Preview

PDF, English Download (48MB) | Terms of use

Abstract

The transition from metal–based cathodes to sustainable, earth-abundant alternatives is constrained by the poor electronic conductivity and high electrolyte solubility of most organic redox materials. This thesis develops an integrated quantum-chemical and machine-learning framework that overcomes these barriers and delivers a fast, data-driven route from molecular concepts to high-performance organic cathodes.

First, a density-functional-theory (DFT) workflow was benchmarked against experimental data and used to create ReSolved, a curated dataset of approximately 19 000 closed- and open-shell organic molecules in five solvents. A graph neural network with a solvent-aware set-transformer read-out reproduces absolute reduction potentials with a mean absolute error of $\approx$0.20 eV and transfers accurately to unseen solvents. Coupled to an evolutionary algorithm, the model enables inverse design, rapidly proposing synthetically accessible molecules tailored to specific battery chemistries.

To extend molecular insights to extended solids, the thesis introduces deCOFpose: an automated, topologically aware fragmentation algorithm that dissects crystalline covalent organic frameworks (COFs) into chemically meaningful nodes and linkers. Applied to the CoRE-COF database, deCOFpose successfully processes >70 % of structures, reveals systematic trends between fragment properties and COF band gaps, and feeds a modified PORMAKE builder that can enumerate millions of plausible frameworks for high-throughput screening.

Addressing composite electrodes, a semi-empirical tight-binding protocol was combined with a symmetry-adapted, P6mm wallpaper-group equivariant graph neural network (WallpaperNet) to predict adsorption geometries, binding energies, and force fields for small molecules on graphene.

Together, these contributions introduce a set of tools that span length scales from single molecules to periodic frameworks and composite materials, significantly reducing the screening time for viable organic cathodes. Beyond batteries, the methodologies, datasets, algorithms, and software are broadly applicable to catalysis and molecular sensing, furnishing a versatile platform for the accelerated discovery of sustainable, carbon-based functional materials.