Cheminformatics and Machine Learning for Transition States

Dey, Anish (2025) Cheminformatics and Machine Learning for Transition States. Masters thesis, Indian Institute of Science Education and Research Kolkata.

Text (MS Dissertation of Anish Dey (20MS026)) 20MS026_Thesis_file.pdf - Submitted Version Restricted to Repository staff only Download (6MB)

Abstract

This thesis explores the intricacies of molecular representation and transition state modeling in computational chemistry. The thesis critically evaluates the SMILES notation, pointing out its shortcomings in representing coordination complexes and transition states. Molecular graph based models conventionally used are susceptible to bonding ambiguities as well as electron delocalization issues. To avoid these, we propose a new methodology that uses a Quantum Theory of Atoms in Molecules (QTAIM)-based graph representation. The method uses electron density topology, focusing on the critical points such as bonds, rings, and cages to establish more substantial chemical connectivity. Our density-driven, canonical graph model better represents intricate systems in coordination chemistry and reaction modelling. The second part of this thesis explores the application of Machine Learning (ML) to predict transition state energetics for hydrogen splitting reactions, focusing on barrier height modelling of Vaska’s complex in the SchNet framework. While the chemical space of potential catalysts includes highly active candidates, their identification is challenging due to the vast combinatorial complexity of ligand-substituent structures. Using the SchNet deep learning model, we explore barrier heights across a wide range of ligand-substituent combinations. This allows us to avoid the computational burden conventionally associated with quantum chemical methods and quickly screen effective catalysts. Finally, this offers a scalable solution to data-driven catalyst discovery and design.

Actions (login required)

View Item