Author: Ryan J. DiRisio

Publisher:

Published: 2022

Total Pages: 0

ISBN-13:

Understanding the role of hydrogen bonding in the structure and dynamics of water is an ongoing challenge in physical chemistry. In particular, understanding how the quantum mechanical effects of molecular vibrations govern the structure and dynamics of water is of interest. The cornerstone method used to study this phenomenon in this work is Diffusion Monte Carlo (DMC), which can be used to obtain the ground state vibrational wave function of any arbitrary molecule or molecular cluster. Instead of attempting to model bulk water and its properties outright, small, gas-phase molecular and ionic clusters of water, which provide model systems to study hydrogen bonding and proton transfer, are studied. To begin, DMC will be reviewed, and PyVibDMC, an open source, general purpose Python DMC software package developed as part of this work, will be discussed. As DMC is rigorously a ground state method, extensions to the DMC approach are required to obtain information about excited states. With excited state information, one can then directly compare simulation to experiment through theoretical and experimental spectroscopy. As such, next, the Ground State Probability Amplitude (GSPA) approximation is presented, and it is applied to protonated water clusters. In the GSPA approach, excited state wave functions are approximated based on simple products of polynomials of vibrational displacements with the ground state DMC wave function. The power of this approach is that one can construct a small basis through which to comprehensively examine the vibrational state space of the chemical system of interest. Extensions to the GSPA approach that incorporate excited state mixing and improved descriptions of higher-order excited states states will be presented as well. These improvements lead to good agreement between the GSPA theoretical and gas-phase experimental vibrational spectra of H7O3+ and H9O4+. Using this rich theoretical approach, we are able to draw connections between the molecular vibrations and structures that govern proton transfer and experimental spectroscopy of the clusters. A methodological procedure is presented next, which is the incorporation of machine learning into the DMC workflow. A potential energy surface is required for DMC simulations. Performing on-the-fly, ab initio potential energy calculations of molecular configurations in DMC simulations for systems beyond a few atoms is computationally intractable. As such, fitted potential energy surfaces are often employed for DMC simulations. However, as systems of interest increase in size, even the evaluations of these fitted surfaces become computationally demanding. To this end, a workflow is developed to use the large amount of data obtained from a small-scale DMC simulation to train a neural network to learn the potential energy surface of interest. Neural network structure, choice of descriptor, and hyperparameter optimization are reviewed and discussed in the context of other machine learning methods, and training data collection strategies are discussed, including the need to sample regions of the potential energy surface that are beyond regions accessed by a typical DMC simulation. Once the neural network surface is trained, it is evaluated in an extremely fast and highly-parallel manner, making DMC simulations significantly more efficient for H2O, CH5+, and (H2O)2. In the final section, DMC is set aside, and an exploration of the correlation between the vibrational spectral signature of an individual water molecule with its surrounding chemical environment is discussed. Specifically, the frequency of a hydrogen-bonded OH stretch in a water dimer pair is correlated to the number of solvating water molecules surrounding it. A quantum mechanical model is constructed to quantify this correlation, and applications of the model to a sample water cluster show the causality between the change in quantum mechanical electron density in the hydrogen bonding region of a particular OH bond and its OH stretch frequency. The application of the quantum model formalizes and explains empirical trends and categorization approaches put forth in previous work to characterize hydrogen bonding environments. This model is then applied to the water network found in a Cs+(H2O)20 cluster, where these trends are again quantified and then related to both the first and second solvation shell of a hydrogen-bond donor/acceptor water pair within the larger network.