Using building block structures and a cooperative approach with neural networks and random forest to identify reactions: a case study on the dissociation of sodiated disaccharides

Abstract

A new search scheme utilizing machine-learning methods has been developed to explore the reactions of di-saccharides. It incorporates structure sampling, neural network potential (NNP) training, and target search methodologies, addressing the challenges of their structural diversity and flexibility. We introduce building block sampling to identify transition state (TS) structures and examine the dissociation mechanism of α-maltose under collision-induced dissociation conditions. With a decent NNP model with a mean absolute error of 5 kJ mol⁻¹ for M06-2X/6-311+G(d,p), the 4 main dissociation channels are explored, and more than 70 000 TSs can be located in an extensive search. To prioritize computational resources for low-energy TSs, a target search using random forest is conducted and low-energy TSs with only 42% of the extensive computational workload are identified. With the NNP-accelerated target search scheme, we demonstrated that di-saccharide reaction exploration can be done within a few days.