UC Irvine

Reaction mechanisms and chemical reactivity are governed by high-dimensional potential energy surfaces (PESs). Exploring these surfaces is essential for discovering novel molecules and reaction pathways, yet exhaustive PES exploration becomes computationally prohibitive with each additional atom in a reaction system. Stoichiometry-preserving reaction networks (RNs) provide a tractable discrete representation of chemical space that enables systematic PES exploration. In this representation, nodes correspond to molecular collections of fixed stoichiometry and edges represent stoichiometry-preserving transformations. This dissertation investigates the structural properties of RNs and develops methods for more efficient pathway searching and targeted PES exploration within the open-source colibri2 software package. Comparative analysis of RNs against generative network models reveals that RNs closely resemble low-dimensional regular lattices with a small degree of random edge rewiring. Exploiting this underlying regularity, embedding methods are developed to construct low-distortion Euclidean embeddings of RNs. Combined with chemically informed edge-weighting, these embeddings preserve meaningful pathway rankings while enabling reaction pathway prediction without exhaustive graph traversal. RNs thus provide a low-dimensional discrete representation of PESs. To handle large chemical systems, this dissertation introduces bounded RN generation through energy threshold checks during node expansion and step limits relative to a starting node. These constraints restrict network growth to chemically relevant regions of the PES for specific reaction studies. Applications to atmospheric chemistry, enabled by the addition of support for radical reaction rules, demonstrate that these bounded RNs efficiently identify feasible pathways while suppressing irrelevant or infeasible ones. This dissertation establishes stoichiometry-preserving RNs as a scalable framework for PES exploration, supporting both broad chemical space exploration and targeted searches focused on specific regions of the PES within colibri2. By exploiting the lattice-like structure of RNs, this dissertation shows that accurate pathway prediction and targeted PES exploration are achievable without exhaustive graph traversal, enabling scalable computational investigation of complex chemical systems.