Automatic differentiable nonequilibrium Green's function formalism: An end-to-end differentiable quantum transport simulator

The state-of-the-art first-principles quantum transport theory and modeling are based on carrying out self consistent atomistic calculations within the Keldysh nonequilibrium Green's function (NEGF) formalism. The atomistic model of the device can be at the tight-binding (TB) or the density functional theory levels, and NEGF determines the nonequilibrium carrier distribution under external bias and gate voltages. In this work, we report an end-to-end automatic differentiable NEGF simulator (AD-NEGF) within the NEGF-TB framework. ADNEGF calculates gradient information by automatic differentiation (AD) and the implicit layer technique while guaranteeing the correctness of forward simulation. The gradient information enables accurate calculations of transport properties that depend on the derivatives of the transmission coefficient and/or charge current. More interestingly, AD-NEGF can be applied to the extremely interesting inverse design problem; namely, with a desired transport property, AD-NEGF inversely finds a possible device Hamiltonian that would produce such a property.