Logo Logo
Help
Contact
Switch language to German
Improving grid based quantum dynamics. from the inclusion of solvents to the utilization of machine learning
Improving grid based quantum dynamics. from the inclusion of solvents to the utilization of machine learning
In this work the development and refinement of methods that allow for a more efficient use of quantum dynamics calculations with the dynamic Fourier method (DFM) and the extension of the DFMs applicability to solvated systems are presented. The systems studied with these methods are primarily molecular reactions. The contents of this thesis can be divided in three parts. In the first, approaches to automate the necessary and difficult construction of reactive coordinates are developed. The reactive coordinates constructed are linear combinations of the Cartesian coordinates of the atoms in two cases and nonlinear combinations in one case. Of the two linear approaches, one uses points along the minimum energy path of the reaction to span a reactive subspace and the other uses classical trajectories that are run along the reaction path to extract essential motions. The nonlinear approach uses an autoencoder that learns an efficient low-dimensional description of the reactive space by using large amounts of trajectory data. All three methods are presented in detail and applied to example systems. The advantages of each method and the immense potential of the nonlinear approach are discussed. In the second part, the inclusion of dynamic solvent effects on the reactive solute is studied. This is important, because dynamic interactions can significantly alter the outcome of reactions. Three methods were developed in the course of this work. The first one treats the solvent implicitly as a continuum that causes frictional forces acting on the solute. This is computationally convenient but, as it cannot describe certain interactions -- such as collisions -- it is not suited for all systems. The second and third method treat the solvent particles explicitly, using frozen and classically propagated environments, respectively. Here, the third approach extends the second to a much larger field of application by means of a quantum-classical TDSCF. The three methods are applied to the practically important problems of the photogeneration of diphenylmethyl cations as reactive intermediates and the photorelaxation of uracil as a way to prevent photodamage in RNA. The third part is comprised of two minor improvements. The first deals with errors that can be introduced due to an incorrect treatment of an approximation within the Wilson G-matrix formalism, a formalism that offers a simple way to perform coordinate transformations. This not only applies to some implementations of the DFM method, but to all methods that use the G-matrix formalism with nonlinear coordinates. The second studies the use of undersampling in the DFM method to reduce the number of grid points and thus computational time. Its potential savings are demonstrated using a model system.
Quantum Dynamics, Machine Learning, Neural Networks, Solvent Effects, Grid Based Methods
Zauleck, Julius Philipp Paul
2018
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Zauleck, Julius Philipp Paul (2018): Improving grid based quantum dynamics: from the inclusion of solvents to the utilization of machine learning. Dissertation, LMU München: Faculty of Chemistry and Pharmacy
[thumbnail of Zauleck_Julius.pdf]
Preview
PDF
Zauleck_Julius.pdf

14MB

Abstract

In this work the development and refinement of methods that allow for a more efficient use of quantum dynamics calculations with the dynamic Fourier method (DFM) and the extension of the DFMs applicability to solvated systems are presented. The systems studied with these methods are primarily molecular reactions. The contents of this thesis can be divided in three parts. In the first, approaches to automate the necessary and difficult construction of reactive coordinates are developed. The reactive coordinates constructed are linear combinations of the Cartesian coordinates of the atoms in two cases and nonlinear combinations in one case. Of the two linear approaches, one uses points along the minimum energy path of the reaction to span a reactive subspace and the other uses classical trajectories that are run along the reaction path to extract essential motions. The nonlinear approach uses an autoencoder that learns an efficient low-dimensional description of the reactive space by using large amounts of trajectory data. All three methods are presented in detail and applied to example systems. The advantages of each method and the immense potential of the nonlinear approach are discussed. In the second part, the inclusion of dynamic solvent effects on the reactive solute is studied. This is important, because dynamic interactions can significantly alter the outcome of reactions. Three methods were developed in the course of this work. The first one treats the solvent implicitly as a continuum that causes frictional forces acting on the solute. This is computationally convenient but, as it cannot describe certain interactions -- such as collisions -- it is not suited for all systems. The second and third method treat the solvent particles explicitly, using frozen and classically propagated environments, respectively. Here, the third approach extends the second to a much larger field of application by means of a quantum-classical TDSCF. The three methods are applied to the practically important problems of the photogeneration of diphenylmethyl cations as reactive intermediates and the photorelaxation of uracil as a way to prevent photodamage in RNA. The third part is comprised of two minor improvements. The first deals with errors that can be introduced due to an incorrect treatment of an approximation within the Wilson G-matrix formalism, a formalism that offers a simple way to perform coordinate transformations. This not only applies to some implementations of the DFM method, but to all methods that use the G-matrix formalism with nonlinear coordinates. The second studies the use of undersampling in the DFM method to reduce the number of grid points and thus computational time. Its potential savings are demonstrated using a model system.