Paper detail

Machine Learning Exchange-Correlation Potential in Time Dependent Density Functional Theory

We propose a machine learning based approach to develop the exchange-correlation potential of time dependent density functional theory (TDDFT). The neural network projection from the time-varying electron densities to the corresponding correlation potentials in the time-dependent Kohn-Sham equation is trained using a few exact datasets for a model system of electron-hydrogen scattering. We demonstrate that this neural network potential can capture the complex structures in the time-dependent correlation potential during the scattering process and provide correct scattering probabilities, which are not obtained by the standard adiabatic functionals. We also show that it is possible to incorporate the nonadiabatic (or memory) effect in the potential with this machine learning technique, which significantly improves the accuracy of the dynamics. The method developed here offers a novel way to improve the exchange-correlation potential of TDDFT, which makes the theory a more powerful tool to study various excited state phenomena.