Abstract
Accurate prediction of reaction energetics remains a fundamental challenge in computational chemistry, as conventional density functional theory (DFT) often fails to reconcile high accuracy with computational efficiency. Here, we introduce Deep post-Hartree-Fock (DeePHF), a machine learning framework that integrates neural networks with quantum mechanical descriptors to achieve CCSD-(T)-level precision while retaining the efficiency of DFT to solve the reaction problems. By establishing a direct mapping between the eigenvalues of local density matrices and high-level correlation energies, DeePHF circumvents the traditional accuracy-scalability tradeoff. Trained on a limited data set of small-molecule reactions, our model demonstrates superior performance across multiple benchmark data sets, exhibiting exceptional transferability. In fact, its accuracy even surpasses that of advanced double-hybrid functionals, all while maintaining O-(N(3)) scaling. DeePHF offers a promising pathway to bridge the gap between high-level quantum chemistry methods and the practical demands for scalable, accurate models in computational chemistry, and with further refinement, it is poised to make significant contributions to the advancement of chemical reaction modeling.