Materials Using Evolutionary Approaches

“…The development of Co/SrCO 3 based catalysts for preferential oxidation of CO in excess H 2 was also reported by Kobayashi et al The catalysts consisted of 17 mol % Co + 1.7 mol % X/SrCO 3 catalyst, where X was one of 10 elements (B, K, Sc, Mn, Zn, Nb, Ag, Nd, Re, and Tl). Sixteen physicochemical properties for 10 elements and the catalytic performance (CO conversion and O 2 conversion by H 2 ) were used to train a neural network model.…”

“…The predicted optimum catalyst was prepared and tested and showed 80% selectivity for preferential oxidation at 200 °C and over 99% conversion of CO with the assistance of methanation of CO at 220 °C (Figure 40). The development of Co/SrCO 3 based catalysts for preferential oxidation of CO in excess H 2 was also reported by Kobayashi et al 87 The catalysts consisted of 17 mol % Co + 1.7 mol % X/SrCO 3 catalyst, where X was one of 10 elements (B, K, Sc, Mn, Zn, Nb, Ag, Nd, Re, and Tl). Sixteen physicochemical properties for 10 elements and the catalytic performance (CO conversion and O 2 conversion by H 2 ) were used to train a neural network model.…”

Knowledge Extraction from Catalysis of the Past: A Case of Selective CO Oxidation over Noble Metal Catalysts between 2000 and 2012

“…Similarly, Baerns and Holena [5] described the use of ANNs for the study of catalytic materials in their book in 2009. Since the mid-1990s, numerous successful applications of ANNs have been reported in the literature in the field of catalysis, [6][7][8][9][10][11][12][13][14] including some of our recent works. [15][16][17][18] In contrast, decision trees can help derive simple but valuable rules, such as finding the variables leading to high or low catalytic performance levels.…”

Hybrid Model of Industrialp-Xylene Oxidation Incorporated Fractional Kinetic Model with Intelligent Models

“…It is impossible to employ SVR method to develop a correction model due to the lack of the industrial rate constants, i.e., the lack of the complete training sample data. As one of the most popular intelligent modeling method, ANN model has been already used for the chemical process modeling successfully. − Thus, an artificial neural network (ANN) correction model is introduced to modify the rate constants of laboratory SBR to that of industrial CSTR due to the high ability of ANN to model nonlinear problems. Because of the lack of the complete training sample data, the traditional backpropagation algorithm, which is based on the input and output sample data, is not suitable for training the ANN.…”