Almost all empirical parametrizations of dispersion corrections
in DFT use only energy errors, thereby mixing functional and density-driven
errors. We introduce density and dispersion-corrected DFT (D2C-DFT), a dual-calibration approach that accounts for density delocalization
errors when parametrizing dispersion interactions. We simply exclude
density-sensitive reactions from the training data. We find a significant
reduction in both errors and variation among several semilocal functionals
and their global hybrids when tailored dispersion corrections are
employed with Hartree–Fock densities.