Neural-network-based transfer learning for predicting cryo-CMOS characteristics from small datasets

Abstract

Transfer learning was examined to predict current-voltage (I-V) characteristics of MOSFETs at cryogenic temperatures. An experimental dataset was obtained from approximately 800 silicon-on-insulator MOSFETs using an automated cryogenic wafer prober to pre-train a 3-hidden-layer neural network (NN) model. Transfer learning based on the NN model was then conducted using another small dataset from 2 bulk MOSFETs. The transfer learning NN model predicted more realistic I-V characteristics and threshold voltages than a control NN model trained using only the small dataset. This study demonstrates cryogenic MOSFET characteristics prediction from a small dataset to reduce time and financial costs for developing cryo-CMOS devices.

Keywords

• cryo-CMOS
• neural network
• transfer learning
• MOSFET
• quantum computing
• device modeling