Salman Khan, University of California Santa Barbara
Electronic structure calculations have greatly advanced our understanding of homogeneous catalysts and crystalline heterogeneous catalysts. However, amorphous heterogeneous catalysts (e.g. Cr/SiO2 for olefin polymerization1 and WO3/SiO2 for olefin metathesis2) remain poorly understood. The principle difficulties are i) The nature of the disorder is quenched and unknown. (ii) Each active site has a different local environment and activity. (iii) Active sites are rare, often less than ~20% depending on the catalyst and preparation method. Few (if any) studies have ever attempted to compute site-averaged kinetics because the Arrhenius dependence on variable activation energies leads to an exponential average that requires an intractable number of electronic structure calculations to. We present a new algorithm using machine learning techniques (metric learning kernel regression) and importance sampling to efficiently learn the distribution of activation energies. We demonstrate the algorithm by computing the site-averaged activity of a model amorphous catalyst with quenched disorder.
