about the role of FC/FS systems in the training set may be correct, it is not the only possible explanation for their observations. The authors then showed that if they exclude the FC/FS systems from the training set, the resulting functional (DM21m) becomes much less accurate for the BBB dataset but somewhat more accurate for GMTKN55 and QM9 ones.Īlthough the conclusion of Kirkpatrick et al. The resulting functional, named DM21 by the authors, was found to be much more accurate than four functionals-SCAN-D3(BJ) ( 2, 3), ωB97X-V ( 4), M06-2X ( 5), PW6B95-D3(0) ( 6)-commonly used on bond-breaking benchmark (BBB), GMTKN55 ( 7), and QM9 ( 8) test datasets. The training set consists of 1161 reaction/molecular energies, 1074 energies of FC/FS systems, and 931 sets of reasonable Kohn-Sham orbitals. To prove this, the authors constructed a local hybrid functional with learnable (neural network–based) enhancement factors and trained it to simultaneously approximate the energies and Kohn-Sham orbitals of the target systems. ( 1) is that the inclusion of FC and FS systems in a training set of a sufficiently flexible density functional allows it to become outstandingly accurate even for systems with FC/FS character, on which density functional theory (DFT) usually fails. The main conclusion of Kirkpatrick et al.

Artificial academy 2 download append