Model discovery of catalyst poisoning for ammonia synthesis through symbolic regression and kinetic parameters estimation

Ammonia is a fundamental component in the modern chemical industry. The industrial synthesis of ammonia mostly relies on heterogeneous catalysts to guarantee high ammonia productivity with mild operating conditions. Therefore, characterizing the performance of catalysts for ammonia synthesis is essential to optimize industrial reactors. This typically requires accurate mathematical models. In this context, high-pressure Temkin-Pyzhev equation is typically recognized as the reference analytical kinetic model for industrial ammonia synthesis. The model describes the reaction kinetics under the assumption that nitrogen adsorption on the catalyst surface is the rate determining step. Nevertheless, this model may offer simplicist characterization of catalyst behavior, specifically when the system is affected by phenomena such as mass transfer limitations and catalyst deactivation due to poisoning by oxygenated compounds. To face these issues, empirical corrections, such as the Dyson–Simon model for mass transfer and the Nielsen model for reversible poisoning, are traditionally applied. However, they exhibit some limitations. The Dyson-Simon model can be applied only in limited range of operating conditions, therefore limiting its applicability outside its domain, while the Nielsen model fails to capture the irreversible poisoning effect and its parameters are dependent on specific experimental conditions, thus they are specific to the catalyst. To overcome these limitations, this Thesis proposes two novel modelling strategies. First, a physically consistent mathematical extrapolation strategy is developed to extend the applicability range of the empirical Dyson-Simon model to lower pressure regimes. Second, an innovative hybrid modelling framework enhances first-principles kinetic model with data-driven symbolic regression methodologies, to improve the modelling of catalyst deactivation by poisoning capturing the non-linear temperature-poisoning effect on the catalyst and of the effect of irreversible reversible deactivation. Results indicate that the proposed model offers a significantly superior accuracy in the prediction of the produced ammonia compared to literature empirical models, and successfully describes the complex dynamics of cumulative catalyst poisoning, proving that the integration of machine learning into mechanistic frameworks successfully provides a comprehensive, dynamic, and reliable tool for the simulation and optimization of industrial ammonia synthesis reactors. These modelling strategies are evaluated on the industrial case study of an extensive dataset collected from high-throughput experimental campaigns.