Prediction Methodology for Next-generation Device Characteristics using Machine Learning

Abstract

In this article, we propose a prediction methodology for next-generation device characteristics for process design kit (PDK) models that utilize various machine learning algorithms to achieve high accuracy and reduction of development turn-around time (TAT). The Berkeley short-channel IGFET model (BSIM) is used for generating datasets, while n-channel MOSFET compact model is used for peripheral circuits in dynamic random-access memory (DRAM) technology. Datasets for training comprise device characteristics that use compact models in present-generation products. In addition, a compact model of next-generation products is used for validating datasets. We demonstrate that our prediction methodology using random forest regression provides high accuracy of less than 0.7% RMSE and reduces development TAT.