Autonomous discovery of reaction networks from active-learning-driven reactive molecular dynamics

Reaction networks underlie complex chemical systems such as combustion, catalysis, and atmospheric chemistry and can comprise thousands of elementary reactions. Ab initio nanoreactor molecular dynamics explores such networks efficiently by accelerated sampling of the potential energy surface, but computational cost limits how far it can sample. This work presents a transferable methodology coupling nanoreactor molecular dynamics with active learning to autonomously generate reaction networks from starting molecules. An agentdeveloped parser extracts elementary reactions from trajectories, and a large-language-model coding agent improves transition-state search strategies. Applied to the atmospheric oxidation of dimethyl sulfide, the methodology yields 980 nanoseconds of simulation, requiring only 290,195 quantum-chemical calculations. Initializing reaction discovery from already discovered species drastically speeds up reaction network exploration leading to 13,271 elementary reactions and 6,995 species. The reaction network covers most of the established mechanism but also unveils novel species such as an isomer of the key intermediate HPMTF.