(376i) Statistical Learning Approaches for Identifying Descriptors of Metal/Support Interactions in Catalysis

In this work, we apply statistical learning (SL) methods to identify physical descriptors for predicting changes in metal/oxide interactions that occur in heterogeneous catalysis. The binding energy of transition metal atoms on oxide surfaces can be influenced by the presence of adsorbates or surface impurities that alter the electronic composition of the oxide support. We use SL approaches to predict the extent to which these surface-modifiers alter metal/support interactions based on readily available physical properties of the supported metal and the oxide. Least absolute shrinkage and selection operator (LASSO) is a well-established tool for feature selection in catalysis and material science. However, a persistent flaw of LASSO is its high rate of false positives, where features are selected for the model that are not true independent descriptors of the data in the training set. To combat this issue, we introduce two additional Bayesian methods, Horseshoe priors and Dirichlet-Laplace priors, in our feature selection process. Unlike LASSO regression, these two Bayesian methods estimate a distribution of feature coefficients and the corresponding error. These distributions then are updated using iterative Markov chain Monte Carlo (MCMC) sampling until the distribution reaches an equilibrium prediction. The descriptors with coefficients greater than a pre-defined cutoff threshold then are selected as the best descriptors for the final model. Well-selected priors in these Bayesian methods reduce false positives, thus leading to more robust descriptor sets for general models of metal/support interactions in catalysis.