Modern energy, environment, and materials challenges hinge on the ability to predict and control vast reaction networks spanning thousands of steps across chemically diverse molecules and radicals. Detailed chemical kinetic models are the computational backbone of this enterprise, but their fidelity is limited by the breadth and quality of underlying properties. Thermochemical parameters, for example, determine equilibrium constants, reverse rate coefficients, and temperature dependences throughout reaction networks; even small biases can cascade into large errors in macroscopic predictions such as ignition delays, flame speeds, emissions, and catalyst performance descriptors.

Historically, researchers have relied on a patchwork of experimental measurements and high-level quantum chemistry for selected species, group-additivity estimates to interpolate across chemical space, and manual curation from dispersed literature. This workflow struggles to keep pace with the scale and complexity of contemporary combustion, atmospheric chemistry, and catalytic systems. 

In response, chemical research is undergoing a transformation driven by the interplay between quantum-mechanical rigor and artificial-intelligence innovation. Traditional quantum chemistry remains foundational for understanding molecular properties and reaction mechanisms, but it faces inherent limitations in scalability, computational cost, and adaptability to complex systems. Meanwhile, the exponential growth of chemical literature and data has outstripped manual analysis, creating an urgent need for advanced tools that can synthesize knowledge and uncover hidden patterns.

This dissertation addresses these gaps by integrating high-throughput electronic-structure calculations, graph-based deep learning, and domain-specific large language models, to (i) generate benchmark-quality thermochemical datasets at scale for fuels and radicals, (ii) propose a novel deep learning architecture that learns transferable molecular representations reliable for larger and more multifunctional molecules and radicals, and (iii) automatically extract structured mechanistic knowledge from the chemical literature to accelerate hypothesis generation and model curation.