(538h) Investigation of Metal Adsorption Trends on Modified MgO Supports Using Density Functional Theory and Statistical Learning

Metal/oxide interactions are known to influence reactivity and stability in numerous heterogeneous catalysts. In this work, we use density functional theory (DFT) and statistical learning (SL) to derive models for describing trends in metal adsorption energy on MgO(100) surfaces that are modified by adsorbates (e.g., H, OH, F and NO₂) and dopants (e.g., Li, Na, B and Al). Since pristine MgO is an irreducible and relatively inert oxide, we examine how these modifications can be used to enhance and control metal/oxide interactions. We find that most transition metals bind to MgO(100) more strongly after the surface is either modified by adsorbates or doped by impurities, as these modifications create electron-rich or electron-poor surfaces (depending on the surface-modification) that are generally more reactive. The adsorbed metal acts as a reservoir that can exchange electron density with the surface, and, intuitively, the absolute electronegativity of the adsorbed metal (i.e., the average of the metal’s first ionization potential and electron affinity) is found to correlate with the metal’s binding energy difference. However, this simple linear correlation is insufficient for fully describing and predicting the effects of differing adsorbates and dopants. As such, we apply SL to identify more general descriptors that capture simultaneous effects stemming from properties of the adsorbed metal, the MgO support, and the surface modifiers. We employ a composite SL approach to select features from a pool of more than one million candidate descriptors to predict metal binding energies on modified MgO(100) surfaces. This composite approach applies three separate SL methods (i.e., LASSO, Horseshoe priors, and Dirichlet-Laplace priors) to independently select features from the candidate feature-space. The final model is composed of only features that were identified by all three models, thus yielding a robust set of descriptors. The final model for predicting binding energies on doped MgO(100) surfaces contains three descriptors and has a root mean square error (RMSE) of 0.37 eV. For adsorbate-modified surfaces, the final model contains four descriptors and has a RMSE of 0.25 eV. In both models we find that the selected features are dominated by electronic properties, such as electronegativity, ionization energy, and electron affinity, thus underscoring the importance of charge transfer in the metal/oxide interaction.