(699c) Insights from Machine Learning on a Large Database of Adsorption Energies

Adsorption energies of important intermediates are often excellent descriptors for catalytic performance. This has spurred significant interest in efficient predictions of adsorption energies, using both physical and data scientific approaches. However, larger data sets are needed in order to ensure the accuracy and generality of the approach. Here, we use a database with over 5,000 adsorption energies of H on stepped Ag alloy surfaces to develop an effective machine learning model to predict adsorption energies. We use an iterative process where we begin with a simple model, examine the errors, and then improve the model based on the errors. This brings physical insight into adsorption along with an accurate model. In particular, we discover counterintuitive trends in how adsorption energies depend on the H atom's first and second neighbor shells. The final model, an extra forest based on only 25 features, provides a high accuracy with a root-mean-square-error of 0.05 eV.