An evaluation of local double hybrid density functionals

First steps towards range-separated local double hybrid (RSLDH) functionals are reported. Optimizations using human-designed local mixing functions (LMFs) for the position-dependence of exact exchange led either to a vanishing perturbational correlation contribution or to a constant LMF, i.e., to the ω DH25 double hybrid. ω DH25-D4 achieves the so far lowest WTMAD-2 value (2.13 kcal/mol) of a gDH for the GMTKN55 test suite. Then an “n-LMF” was optimized as a shallow neural network, giving the first true RSLDHs, so far in a post- ω DH25 manner. A well-structured LMF and WTMAD-2 values down to 1.88 kcal/mol are obtained, without calibrating exchange-energy densities. • We explore if position-dependent exact-exchange admixture is useful for double hybrids. • As a side result, we obtained the excellent ω DH25-D4 double hybrid. • Using a neural network local mixing function provides the first local double hybrids.