Deep Learning Functionals based on the Adiabatic Connection

Density functional theory (DFT) is vital for advancing our understanding of molecular and material properties, offering a unique balance between predictive power, versatility, and computational efficiency, despite its challenges in accurately describing strongly correlated systems. The adiabatic connection (AC) formalism offers a unique advantage in density functional approximations (DFAs) construction by allowing for direct and explicit consideration of strong correlation effects, enabling the development of interpolation models that depend on both strongly interacting and non-interacting limits of the AC, providing a more accurate estimation of exchange-correlation energies [1,2]. Our method combines the rigorous principles of the density-fixed AC with the power of deep learning, offering a transformative solution for improved accuracy and efficiency in DFT calculations. By leveraging deep learning techniques in the local interpolation along the AC, we have developed models ensuring size consistency and the accurate capture of strong correlation effects. The results of extensive testing on representative chemical systems are presented, showcasing significant advancements over state-of-the-art DFT.

References

[1] S. Vuckovic, T. J. P. Irons, A. Savin, A. M. Teale, P. Gori-Giorgi, J. Chem. Theory Comput. 2016, 12, 2598–2610.

[2] S. Vuckovic, T. J. P. Irons, L. O. Wagner, A. M. Teale, P. Gori-Giorgi, Phys. Chem. Chem. Phys. 2017, 19, 6169–6183.