When performing computational studies of catalysts, often many calculations are performed to test various sites, reaction pathways, alloys, etc. In many cases, only a few of these calculations are needed in the end: the most stable sites, the most kinetically relevant pathways, or the best-performing alloys. Hence, efficient estimates could dramatically reduce the number of calculations needed, and may also allow more intensive computational methods to be used for this smaller number of calculations. However, this scheme requires a pre-trained, general model that can predict the energetics of many reaction intermediates in different sites on many different alloy surfaces. In this work, we develop and apply such a model.<br/><br/>To develop a general model, we used a new machine learning architecture made up of sub-models. This framework partially decouples the adsorbed species from the surface by introducing latent variables, which are a set of surface properties that control how the surface interacts with all adsorbates. These latent variables can be explicitly defined or can be learned implicitly. We create separate sub-models for predicting these latent variables for each surface element, and separate sub-models for how different adsorbates respond to these latent variables. Our framework takes advantage of the fact that elements are discrete entities, and greatly simplifies the huge combinatorial challenge of considering many possible adsorbed species and many possible alloy surfaces. In the end, our method allows efficient prediction of the energetics of entire reaction pathways on alloy surfaces, and can be reused in new contexts (i.e., for a new reaction). We apply this framework to multiple reactions and show how it can significantly improve the efficiency of computational catalysis studies. Further, this work shows the tradeoffs between accuracy, complexity, data efficiency, and generality in machine learning models.

Reducing computational effort in catalysis with a general machine learning model

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.