Surrogate models for device design using sample-efficient deep learning

Abstract

Generation of training dataset for machine learning-based device design algorithm is expensive. To address this, we propose an active learning approach. Its efficiency is demonstrated through a Deep Neural Network (DNN) based Laterally Diffused Metal Oxide Semiconductor Field-effect Transistor (LDMOSFET) off-state breakdown voltage (BVDS,off) and specific on-resistance (Rsp) predictor. Our results show the possibility of 50% reduction in the training dataset size without compromising the baseline accuracy. Specifically, we compared eight sampling techniques and found that Informative-Query by Committee (I-QBC) and Diverse Informative-Greedy Sampling (DI-GS) techniques work best with Euclidean Norm of Prediction Error (ENPE).