Automated learning of chemical reaction networks

Abstract

The identification of kinetic models from process measurements relies on obtaining expensive data from experimental designs in order to estimate kinetic rate parameters and discern which mechanisms fit the data most appropriately. This paper proposes an approach that utilizes a mixed-integer nonlinear programming formulation to estimate and identify kinetic rate parameters from a postulated superset of reactions. The proposed methodology can be applied to steady state systems, or dynamic systems through explicit derivative estimation. An adaptive design of experiments methodology is proposed to facilitate intelligent sampling across the experimental domain. The proposed methodology is applied to three problem settings: isothermal CSTRs where initial concentrations are varied to learn the reaction network, dynamic chemical looping combustion reactors where semi-empirical mechanistic models are used to model catalyst conversion, and an expanded example that considers alternative reaction mechanisms and stoichiometric relationships and applies the proposed adaptive sampling technique to a higher dimensional domain.