Application of density functional theory to heterogeneous catalysis and development of computational methods for photochemistry

Abstract

This Ph. D. thesis focuses on modeling heterogeneous catalysis and developing methods for quantum mechanical photochemistry. Density functional theory (DFT) is one of the most widely used methods for structural and mechanistic studies of chemical reactions in heterogeneous catalysis. DFT has been powerful for the discovery of new structures, functionalities of new materials, and in the determination of possible reaction mechanisms. Despite the efforts to study Cu-catalyzed methane oxidation to methanol, there has been no consensus thus far about the actual structure of the active site(s) or the mechanism(s). In Chapter 2, we have modeled the active site of and mechanistic pathways for catalysis of methane to methanol through DFT calculations in Gaussian and through enzyme biomimicking enabling the identification and validation of the material Cu-ZIF for controlled C-H bond activation. Photochemistry is very important for organic synthesis and photosynthesis. We need quantum mechanical methods which are accurate as well as computationally affordable. For direct dynamics, in Chapter 3, we focus on the usage of the curvature-driven coherent switching with decay of mixing dynamical method Like Trajectory Surface Hopping, it predicts physically realistic asymptotic behavior, and like the Ehrenfest method, it allows continuous trajectory propagation (without hops) in regions of strong state interaction, making it more robust with respect to the choice of representation. Since the method is driven by the curvature of the adiabatic surfaces, it saves the cost of calculating exact nonadiabatic coupling vectors by electronic structure theory. Dynamics calculations on fitted potential energy surface (PES) are orders of magnitude faster than direct dynamics, but accurate PES prediction is computationally very expensive for even small systems. Our group has developed a machine-learning method called diabatization by deep neural network (DDNN), which can provide learned adiabatic PESs from input points on adiabatic surfaces. However, a notorious drawback for machine-learned potentials is their unreliable behavior when applied to data quite different from that used for training. In Chapter 4, we have developed a physics-based machine learning technique called parametrically managed diabatization by a deep neural network (PM-DDNN), enabling construction of analytical machine-learned smooth ground and excited state PES and extrapolation of the PES for O3 outside the training region. The PES from PM-DDNN gives excellent results for collision dynamics of O + O2 calculated with the ANT program. In Chapter 5, we benchmarked and developed DFT datasets for studying the photodissociation of o-fluorothiophenol. In particular, we presented a combined ∆-learning method (using a higher-level and lower-level DFT dataset) and physics-based NN, called dual-level parametrically managed neural network (DL-PMNN), to obtain the 33-dimensional ground state PES, extrapolated the PES outside the training region, and validated the PES by running trajectories using the SHARC-MN program.