Reproduce “Stable and Accurate Orbital-Free DFT Powered by Machine Learning” on OpenResearch

About this paper

Roman Remme, Tobias Kaczun, Tim Ebert, Christof A. Gehrig, Dominik Geng, Gerrit Gerhartz, Marc K. Ickler, Manuel V. Klockow and 5 more

Researchers at Heidelberg University developed STRUCTURES25, a machine-learned functional for orbital-free density functional theory that achieves stable and convergent variational density optimization, providing chemical accuracy (0.64 mHa MAEE) for molecular energies and densities across the QM9 dataset, and demonstrating nearly an order of magnitude speedup compared to Kohn-Sham DFT for larger drug-like molecules.

Key insights

• Generating diverse training data by perturbing KS-DFT effective potentials is crucial for learning a stable energy functional that enables fully convergent variational density optimization.
• Learning the aggregated $E_{TXC}$ functional, rather than individual components, simplifies the learning task, improves efficiency by avoiding quadrature grids, and enhances stability.
• Equivariant graph neural network architectures with tensorial messages and local message-passing cutoffs are essential for achieving both accuracy and favorable (O(N)) computational scaling, allowing generalization to larger molecules.

Summary by alphaXiv.