(521dy) Developing Predictive Adsorption Energy Descriptors for Unary and Binary Transition Metals By Tailoring Feature Sets for Machine Learning

Computational techniques such as first-principles calculations including density functional theory, and statistical learning are used as tools to screen through large numbers of potential catalysts to find viable candidates for numerous chemical reactions, as well as descriptors to measure the viability of a catalyst candidate. However, many machine learning tools operate as a black box and the identified results are difficult to interpret through a physical lens. We attempt to elucidate some of the mysteries of these black-box methods to find physically interpretable descriptors for reaction intermediate adsorption energies, using unary and binary metals as a model system, and compare these statistically generated descriptors to existing analytical models. Specifically, we use the sure independence screening and sparsifying operator (SISSO) method which takes in a matrix of material properties and combines them with each other using operators to form and identify physically interpretable descriptors.

By manipulating feature choices considered in the training of the SISSO model, we show that the resulting descriptors and error trends can inform us about the types of material properties that contribute to catalytic activity. Figure 1 shows the scheme used to identify trends in descriptor types by using various types of feature sets. For instance, features can be split into categories of tabulated, geometric, and electronic features, and the error trends of models trained on these individual feature sets can pinpoint us in the right direction when looking for adsorption energy descriptors. We conclude that site-specific, electronic descriptors contribute heavily to the adsorption energy, regardless of which reaction intermediate is considered. This compares well to what is known from analytical models for these metal systems. To expand upon this work, other materials are considered, and stability effects are also targeted by expanding the dataset to include information about both activity and stability of different material classes.