Machine Learning Based Density Functional Method for Chemical Reactions I: Organic Reactions

The accurate computation of reaction energy and reaction barrier height is indispensable for the investigation of chemical reaction mechanisms and rates. In this research, we employ the Deep post-Hartree Fock (DeePHF) model in conjunction with the local density matrix derived from different base quantum chemistry methods (HF, PBE, B3LYP, M06-2X and $\omega$B97M-V). Through extensive training on high-precision small molecule reaction data, including reaction structures and high-precision NEB (Nudged Elastic Band) trajectories, our model showcases better performance across various reaction barrier height datasets compared to the common density functional theory(DFT). The precision of the model rivals that of the double hybrid functionals. Notably, DeePHF maintains high accuracy even when applied to small molecules, medium-sized molecules, and standard benchmark reactions datasets, displaying its substantial advantage in exploring reaction potential energy profiles.