Acceleration of Semiconductor Device Simulation With Approximate Solutions Predicted by Trained Neural Networks

In order to accelerate the semiconductor device simulation, we propose to use a neural network to learn an approximate solution for desired bias conditions. With an initial solution (predicted by a trained neural network) sufficiently close to the final one, the computational cost to calculate several unnecessary solutions is significantly reduced. Specifically, a convolutional neural network for the metal-oxide-semiconductor field-effect transistor (MOSFET) is trained in a supervised manner to compute the initial solution. In particular, we propose to consider a device template for various devices and a compact expression of the solution based on the electrostatic potential. We empirically show that the proposed method accelerates the simulation significantly.

Automated summary

A method for accelerating semiconductor device simulation using neural networks has been proposed, where a trained neural network predicts an initial solution to reduce computational costs. Specifically, a convolutional neural network computes the initial solution for MOSFET, considering device templates and a compact expression based on electrostatic potential. Empirical results show significant acceleration in simulation with this proposed method.