Solving the many-body problem in quantum mechanics requires the use of approximate methods. These methods come in several flavors that differ in accuracy and computational cost . For small molecules, highly accurate, explicitly correlated methods such as coupled cluster are usually feasible. For larger systems, these become too expensive and one has to revert to using faster low-level approximations such as density functional theory (DFT).

We are trying to improve the accuracy of DFT by incorporating machine learned density functionals that are trained on data from reference methods. To achieve this, we take an atom-based approach, projecting the electron density onto a set of localized atomic orbitals to create descriptors that serve as input to our machine learning models. Doing so we can generate specialized functionals that are size extensive and can achieve close to coupled-cluster accuracy on selected systems.

Ongoing research addresses how such a functional can be made quasi-universal, i.e. valid for a wide variety of systems both from the realm of molecules as well as solids.

Our code, neuralxc is available under a public license.