Density-driven error (DC-DFT)

🌱 Seedling — sparse and possibly wrong.

Density-corrected DFT decomposes the total error of an approximate functional evaluated on its own self-consistent density into a functional error (the functional applied to the exact density) and a density-driven error (the extra error from the approximate density), $\Delta E=\underset{\text{functional}}{\underbrace{E_{\mathrm{XC}}[\tilde{n}]-E_{\mathrm{XC}}[n_{\mathrm{exact}}]}}+\underset{\text{from }\tilde{n}\ne n_{\mathrm{exact}}}{\underbrace{(\text{density-driven})}}$. For abnormal (density-sensitive) cases — anions, charge-transfer, some transition-metal systems — swapping in a more accurate density (HF-DFT) removes the density-driven part and can beat the self-consistent result.

The angle I care about: density-driven error in forces and curvatures, hence in phonons (the PRB-targeted methods direction), and in the reference data used to train MLIPs — where a density-sensitive functional can poison the learned potential energy surface in a way that’s invisible to energy-only metrics.

Prerequisites

hybrids · self-consistency and the density sensitivity diagnostic

Builds toward: phonon workflows · MLIP training data quality