Bitte benutzen Sie diese Kennung, um auf die Ressource zu verweisen:
Autoren: Kanekal, Kiran H.
Titel: Investigating the Effect of Coarse-Graining on Chemical Compound Space
Online-Publikationsdatum: 8-Apr-2021
Sprache des Dokuments: Englisch
Zusammenfassung/Abstract: Chemical structure-property relationships are essential for the development of new materials used in all facets of life. Practically, this process amounts to projecting regions of the chemical compound space (CCS) onto certain descriptors related to the property of interest, allowing the structure-property relationship to be inferred. The challenge in constructing these relationships usually stems from a lack of data, as their accuracy and transferability will depend on how well-sampled CCS is with respect to the chosen descriptors. High-throughput screening, in which the properties of compounds are determined in an automated fashion, is one strategy used to overcome this problem. However, for many properties of soft-matter systems, this approach is difficult to implement computationally. This difficulty arises due to the large costs associated with adequately sampling complex free energy landscapes at atomistic resolutions using established tools such as molecular dynamics (MD) simulations. Coarse-grained (CG) models, parameterized at lower resolutions compared to their atomistic counterparts, provide a means to circumvent these costs. However, many of these models are constructed in order to specifically reproduce the properties of a small number of compounds, making it difficult to generalize across CCS. In this work, we demonstrate that the coarse-grained Martini model reduces the size of CCS, and can be used in computational high-throughput screening methods to efficiently construct chemical structure-property relationships over wide ranges of CCS. We find that this reduction of CCS is due to a limited number of Martini interaction types, with multiple atomistic chemical fragments mapping to the same CG interaction type. We then investigate the relationship between unsupervised machine learning and coarse-graining, yielding strategies for parameterizing chemically transferable CG models from both a top-down and bottom-up perspective. We employ these data-driven techniques to parameterize new top-down CG models and quantify their transferability and accuracy as a function of the number of CG interaction types for each model. Finally, we develop a method that uses unsupervised machine learning in combination with the bottom-up multiscale coarse-graining technique to generate chemically-transferable CG models with high structural accuracy. We examine the limitations of both top-down and bottom-up approaches and make recommendations for the future development of these methodologies. Overall, our work demonstrates the means by which chemically-transferable CG models can be both constructed and utilized to efficiently infer chemical structure-property relationships for materials discovery.
DDC-Sachgruppe: 004 Informatik
004 Data processing
500 Naturwissenschaften
500 Natural sciences and mathematics
530 Physik
530 Physics
540 Chemie
540 Chemistry and allied sciences
570 Biowissenschaften
570 Life sciences
660 Technische Chemie
660 Chemical engineering
Veröffentlichende Institution: Johannes Gutenberg-Universität Mainz
Organisationseinheit: FB 09 Chemie, Pharmazie u. Geowissensch.
Veröffentlichungsort: Mainz
Version: Original work
Publikationstyp: Dissertation
Nutzungsrechte: in Copyright
Informationen zu den Nutzungsrechten:
Umfang: xvi, 229 Seiten
Enthalten in den Sammlungen:JGU-Publikationen

Dateien zu dieser Ressource:
Datei Beschreibung GrößeFormat 
kanekal_kiran-investigating_-20210222180034713.pdf58.15 MBAdobe PDFÖffnen/Anzeigen