(387a) A Combined Graph Theory and Machine Learning Based Method for Estimating Complex Adsorbate Configurations on Model Catalytic Surfaces

Heterogeneous catalysts form a crucial component of various vital processes, ranging from ammonia synthesis to cleanup of exhaust gases, and therefore are an important part of our society. To gain an understanding of the atomic scale phenomenon for these catalytic systems, ab-initio Density Functional Theory (DFT) is a widely used tool, and it has been successfully used to model many catalytic systems of varying degrees of complexity. Recently, growing computational power has begun to enable the extension of DFT analyses to understand intricate reaction networks involving high adsorbate coverages, multidentate adsorbates, complex catalyst morphologies, or combinations thereof. Although promising results have emerged, the vast combinatorial space implies that large numbers of explicit simulations are required to treat such systems, and these studies can therefore benefit from a more algorithmic and machine learning based data driven approaches.

To systematically account for the large and complex sample spaces described above, herein we present a generalized Python-based graph theory approach, combined with a convolution graph neural network-based machine learning workflow. The approach converts atomic scale models into undirected graph representations, which can then be used to (i) enumerate the large phase space of configurations, and (ii) train machine learning based surrogate model to the target property of choice. We showcase the utility of such an approach by systematically constructing high coverage phase diagrams for two cases: (i) NO* adsorbed on low-symmetry Pt-Sn alloy surfaces, and (ii) OH* adsorbed on non-close packed Pt surfaces. These cases encompass a large phase of 3500 and 11500 unique configurations each and our approach enables the estimation of most relevant high coverage configurations by performing DFT simulations on less than 10% of the total configurations in both the cases.