Derivative Informed Learning of Exchange-Correlation Functionals

2026-06-02

AI Summary

This work addresses the longstanding challenge in machine learning exchange-correlation (ML-XC) functionals of simultaneously achieving high accuracy and computational efficiency, as existing approaches have yet to consistently surpass costly hybrid functionals. The authors propose a hybrid distillation framework that trains low-complexity ML-XC functionals to reproduce B3LYP/def2-SVP reference data, introducing a novel DI-Loss that, for the first time, incorporates supervision on both first- and second-order derivatives of the energy with respect to the density matrix on the Grassmann manifold. This enables alignment at the level of functional response rather than merely matching fixed-point energies. The method yields substantial improvements: mean absolute error (MAE) in total energy is reduced by 66%, mean-field energy error drops from 1.2 to 0.8 mEh, self-consistent field (SCF) iterations are cut by up to 50%, and excitation energy prediction MAE decreases by 19–35%.

Abstract

Machine-learned (ML) exchange-correlation (XC) functionals aim to replace human-designed density functional approximations by learning directly from reference data, but they still do not consistently outperform traditional $\mathcal{O}(N^4)$-scaling hybrid functionals. We study a hybrid-distillation setting in which $\mathcal{O}(N^3)$-scaling ML-XC functionals are trained to reproduce B3LYP/def2-SVP targets. We introduce Derivative Informed XC-Loss (DI-Loss), a loss that incorporates additional information from the reference hybrid functional by supervising first and second derivatives of the energy on the Grassmannian of admissible density matrices. Rather than only matching the self-consistent fixed point, DI-Loss aligns the local first- and second-order response of the learned functional with that of the target functional. Across four evaluated architectures, DI-Loss consistently improves the main energy metrics. Averaged uniformly across architectures, the total-energy MAE decreases by 66% relative to energy and density supervision alone. The density-sensitive mean-field energy metric $E_ρ$ improves from $1.2$ to $0.8$ mEh on average, while dipole and $\mathcal{L}_2$ density errors do not improve uniformly. We further show that densities from the distilled functionals reduce hybrid-functional SCF iterations by up to 50%. In downstream TDDFT calculations, Hessian supervision improves excited-state predictions, with XCdiff reducing the mean excitation-energy MAE by 19 - 35%.

Problem

• exchange-correlation functional
• machine learning
• density functional theory
• energy derivatives
• hybrid functional distillation

Innovation

• Derivative Informed Loss
• Machine-Learned XC Functional
• Grassmannian Derivatives
• Hybrid Functional Distillation
• TDDFT Excitation Energy