Locality meets machine learning: Excited and ground-state energy surfaces of large systems at the cost of small ones
Abstract
It is demonstrated that supervised machine learning of local environments of the atoms in a molecule can be very effectively combined with the density functional-based tight-binding method to provide a fast and size-extensive scheme for electronic structure calculations. We train our machine learning model on small and basic molecules and then run it successfully for large and complicated molecules. This facilitates investigations of structural, electronic, and optical properties of large model systems for which the conventional iterative self-consistent procedure becomes too costly. The fruitfulness of this shortsightedness-based scheme is shown for describing the energy landscape of the ground and low-lying excited states of several model molecules of variant sizes and complexity. The achieved accuracy in the tests supports the locality view to the electronic redistribution in a molecule and promises the efficiency of the machine learning equipped divide-and-conquer approach for solving the Schrödinger equation.
- Publication:
-
Physical Review B
- Pub Date:
- March 2020
- DOI:
- 10.1103/PhysRevB.101.115132
- Bibcode:
- 2020PhRvB.101k5132B