Summary

Computational chemistry is the key to rationalizing experimental procedures for producing in a more sustainable way. Our best tool for predicting and understanding chemical processes is density functional theory (DFT). However, this method is expensive in computational time terms. This issue could be improved using statistical learning techniques, like the artificial neural network (ANN) method as it is shown in this study. This method has been applied in two different systems: metal oxides oxygen diffusion and direct synthesis of hydrogen peroxide from oxygen and hydrogen on palladium (Pd) and palladium-gold (PdAu) alloy surfaces. Our ANN1 code was trained for predicting activation energy values. Those values were the input for our Kinetic Monte Carlo (KMC) machine, which whom we have tested the accuracy of our predictions calculating chemical parameters as the diffusion constant for cerium oxide doped with gadolinium (CEO₂) at different temperatures and the selectivity for pd and pdau.

The remarkable accuracy of our predictions allows us to asses the usefulness of statistical learning in theoretical and computational chemistry and give some insights into its limitations.