Ferroelectric HfZrO2 With Electrode Engineering and Stimulation Schemes as Symmetric Analog Synaptic Weight Element for Deep Neural Network Training

K. Y. Hsiang, C. Y. Liao, K. T. Chen, Y. Y. Lin, C. Y. Chueh, C. Chang, Y. J. Tseng, Y. J. Yang, S. T. Chang, M. H. Liao, T. H. Hou, C. H. Wu, C. C. Ho, J. P. Chiu, C. S. Chang, M. H. Lee*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

1 Scopus citations

Abstract

Atomic layer deposition (ALD)-based TiN electrode on ferroelectric HfZrO2 metal/ferroelectric/metal (MFM) capacitor and ferroelectric field-effect transistor (FeFET) is demonstrated experimentally with weight transfer, that is, Delta P, per pulse analysis through consecutive alternating potentiation/depression (Pot./Dep.) training pulses. The weight training pulse schemes are studied to have symmetric and linear synapse weight transfer to increase the accuracy and accelerate the deep neural network (DNN) training. With ALD TiN inserted, alpha(p)/alpha(d) = -0.63/-0.84, asymmetry vertical bar alpha(p) - alpha(d)vertical bar = 0.21, and polarization modulation ratio (Pot./Dep.) = 97%/98% are achieved for MFM capacitor, and alpha(p)/alpha(d) = -1.32/-1.88, asymmetry vertical bar alpha(p) - alpha(d)vertical bar = 0.56, and G(max)/G(min) > 10x are delivered for FeFET.

Original languageEnglish
Article number9180313
Pages (from-to)4201-4207
Number of pages7
JournalIEEE Transactions on Electron Devices
Volume67
Issue number10
DOIs
StatePublished - Oct 2020

Keywords

  • Training
  • Synapses
  • Electrodes
  • Capacitors
  • Tin
  • Field effect transistors
  • Modulation
  • Ferroelectric memories
  • hafnium
  • neural networks

Cite this