Year: 2026
Type: preprint
Abstract: Machine-learned exchange–correlation functionals promise accuracy beyond traditional forms, yet so far they trade interpretability for flexibility. Here we introduce DL-xDH26, a glass-box foundation model for density functional development on the fifth rung of Jacob’s ladder, with separate exchange and correlation expert towers, interpretable enhancement-factor and shift parameters relative to a physical base, and exact degree-one homogeneity. Trained exclusively on relative energies, it recovers atomic exchange and correlation energies and their quantum shell structure, and produces smooth, size-consistent dissociation curves without dispersion corrections, while reaching errors below 1 kcal/mol on the GMTKN55 and MGCDB84 benchmark suites. Lightweight fine-tuning then specializes the same backbone to cyclization reactions, distorted organic geometries, and organometallic reaction energetics while retaining its broad accuracy.
Source: ChemRxiv
Authors: Sheng Bi, Shiqi Liu, Igor Ying Zhang, Xin Xu
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7211931810