Data Driven Modeling In Understanding The Mechanism Of Heterogeneous Catalytic Reactions

Understanding the mechanism of a catalytic system, i.e., the active site, the sequence of bond breaking/formation and associated rates, and the controlling factor of catalyst activity is the key to designing and identifying the optimal catalyst that can substantially improve its performance, in terms of rate or selectivity. Traditionally, the task of elucidating the mechanism has been done in multiple ways, for example, (1) the surface science measurement in in-situ and operando reactor, (2) the steady-state and time-variant kinetic measurement, (3) or the computational simulation from quantum mechanics to continuous microkinetic modeling. Sometimes, the combination of different methods is employed, such as developing a thermodynamic consistent mean-field microkinetic model parameterized by density functional theory and re-estimated by experimental kinetic data. Such methodology is limited to computational modeling and kinetic experiments on well-defined surfaces due to the complexity of the catalyst surfaces and the reaction mechanism. The data generated from different sources, i.e., different types of experiments and different levels of computational simulations, are inherently disparate and heterogeneous. The processing, analysis, information extraction, and development of a unified model to understand the mechanism of heterogeneous catalytic systems from those disparate datasets is still an ongoing process.There are three aims in the thesis (1) to integrate and improve the microkinetic models with data-driven approaches, (2) to identify and develop machine learning algorithms for heterogeneous data sources, (3) to apply statistical numerical methods, i.e., uncertainty quantification, sensitivity analysis, experimental design, to facilitate the mechanistic knowledge extraction from the data-augmented modeling framework. Ultimately, this thesis posits that the integration of mechanistic model with emerging data-driven approaches and machine learning algorithms can be used in concert with heterogeneous experimental and computational chemistry data that enables the elucidation of the mechanism, the optimization of catalysts and working conditions, and the design and discovery of next-generation heterogeneous catalytic processes.This thesis presents several case studies of applications of this data-centered modeling paradigm. First, we develop a data-driven correction of ab-initio calculations and then augment it with microkinetic models to improve predictions and quantify uncertainty. Second, we apply global sensitivity analysis on mechanistic modeling to understand the importance of kinetic factors as an extension of the traditional local sensitivity method. Third, we customize and design various machine algorithms for the critical components in formulating microkinetic models, such as entropy, binding energy, and reaction propensity, to correct the oversimplified approximation, to improve model performance with fewer data, and to provide uncertainty quantification capability. Fourth, we develop regression and sparsification based methods to recover the kinetic information from modulation excitation spectroscopy. We highlight these methods and applications with illustrative examples.