Adaptive designs for the exploration of reaction kinetic phase transitions

Abstract

In heterogeneous catalysis, changes of the reaction mechanism or of the catalyst’s surface structure and composition often lead to drastic changes in the effective kinetics over a small range of reaction conditions. The topology of kinetic phase diagrams that summarize this effective kinetics is therefore typically characterized by extended regimes of smooth kinetic behavior separated by narrow such phase transitions. This discontinuous topology prevents the efficient exploration of kinetic phase diagrams with traditional design-of-experiment approaches that assume a smoothly varying measurement function over the entire investigated domain. Bridging towards modern statistical learning and optimization, we here propose an adaptive design algorithm that tackles this issue in a data-efficient way. A support vector machine classification learns and iteratively refines the a priori unknown positions of the phase transitions by optimally designing new data points, i.e. the reaction conditions at which the next kinetic measurements are to be taken. Using a variety of analytic toy systems, we illustrate the potential of this approach in analyzing experimental design spaces composed of multiple distinct regimes. Finally, we reconstruct the kinetic phase diagram for the CO oxidation over a RuO${}_2$(110) model catalyst from a minimum number of simulated kinetic data from a microkinetic model.