✨ TL;DR
This paper introduces surrogate functionals, machine-learned energy functionals for orbital-free density functional theory that are optimized to produce correct ground-state densities rather than matching a physical reference. The approach requires only ground-state density data and achieves competitive accuracy while improving computational efficiency.
Orbital-free density functional theory (OF-DFT) is computationally efficient but requires accurate energy functionals to predict ground-state densities. Traditional machine-learned approaches for OF-DFT require expensive fully supervised training with reference energies and gradients at many points away from the ground state. Additionally, prior methods often require expensive orthonormalization steps that scale as O(N³), limiting applicability to larger systems.
The authors propose surrogate functionals that are trained with a novel gradient-descent-improvement loss function, which directly optimizes for the ability to recover ground-state densities through density optimization rather than matching a universal reference functional. They combine this loss with an adaptive sampling scheme that focuses learning on the optimization trajectories actually visited during inference. The gradient-descent-improvement loss is designed to guarantee exponential convergence of densities to the ground state.
What the paper shows.
On QM9 and QMugs benchmarks, surrogate functionals achieve density errors competitive with or better than state-of-the-art fully supervised machine-learned OF-DFT methods. The method demonstrates improved computational efficiency for larger systems by removing the expensive orthonormalization step, resulting in better runtime scaling compared to prior approaches.
The paper does not explicitly discuss limitations, but implicit constraints include: the approach is evaluated only on QM9 and QMugs datasets; the generalization to systems significantly larger than those in the benchmarks is not demonstrated; and the practical applicability to real materials discovery workflows beyond the tested benchmarks remains unclear.
✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.