Improving the predictive power of microkinetic models via machine learning

Microkinetic modeling is commonly used in heterogeneous catalysis to study reaction mechanisms and compute information such as the reaction rates, selectivity, degrees of rate control, surface coverages, and so on, under reaction conditions. Typically, these models are formulated by invoking many approximations that ultimately lower their quantitative accuracy. Here, we discuss how some of the approximations commonly employed can be improved by using machine-learning techniques. In particular, we discuss (i) improved estimation of enthalpy and entropy of adsorbates and transition states, (ii) overcoming the mean-field assumption at the fast-diffusion limit, and (iii) quantification of prediction uncertainties to improve the predictive accuracy of microkinetic models and assess their reliability. We finally present our outlook on how machine learning holds great promise in modeling complex catalytic systems and specifically the need for advances in building data-driven models when the underlying data are imperfect, that is, they are sparse, collated from disparate sources, and are of differing accuracies.