Orbital-Free Surrogate Functionals Yield Transferable Interatomic Potentials and Electron Densities

Abstract

Orbital-free density functional theory seeks to compute the energy of an electronic system directly from its electron density, avoiding one-electron wave functions and thereby offering a route to scalable electronic structure calculations. Machine-learned orbital-free density functionals have recently achieved promising results on small organic molecules, predicting energies with sub-millihartree accuracy. However, their convergence in density optimization remains sensitive to hyperparameter tuning and architectural choices. Here, we extend the recently introduced (weak) surrogate functional framework - designed to predict ground-state electron densities only - to also yield their energy, resulting in "strong" surrogate functionals. We find that these learned functionals enable stable convergence across all tested neural network backbones, reducing electron density errors relative to the Kohn-Sham reference by an order of magnitude compared to previous OF-DFT methods. More importantly, the predicted energies are competitive with state-of-the-art machine-learned interatomic potentials (MLIPs) trained only on energies, while exhibiting superior generalization to larger, unseen systems.