Learning Orbital-Free Density Functional Theory

Ziel

This project is motivated by a doubly frustrating gap. On one side, Hohenberg and Kohn have proven that the electron density is all that is needed to describe the ground state of a molecular system; and yet we recur to the use of orbitals, and hence Kohn-Sham DFT (KS-DFT), because an exact expression (or merely sufficiently accurate approximation) for the kinetic energy of an electron density remains elusive. The cost of KS-DFT remains prohibitive for very large systems, or for very high throughput screening scenarios. On the other side, machine-learned interatomic potentials are fast, but until now remain unreliable for more complex chemistries. LearningOFDFT aims to close this gap by learning the kinetic energy functional.

I will use geometric machine learning to develop an orbital-free spin density energy functional that obeys select exact constraints and that affords fully variational density optimization across charge, spin, and geometry. Steps include leveraging and expanding the OMol25 corpus, adding non-covalent interactions and exotic species to ensure coverage of charged, open-shell, and distorted systems; developing surrogate functionals that relax exactness away from the ground state to accelerate convergence and minimize error only where it matters most; training inference dynamics, by learning vector fields that guide densities toward the ground state and sampling these stochastically to obtain fast uncertainty estimates; hardening and releasing open-source code for immediate community use and further development.

If successful, the learned functional will prove an inspiration and a challenge for theoretical chemistry; and it will be a game changer for applied research, from biomolecular reactions to green materials, where existing methods are too inaccurate, or too slow.