Zum Hauptinhalt springen

Natur- und Lebenswissenschaften

Viktor Zaverkin

E-Mail

zaverkin@theochem.uni-stuttgart.de

Alter

26

Aktuelle Tätigkeit

Promovend*in


Institution

University of Stuttgart, Institute for Theoretical Chemistry, Computational Chemistry Group

Stuttgart

Baden-Württemberg


Biographie

My name is Viktor Zaverkin and since March 2019 I am a Ph.D. student at the University of Stuttgart with Prof. Johannes Kästner. While at school, I have developed a passion for math and natural science, especially for physics and chemistry. In October 2014 this passion led me to the University of Stuttgart where I studied materials science, an interdisciplinary subject comprising both physics and chemistry. During my undergraduate studies, I became interested in theoretical physics and chemistry which turned around my world. I have graduated in February 2019 from the Institute for Theoretical Chemistry, where I worked on instanton theory, the aim of which is the description of the quantum-mechanical tunneling effect in chemical reactions. The work in the group of Prof. Johannes Kästner awakened my interest in molecular machine learning, which is the subject of my current work. In 2019 I have been awarded the Artur Fischer Prize for outstanding graduation and since 2020 I receive a scholarship from the “Studienstiftung des Deutschen Volkes”.

Fragestellungen im Themenfeld Künstliche Intelligenz

The chief goal of my research project is the application of machine learning methods to construct potential energy surfaces with the aim of modeling chemical reactions. The main prerequisite for a proper ML model for molecules and solids is the incorporation of manifold symmetries: the invariance of a chemical system with respect to translation, reflection, or rotation of the whole molecule and to permutation of atoms with the same nuclear charge. This can be achieved by a proper coordinate transformation which I developed by exploiting the mathematical properties of Gaussian type orbitals [JCTC 2020, 16, 8, 5410–5421]. The efficiency of ML methods depends on the quality and expressiveness of training data. I'm using the method of active learning derived in the framework of optimal experimental design to construct highly informative chemical data sets [submitted to MLST]. The specific applications of the developed methods range from the description of nitrogen atom dynamics on top of amorphous solid water at astrochemical conditions [MNRAS 2020, 499, 1, 1373–1384] to catalytic research questions, e.g., the simulation of covalent organic frameworks.