An Automated Framework for Analyzing Structural Evolution in On-the-fly Non-adiabatic Molecular Dynamics Using Autoencoder and Multiple Molecular Descriptors

A major challenge in nonadiabatic molecular dynamics is to automatically and objectively identify the key reaction coordinates that drive molecules toward distinct excited-state decay channels. Traditional manual analyses are inefficient and rely heavily on expert intuition, creating a bottleneck for interpreting complex photochemical processes. To overcome this, we introduce a fully automated machine-learning framework that directly extracts these coordinates from on-the-fly trajectory surface hopping data. By combining an Autoencoder for nonlinear dimensionality reduction with clustering and information entropy analysis, our method autonomously maps reaction channels and pinpoints their governing structural motions. When applied to keto isocytosine and the methaniminium cation, the framework objectively revealed invovled reaction channels and corresponding active coordinates with high efficiency and accuracy. This work establishes an effective paradigm for mechanistic insight in excited-state dynamics, transforming raw trajectory data into clear, interpretable reaction mechanisms.