ABSTRACT Due to their position‐dependent admixture of the exact‐exchange (EXX) energy density, local hybrid functionals (LHs) enable a flexible balance between reduced self‐interaction errors and smaller static‐correlation errors, allowing an escape from the usual zero‐sum game between these two central aspects of the development of density functional approximations. Recent LHs with strong‐correlation factors incorporated into their local mixing functions (LMFs) governing the position‐dependence of EXX admixtures have been particularly successful in this context. As only few exact constraints for LMFs are known regarding valence‐space behavior, some recent efforts have used machine learning in this context, and the recent LH24n functional with a “neural‐network LMF” (n‐LMF, DOI: 10.1021/acs.jctc.4c01503) has shown excellent performance for the large GMTKN55 test suite of main‐group energetics. However, so far the construction of n‐LMFs that also cover strong‐correlation effects has not been successful. Here we report the LH25nP functional that has an n‐LMF optimized in the presence of a fixed strong‐correlation factor. LH25nP‐D4 achieves a remarkable self‐consistent WTMAD‐2 value of 2.47 kcal/mol for the GMTKN55 set, the so far lowest value for a rung 4 functional. Mean absolute deviations of 2.4 kcal/mol for the large W4‐11RE reaction‐energy set are also the lowest known currently for rung 4. At the same time, very low fractional‐spin errors and excellent performance for the spin‐restricted dissociation of covalent bonds, as well as a curing of spin‐contamination problems in open‐shell transition‐metal complexes has been found, suggesting a clear deviation from the usual zero‐sum behavior. Transferability to organometallic transition‐metal energetics is so far less favorable, suggesting the need for a wider training of n‐LMFs that includes data for transition‐metal systems.
Wodyński et al. (Sat,) studied this question.