PaperlayerView exact revision on arXiv

Source-linked AI summary

Hafnia-based Double Layer Ferroelectric Tunnel Junctions as Artificial Synapses for Neuromorphic Computing

Benjamin Max, Michael Hoffmann, Halid Mulaosmanovic, Stefan Slesazeck, Thomas Mikolajick

arXiv:2107.00945v1physics.app-phcond-mat.mtrl-scics.ET

TL;DR

Conventional von Neumann systems face challenges in implementing brain-like analogue computing, motivating devices for spiking neural networks. This paper studies a double-layer HZO-based FTJ as an artificial synapse, demonstrating pulse-controlled resistance changes and spike-timing-dependent plasticity behavior.

  • Problem

    Implementing brain-like analogue computing on prevalent von Neumann systems faces high power consumption, limited parallel processing, long computation times, and scalability problems.

    Source: INTRODUCTION · p. 2

  • Method

    The authors investigate TiN/HZO/Al2O3/TiN double-layer FTJs, varying voltage-pulse amplitude, width, number, and timing to emulate synaptic behavior.

    Source: INTRODUCTION · p. 4, RESULTS AND DISCUSSION · p. 14

  • Results

    The FTJs exhibit gradual resistance tuning with voltage pulses and STDP behavior under time-delayed pre- and post-neuron signals.

    Source: RESULTS AND DISCUSSION · pp. 12, 15, 16

  • Takeaways & Limitations

    The demonstrated resistance tuning and STDP behavior support the potential use of hafnia-based FTJs as artificial synapses in spiking neural networks.

    Source: CONCLUSION · p. 19

    At device diameters in the 5–30 nm HZO grain-size range, switching may become more abrupt, and further study is needed.

    Source: RESULTS AND DISCUSSION · p. 14

Abstract

from arXiv · show

Ferroelectric tunnel junctions (FTJ) based on hafnium zirconium oxide (Hf1-xZrxO2; HZO) are a promising candidate for future applications, such as low-power memories and neuromorphic computing. The tunneling electroresistance (TER) is tunable through the polarization state of the HZO film. To circumvent the challenge of fabricating thin ferroelectric HZO layers in the tunneling range of 1-3 nm range, ferroelectric/dielectric double layer sandwiched between two symmetric metal electrodes are used. Due to the decoupling of the ferroelectric polarization storage layer and a dielectric tunneling layer with a higher bandgap, a significant TER ratio between the two polarization states is obtained. By exploiting previously reported switching behaviour and the gradual tunability of the resistance, FTJs can be used as potential candidates for the emulation of synapses for neuromorphic computing in spiking neural networks. The implementation of two major components of a synapse are shown: long term depression/potentiation by varying the amplitude/width/number of voltage pulses applied to the artificial FTJ synapse, and spike-timing-dependent-plasticity curves by applying time-delayed voltages at each electrode. These experimental findings show the potential of spiking neural networks and neuromorphic computing that can be implemented with hafnia-based FTJs.

INTRODUCTION

Neuromorphic systems seek efficient learning and computation, but mapping brain-like analogue computing onto von Neumann systems brings power, parallelism, and scalability challenges. The introduction presents FTJs as SNN synapses whose polarization-dependent resistance can be tuned, with a double-layer design addressing thin-film fabrication challenges.

INTRODUCTION

  • Neuromorphic computing faces power, parallel-processing, and scalability challenges when analogue brain-like computing is implemented on prevalent digital von Neumann systems.Emerging memory and logic devices, including FTJs, are proposed for circuits implementing spiking neural networks.Source: INTRODUCTION · p. 2
  • FTJs are proposed as artificial synapses for SNNs, where gradual resistance switching can emulate learning through synaptic plasticity.Multilevel resistance states support this emulation, while reproducible, non-volatile states are important for Hebbian learning and memory.Source: INTRODUCTION · p. 3
  • A polarization reversal changes the effective tunneling barrier in an FTJ, producing distinct high- and low-resistance states.HZO FTJs are described as offering CMOS-compatible processing, low power, fast switching, and long data retention.Source: INTRODUCTION · p. 3
  • The artificial FTJ synapse's resistance can be tuned by varying pulse amplitude, width, or number, while STDP is tested with time-delayed electrode voltages.Source: INTRODUCTION · p. 4

RESULTS AND DISCUSSION

The TiN/HZO/Al2O3/TiN FTJ supports gradual resistance tuning and time-dependent synaptic plasticity, with switching behavior linked to ferroelectric polarization and pulse conditions. Its intermediate resistance states are described by a parallel-resistor model, while retention and scaling constrain practical use.

RESULTS AND DISCUSSION

  • HZO polarization reversal changes the double-layer tunneling barrier, switching the device between an approximately 1 GΩ LRS and a 14 GΩ HRS.The reported current mechanisms are tunneling through the dielectric for LRS and modified Fowler–Nordheim tunneling through the dielectric and part of the ferroelectric for HRS.Source: RESULTS AND DISCUSSION · p. 6
  • Intermediate states agree with a parallel-resistor model, but retention loss from depolarization fields narrows the memory window and may worsen at higher temperatures.Retention measurements were made at room temperature; the model relates resistance to the fraction of switched domains.Source: RESULTS AND DISCUSSION · pp. 7–8, Figure 1 · p. 9
  • Switching kinetics follow a nucleation-limited voltage-time trade-off, with 5–7 V needed for switching times from nanoseconds to milliseconds and a fitted exponent n=1.5.The reported fit to the switching model is observed in Figure 2.Source: RESULTS AND DISCUSSION · pp. 10–11, Figure 2 · p. 10
  • Repeated voltage pulses tune resistance between LRS and HRS, with up to 40 reset pulses producing a nearly linear resistance increase and pulse amplitude controlling tuning granularity.Higher set amplitudes switch more quickly; 4 V permits finer tuning than 5 V under the reported 10 µs pulse conditions.Source: RESULTS AND DISCUSSION · pp. 11–12, Figure 3 · p. 13
  • Small positive pre-to-post delays decrease resistance and strengthen the synapse, while small negative delays increase resistance and weaken it.The measured resistance change depends on the time delay between sawtooth action-potential pulses applied at the two electrodes.Source: RESULTS AND DISCUSSION · pp. 14–15, Figure 4 · p. 17
  • Different action-potential waveforms produce distinct predicted STDP curves, including asymmetric, plateau-like, and symmetric resistance-change patterns.The asymmetric curve is attributed to the shifted resistance hysteresis; rectangular pulses yield a plateau, while opposite-sign pulses yield a symmetric curve that decreases resistance.Source: RESULTS AND DISCUSSION · p. 18, Figure 5 · p. 19

CONCLUSION

The two-layer TiN/Hf0.5Zr0.5O2/Al2O3/TiN FTJ separates polarization storage from tunneling and is investigated as an artificial synapse. Its resistance can be tuned through pulse schemes and its STDP behavior is flexible, but gradual switching at nanoscale remains unresolved.

CONCLUSION

  • The double-layer FTJ separates polarization storage and tunneling, while avoiding the high leakage and reduced ferroelectric properties reported for very thin ferroelectric-only junctions.The authors describe this separation as yielding a stable and efficient ferroelectric switching device.Source: CONCLUSION · p. 19
  • Pulse width, amplitude, or number can be varied to gradually change resistance, supporting long-term depression and potentiation; STDP curves are also demonstrated.The STDP learning-rate timing and scaling parameters can be tuned by action-potential waveform amplitude and width.Source: CONCLUSION · p. 20
  • Two metal electrodes support seamless back-end-of-line integration, and the authors describe the FTJ's low current as beneficial for very large neural networks.The paper also reports scaling potential by analogy with HfO2-based FeFETs, while noting that FTJs are expected to follow a similar trend.Source: CONCLUSION · p. 20
  • Whether nanoscale FTJs will retain the gradual switching observed in larger devices remains to be established.The authors state that further research is needed on this and related issues.Source: CONCLUSION · p. 21

EXPERIMENTAL

The experimental section describes fabrication of TiN/HZO/Al2O3/TiN capacitors on p-doped silicon and the electrical and ferroelectric measurement methods used to characterize them.

EXPERIMENTAL

  • TiN/Hf0.5Zr0.5O2/Al2O3/TiN capacitors were deposited on p-doped silicon, with 12 nm TiN electrodes, 12 nm HZO, and a 2 nm aluminum oxide dielectric.The HZO and aluminum oxide were grown by ALD at 260 °C, followed by a 600 °C, 20 s nitrogen anneal to crystallize HZO.Source: EXPERIMENTAL · p. 21
  • Circular 200 µm capacitor structures were formed using Ti/Pt dots and subsequent SC1 etching of the top TiN electrode.The Ti/Pt dots were deposited by ebeam evaporation through a shadow mask.Source: EXPERIMENTAL · p. 21
  • Current-voltage response was measured with a Keithley 4200 SCS, while ferroelectric properties were measured using 10 kHz triangular voltage pulses.Measurements used a semi-automatic probe station with pulse measurement units and an Aixacct TF3000 Analyzer, respectively.Source: EXPERIMENTAL · p. 21

AUTHOR INFORMATION

All authors contributed to writing the manuscript and approved its final version.

AUTHOR INFORMATION

Loading 2107.00945v1…