Directional Roaming in Formaldehyde Revealed by Dynamics Simulations on a Machine-Learnt Potential

The photodissociation study of formaldehyde (HCHO) is a benchmark for unveiling different types of unimolecular reactions. The roaming mechanism in HCHO is one such reaction that has been of great interest to chemists. Here, we utilize a machine learning approach, combining active and Δ-machine learning to achieve coupled-cluster-level accuracy, to explore the full configurational space of the HCHO → H2 + CO reaction. Accelerated dynamics reveals that roaming trajectories have certain structure and directionality. It is observed that H predominantly roams around the O atom with or without forming an O–H bond. It has been shown that roaming around the O atom is stabilized by electronic interaction terms. Two distinct abstraction mechanisms for the formation of H2 are revealed that are in line with recently reported cis and trans H-abstraction mechanisms. This is the first time the directionality of roaming has been observed via machine learning-accelerated dynamics.