Permutation-Invariant-Polynomial Neural-Network-based Δ-Machine Learn-ing Approach: A Case for the HO2 Self-reaction and its Dynamics Study

The potential energy surface (PES) plays a central role in chemistry. As the size of the reaction system increases, it would be more and more difficult to develop its globally accurate full-dimensional PES. One unavoidable difficulty is that it is too expensive to calculate electronic energies of ample configurations for complicated reactions. Δ-machine learning is a highly cost-effective method as only a small number of high-level ab initio energies are required to improve a potential energy surface (PES) fit to a large number of low-level points. Here, we propose a permutation-invariant-polynomial neural-network (PIP-NN)-based Δ-machine learning approach to construct full-dimensional accurate PESs for complicated reactions. The approach is applied to the HO2 + HO2 → H2O2 + O2 reaction, a key process in combustion and atmosphere. The full-dimensional triplet state PES is constructed with a large number of density functional theory (DFT) points, which cover all dynamically relevant regions. Only 14% of the DFT dataset are used to successfully bring the DFT PES to the UCCSD(T)-F12a/AVTZ quality. On this PES of high quality, quasi-classical trajectory (QCT) calculations are performed to study the dynamics of the title reaction. A surprising mode-specific dynamics is observed, in which exciting a spectator mode leads to significant enhancement of the reactivity at low collision energy. This special mechanism can be attributed to increased attraction potential caused by the excited spectator mode. Such mode specificity may be quite prevalent in free radical reactions involving HO2, which is common in combustion and atmosphere.