Molecular dipole moment learning via rotationally equivariant derivative kernels in molecular-orbital-based machine learning

This study extends the accurate and transferable molecular-orbital-based machine learning (MOB-ML) approach to modeling contribution of electron correlation dipole moments at cost Hartree–Fock computations. A MOB pairwise decomposition part moment is applied, these pair could be further regressed as a universal function MOs. The features consist energy their responses electric fields. An interpretable rotationally equivariant derivative kernel for Gaussian process regression (GPR) introduced learn more efficiently. proposed problem setup, feature design, ML algorithm are shown provide highly models both energies on water 14 small molecules. To demonstrate ability MOB-ML generalized density-matrix functionals molecular organic molecules, we apply train test molecules from QM9 dataset. application local scalable GPR with mixture model unsupervised clustering scales up large-data regime while retaining prediction accuracy. In addition, compared literature results, provides best mean absolute errors 4.21 mD 0.045 kcal/mol models, respectively, when training 110 000 excellent transferability resulting also illustrated by predictions four different series peptides.