(219e) Improving Microkinetic Models of Catalytic Reactions Using Machine Learning

Microkinetic modeling is a common tool in computational catalysis to predict or infer mechanistic details of catalytic systems. The state-of-the-art in the field is to develop ab initio microkinetic models where the kinetics and thermochemistry information is obtained from density functional theory (DFT) calculations. Several assumptions and approximations are often employed in this process, however, these lead to large predictive inaccuracies – (i) intrinsic errors from DFT propagate to errors in kinetic/thermodynamic parameters that are on the order of a factor of 10 – 10³; (ii) entropies are often calculated assuming the quantum harmonic oscillator approximation which results in significant errors in the calculation of free energies for intermediates with weak vibrational modes that are neither as free as translational/rotational modes nor as restricted as those corresponding to strong covalent bonds; (iii) rates are calculated using the mass-action kinetics and mean field concentrations of species on the surface.

In this talk, we show that machine learning techniques can be employed to address each of these three issues systematically to rigorously improve the predictive accuracy of microkinetic models while being computationally tractable. We will show that interesting challenges emerge here in the context of machine learning; in particular, data-driven models have to be built using “sparse” and “heterogeneous” data as opposed to “big” and “homogeneous”.

Examples from transition metal catalysis will be used to discuss each of the proposed improvements.