Deep Neural Network for Generation of the Initial Electrostatic Potential Profile

A deep neural network is trained to learn the electrostatic potential of the semiconductor device. In order to demonstrate its feasibility, pn diodes are considered. Various pn diodes with different doping densities are generated and the numerical solutions are calculated. The resultant electrostatic potential profiles are used in the training phase. Our numerical results clearly demonstrate that the trained neural network can provide the initial electrostatic potential reasonably well. Since the initial electrostatic potential is improved, the Newton-Raphson loop for the nonlinear Poisson equation can be converged within a smaller number of iterations.