Date of Award

January 2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Chemistry

First Advisor

Mark R. Hoffmann

Abstract

Computational electronic structure theory has become an inexorable engine for understanding and even predicting the properties of novel molecules and materials. In this dissertation, two types of novel metal-containing systems, one a material and one a molecule, were examined using ab initio methods. In addition, progress on using machine learning (ML) to systematize the selection of qualitatively correct descriptions of molecular systems were examined.The ever-increasing demand for smaller, faster, and cheaper electronics has fueled the ongoing search for new materials. Silicene, a two-dimensional (2D) monolayer of silicon (Si) atoms arranging in a honeycomb lattice, has attracted much attention because of its compatibility with the current Si-based technologies. To enrich the properties of silicene, transition metals (TMs) are often integrated into the silicene network. Binary monolayers of Si with different TMs are known as TM silicides. By using density functional theory (DFT) calculations, the structural, electronic, and mechanical properties of iridium (Ir)-silicide monolayers were investigated. Different plausible 2D structures of Ir-Si with various atomic ratios were modelled and the cohesive energies were then calculated for the geometry optimized structures with the lowest equilibrium lattice constants. Among a large number of candidate structures, we identified several mechanically (via elastic constants and Young’s modulus), dynamically (via phonon calculations) and thermodynamically stable Ir-Si monolayer structures. Ir2Si4 (called r- IrSi2) with a rectangular lattice (Pmmn space group) had the lowest cohesive energy of -0.248 eV (per IrSi2 unit) with respect to bulk Ir and bulk Si. The band structure suggested that metallic properties could be detected within Ir2Si4 monolayer. Hexagonal (P-3m1) and tetragonal (P4/nmm) cell structures were also found to be stable structures with 0.12 and 0.20 eV (per formula unit) higher cohesive energies, respectively. The interactions of stable Ir-Si monolayers with O2 and H2O molecules were also investigated. We found that Ir-Si monolayers are reactive to these molecules. The search for new materials does not stop at Earth. Many astrophysical molecules have been detected and studied in recent years. Alkaline earth metal monohydroxide radicals, especially calcium monohydroxide (CaOH) and strontium monohydroxide (SrOH), have attracted much attention due to their expected presence on rocky exoplanets, and by their potential applications in laser cooling and trapping technologies. Since samples of these interstellar molecules are not easily accessible, it is very challenging to study them using experimental methods. Theoretical approaches often become the more viable option, especially for providing detailed insights into the mechanisms of formation and dissociation of these radicals. Multireference perturbation theory (MRPT) methods are better alternatives to DFT for studying these electronic structures because a carefully balanced dynamical and non-dynamical (or static) electron correlations, starting from multiconfigurational self-consistent field (MCSCF), is needed. The second-order generalized van Vleck perturbation theory (GVVPT2), one of the MRPT approaches, was utilized to study CaOH and SrOH monomers. The dimerization of CaOH was also considered. The optimized geometry parameters of the ground state CaOH monomer as well as the vertical excitation energy of the first low-lying excited state are in good agreement with other ab initio methods and experimental data. On the other hand, the optimized geometry of the ground state of SrOH appears to be quasilinear as opposed to the linear geometry described in published literature. In addition, by applying relativistic corrections, the GVVPT2 optimized geometry of SrOH became more bent by approximately 6°. The biggest challenge of MRPT methods is to find a proper active space which is unique for each electronic structure and heavily dependent on user’s choice. If the active space is too big, the computational cost may be prohibitively expensive. If the active space is too small, the chemical properties of the structure may not be described appropriately, leading to inaccurate results. Hence, an improvement to these methods is much needed. Recently, ML has entered and is revolutionizing the quantum chemistry field. With the ability to automatically process a large amount of data and improve through experience, ML offers a solution to the user-specified active space problem of MRPT methods. This means that studying challenging and extensive chemical systems could become feasible in the near future while maintaining the accuracy of the underlying methods. In this study, we built our ML protocol to find a list of feasible active space configurations to be starting inputs for the GVVPT2 method. By implementing the reinforcement learning algorithm within ML, the machine was allowed to explore the unknown dynamical environment of the input chemical systems and receive feedback for each action it took. As a result, the machine gradually learned how to choose a proper active space. So far, we have used our ML algorithm to find the active space of the ground state water, triplet ground state methylene, and stretched water model systems. Good starting point active space configurations were found for all these three molecules which confirmed the active space selection ability of our ML model.

Download

Share

COinS