Completing density functional theory by machine learning hidden messages from molecules

This study demonstrates a systematic construction of the exchange-correlation energy functional in Kohn-Sham density functional theory (DFT) using machine learning. A feed-forward neural network relates density and energy in reference molecules, allowing for functional derivative calculation via back-propagation. Surprisingly, a functional trained on only a few molecules accurately predicts properties for hundreds of molecules with various first- and second-row elements, achieving accuracy comparable to standard functionals. The inclusion of a nonlocal density descriptor further enhances accuracy. This approach leverages machine learning to enrich the DFT framework.

Ryo Nagai, Ryosuke Akashi, Osamu Sugino