A Novel Machine-Learning Based Mode Space Method for Efficient Device Simulations

In this work, we present a machine-learning based mode space method for efficient quantum transport simulations. By introducing a novel concept of projectability in training a machine learning model, our method reduces the size of the Hamiltonian more effectively compared to previous methods, while faithfully reproducing both the real and imaginary bands of interest. Through performing InAs nanowire FET simulations using the density functional theory (DFT) Hamiltonian and non-equilibrium Green's function (NEGF) method, we demonstrate that our method enables highly efficient simulations without losing accuracy.