A density based machine learning force field for molecule-molecule nonbonded interactions

A new machine learning force field is introduced. In this force field, the intermolecular interaction is calculated based on tted electron charge density, while the intramolecular energy is modeled using a traditional atomic-structure-based machine learning force field model. The electron density of the molecular charge is represented by spherical charges centered at atoms and bond centers. The intermolecular interaction energy is calculated through the Coulomb interaction, exchange correlation interaction and Thomas-Fermi van Weiszaker kinetic energy, all based on the electron charge density. A polarization model is also introduced based on the local electric potentials near the atoms calculated based on the charge density. These interaction energies are combined to form the density based machine learning force field (DB-ML-FF). We show that the DB-ML-FF can accurately represent the intermolecular interaction, and it is practical to use this linear scaling model to carry out large system molecular dynamics simulations despite the fact the charge density is evaluated on a numerical grid.