(696e) Accelerated Phase Diagrams for Metal Carbide Catalysts Under Reaction Conditions Using a Graph-Based Approach | AIChE

(696e) Accelerated Phase Diagrams for Metal Carbide Catalysts Under Reaction Conditions Using a Graph-Based Approach


Choksi, T. - Presenter, Nanyang Technological University
Rekhi, L., NTU Singapore
Le, L. Q., Nanyang Technological University
Prabhu, A., Nanyang Technological University
Liu, W., Nanyang Technical University
Canepa, P., National University of Singapore
Abd Ghaffar, N. F. B., Nanyang Technological University
Metal carbide catalysts evolve into oxy-carbide phases during the hydrodeoxygenation of biomass, dry reforming of methane, and ethanol dehydration. Prior studies indicate that oxy-carbide formation is dictated by the chemical potentials of the reductant (e.g., CH4) and oxidizing gases (e.g., CO2). The atomic structure of such oxy-carbide phases has however not yet been determined. We present a graph-based approach that determines the equilibrium coverage of oxy-carbide surfaces as a function of the chemical potential of the environment (CH4:CO2 ratio) on (100), (111), and (110) surfaces of early transition metal carbides (VC and TiC). Co-adsorbate interactions between oxygen atoms on the surface are partitioned as pair-wise interactions. The pair-wise interactions are in turn, determined using a distance-dependent quadratic function, whose form is inspired by generalized additive models in machine learning. The functions are constructed based of three observations: (a) co-adsorbate interactions rapidly decay with distance, (b) co-adsorbate interactions are confined to within three nearest neighbours of the active site, and (c) co-adsorbate interactions have weak structure sensitivity. The quadratic functions are trained on density functional theory calculations of O* adsorbed across top, bridge, and hollow sites having coverages ranging from 1/8 to 1 monolayer. The functions yield surface energies under reaction conditions with errors of 4 meV/Å2. These errors are comparable with more sophisticated yet less interpretable neural networks. Upon training the model, we then determine surface energies of 4000+ crystal planes on low index surfaces of TiC, essentially on-the-fly. We down select the most stable surfaces under a given reaction environment and validate their energies with explicit DFT calculations. We employ this tool to determine the most stable (100), (111), and (110) surfaces of VC, as the CH4:CO2 ratio is varied. These surface energies are inputted into Wulff constructions, thus revealing how the formation of oxy-carbides alters the carbide morphology.