Neural-network Kohn-Sham exchange-correlation potential and its out-of-training transferability

Abstract: We incorporate in the Kohn-Sham self consistent equation a trained neural-network projection from the charge density distribution to the Hartree-exchange-correlation potential $n \rightarrow V_{\rm Hxc}$ for possible numerical approach to the exact Kohn-Sham scheme. The potential trained through a newly developed scheme enables us to evaluate the total energy without explicitly treating the formula of the exchange-correlation energy. With a case study of a simple model we show that the well-trained neural-network $V_{\rm Hxc}$ achieves accuracy for the charge density and total energy out of the model parameter range used for the training, indicating that the property of the elusive ideal functional form of $V_{\rm Hxc}$ can approximately be encapsulated by the machine-learning construction. We also exemplify a factor that crucially limits the transferability--the boundary in the model parameter space where the number of the one-particle bound states changes--and see that this is cured by setting the training parameter range across that boundary. The training scheme and insights from the model study apply to more general systems, opening a novel path to numerically efficient Kohn-Sham potential.