Please use this identifier to cite or link to this item: http://doi.org/10.25358/openscience-7572
Authors: Stieffenhofer, Marc
Title: Multiscale modeling and deep learning: reverse-mapping of condensed-phase molecular structures
Online publication date: 22-Sep-2022
Year of first publication: 2022
Language: english
Abstract: Molecular processes can be studied at various levels of resolution that range from the fundamental, quantum mechanical description of electronic degrees of freedom up to the classical thermodynamic description of macroscopic quantities. For many systems, and in particular for those incorporating macromolecules, a single model is not able to capture all the relevant length- and timescales to thoroughly study a phenomena of interest. Multiscale modeling (MM) offers a solution by combining molecular models at different resolutions to address phenomena at multiple scales. On the low-resolution end, coarse-grained (CG) models are deployed to study the large-scale behavior of the system. These CG models are constructed by averaging over atomistic degrees of freedom. Their low resolution reduces the computational effort of the simulation and enables a faster exploration of configuration space. In addition to coarse-graining, a tight and consistent link between models of different resolutions calls for a reverse-mapping capable of reintroducing degrees of freedom as well. Reverse-mapping is routinely applied in the MM community, for example to compare simulation results with experimental data, to rigorously analyze the simulation results on a local scale, or to assess the stability and accuracy of the obtained CG structures. At the heart of this work is the development of deepbackmap (DBM), an approach for the reverse-mapping of condensed-phase molecular structures. The new method is based on machine learning (ML), a study of computer algorithms that use data to construct statistical models. Traditional schemes start from a rough coarse-to-fine mapping, which requires further energy minimization and subsequent molecular dynamics simulations to equilibrate the system. DBM directly predicts equilibrated molecular configurations that agree with the Boltzmann distribution. Moreover, DBM requires little human intervention, as the reintroduction of details is learned from training examples. During the course of this thesis, DBM is applied to various tasks involving reverse-mapping: The general performance and transferability of DBM is evaluated at the example of a polymeric system consisting of polystyrene molecules. Beside an excellent accuracy of structural properties for reverse-mapped configurations, DBM displays a remarkable transferability across different state points and chemical space. Moreover, reverse-mapping with DBM is performed to assess the quality of CG models at the atomistic resolution. In addition, DBM is applied to adjust local structural properties, such as bond lengths and angles, of configurations obtained with top-down molecular models in order to resemble target distributions obtained with structure-based models more closely. Finally, a ML-based scheme inspired by DBM is applied for temporal coherent reverse-mapping of molecular trajectories. Overall, this thesis demonstrates the advantages of integrating generative ML methods into the framework of MM, especially for problems that are difficult to solve from a pure physics-based perspective.
DDC: 530 Physik
530 Physics
540 Chemie
540 Chemistry and allied sciences
Institution: Johannes Gutenberg-Universität Mainz
Department: FB 09 Chemie, Pharmazie u. Geowissensch.
MaxPlanck GraduateCenter
Place: Mainz
ROR: https://ror.org/023b0x485
DOI: http://doi.org/10.25358/openscience-7572
URN: urn:nbn:de:hebis:77-openscience-2f397422-7012-4018-9f86-c34ddba4c7072
Version: Original work
Publication type: Dissertation
License: CC BY
Information on rights of use: https://creativecommons.org/licenses/by/4.0/
Extent: XXV, 181 Seiten (Illustrationen, Diagramme)
Appears in collections:JGU-Publikationen

Files in This Item:
  File Description SizeFormat
Thumbnail
multiscale_modeling_and_deep_-20220819094012393.pdfdissertation41.85 MBAdobe PDFView/Open