Application of density functional theory and machine learning in the prediction of efficient catalysts for the oxidative coupling of methane with reduced CO2 production at low temperature

Abstract

The oxidative coupling of methane (OCM) remains a promising means for the production of ethylene. Though high temperatures of 900 oC and above lead to increased conversion of methane, at higher temperatures, the OCM reaction tends to favour a competing reaction that leads to the production of COx product. This thesis focuses on the generation of catalyst component electronic properties via density functional theory (DFT) and the analysis via machine learning (ML) techniques of the combination of the DFT-computed electronic properties and high-throughput experimental data comprising 12708 data points. The DFT data includes the catalyst components' bandgap, Fermi energy and magnetic moment. Variations of a dataset comprising experimental conditions, experiment performance and electronic catalyst properties were used to build a dataset for data modelling and analysis as well as ML analysis for the prediction of CO2 yield, C2H4 yield C2H6 yield, and CH4 conversion. With the aid of multi-linear regression models, Rh, Pt, Ru and Ir were found to be new catalyst promoters that enhance catalyst performance in OCM, particularly with improved methane conversion. A minimum of 58 new bimetallic combinations and 2784 unique catalytic materials with minimum CH4 conversion at 700 oC of 38.5% were identified and proposed as effective catalysts for OCM reaction, an improvement on the 36% CH4 conversion limit from previous studies. The Rh-Li3VO4/SBA15 and Ru-BaZrO3/SBA15 have been proposed as efficient catalysts for the OCM reaction with a predicted C2y of 30% and 29%, respectively. In comparing predictive model using a dataset containing a combination of catalyst electronic properties, deep neural networks (DNN) configured as deep feed-forward networks with back-propagation, along with random forest regression (RFR), support vector regression (SVR) and extreme gradient boost regression (XGBR), were compared on the basis on their mean-absolute-error, mean-squared-error and coefficient of determination for the prediction of reaction outcomes including ethane yield, ethylene yield, carbon dioxide yield and methane conversion (C2H4y, C2H6y, CO2y, CH4_conv, respectively) and C2y (a combination of C2H4y and C2H6y). The inclusion of electronic properties of the catalyst components into the dataset improved the performance of the models by approximately 10% compared to a dataset with only reaction conditions. RFR models had better accuracy compared to other modeling techniques, with an average R2 of 0.98 for the predictive models of all five reaction outcomes. The mean squared error and mean absolute error of the RFR models were from 0.12 to 9.03 for MSE and 0.21 to 2.02, respectively. The order of performance of the modeling techniques was RFR > XGBR > SVR > DNN. The order of data fit for the labels for the given modeling techniques was C2H6y > C2H4y > C2y > CO2y > CH4_conv. In the analysis of model feature impacts to identify descriptors for catalytic activity in OCM reactions, it was observed that the Fermi energies of the catalyst promoter, its atomic number and the bandgaps of the bimetallic oxide and the catalyst support emerged as effective descriptors. Specifically, in relation to the C2y predictive model, C2y increases with an increase in dataset features, including the number of moles of the alkali/alkali-earth in the metallic oxide, atomic number of the catalyst promoter and Fermi energy of the promoter and just relatively in the case of temperature, suggesting a highly non-linear relationship between C2y and temperature. It, however, reduces with an increase in the bandgap of the active metal oxide and the methane-to-oxygen ratio. Using the RFR, the Fermi energy of the promoter had a 4.31% impact on the model, while its atomic number had 6.24%, the number of moles of the alkali/alkali-earth in the metallic oxide was 13.69%, and temperature was 33.70% on the C2y predictive model. catalysts with active metal oxides with lower bandgap energy and promoters with magnetic moments may not be as effective as OCM reaction catalysts with less ferromagnetic properties and higher bandgap energy.