Machine Learning Prediction of H Adsorption Energies on Ag Alloys

Adsorption energies
on surfaces are excellent descriptors of their
chemical properties, including their catalytic performance. High-throughput
adsorption energy predictions can therefore help accelerate first-principles
catalyst design. To this end, we present over 5000 DFT calculations
of H adsorption energies on dilute Ag alloys and describe a general
machine learning approach to rapidly predict H adsorption energies
for new Ag alloy structures. We find that random forests provide accurate
predictions and that the best features are combinations of traditional
chemical and structural descriptors. Further analysis of our model
errors and the underlying forest kernel reveals unexpected finite-size
electronic structure effects: embedded dopant atoms can display counterintuitive
behavior such as nonmonotonic trends as a function of composition
and high sensitivity to dopants far from the adsorbing H atom. We
explain these behaviors with simple tight-binding Hamiltonians and d-orbital densities of states. We also use variations among
forest leaves to predict the uncertainty of predictions, which allows
us to mitigate the effects of larger errors.