Ferroelectric memristors for neuromorphic applications: design, fabrication, and integration
Open access
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
An artificial synaptic element consisting of a three terminal Ferroelectric Field-Effect Transistor (FeFET) with an oxide channel is presented in this thesis. Bio-inspired computing emerged as the forefront technology to harness the growing amount of data generated in an increasingly connected society. Dedicated hardware solutions are required to leverage its full potential, especially regarding power consumption and parallelism by co-locating memory and computing. A common denominator among most proposed neuromorphic computing architectures is a neural network consisting of neurons and synapses. In the analog domain, the state of a synapse is emulated by a programmable and persistent electrical conductance, for which multiple physical effects can be exploited. Among them, the ferroelectric effect promises a low power operation and high endurance due to the electrostatic nature of the polarisation switching.
In the first part of this thesis, the process development for the materials is reviewed. A Back-End-Of-Line (BEOL) compatible crystallisation of HfZrO4 (HZO), a CMOS friendly and scalable material, in the metastable ferroelectric phase is demonstrated by a millisecond flash lamp anneal. Also, the effect of the electrodes and film thickness is studied. It is found that TiN and WOx electrodes both support the stabilisation of the metastable ferroelectric phase and that the ferroelectricity vanishes for very thin HZO. Furthermore, the development of a semiconducting WOx channel is presented, including the effect of the deposition method and processing conditions on its electrical properties.
In the second part of the manuscript, the developed materials are combined in a FeFET device: a simple gate-first device layout is designed and then used to establish a direct link between the ferroelectric polarisation and the channel conductance. The fine-grained domain structure of HZO is used to demonstrate a programmable and persistent multi-state conductance. Moreover, the FeFETs display a good linearity and symmetry, a low cycle-to-cycle noise, fast programming speed, and low write energy. The device area, dynamic range, endurance, and large device-to-device variability call for additional improvement.
In the last part, the process is further developed with the objective of decreasing the device area, reducing the device-to-device variability, and increasing the dynamic range and endurance. An important change to allow for such improvements is the growth of WOx by atomic layer deposition instead of sputtering. In addition, a more complex design enables the integration in cross bar arrays. The result is a sub-μm size artificial synaptic element with a quasi-continuous resistance tuning and a fine-grained weight update. Moreover, the change of conductance appears over two timescales. It is found that a fast, saturating ferroelectric effect and a slow, less saturating ionic drift and diffusion process are responsible for the multi time scale behaviour. The FeFET exhibits an excellent endurance and ferroelectric retention thanks to the good interface between the ferroelectric and the oxide channel. Its reduced footprint is an important step towards dense integration. Also, it is found that as a consequence of the two physical effects leading to different timescales, the symmetry and linearity of the device deteriorate. Taking all these characteristics into account, the performance of the FeFET is assessed by simulating the classification of the MNIST dataset, resulting in an excellent accuracy of 88 % accuracy, making it well suited for neuromorphic and cognitive computing. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000552280Publication status
publishedExternal links
Search print copy at ETH Library
Contributors
Examiner: Luisier, Mathieu
Examiner: Noheda, Beatriz
Examiner: Offrein, Bert Jan
Examiner: Fompeyrine, Jean
Publisher
ETH ZurichSubject
neuromorphic computing; memristors; oxides; ferroelectric; hafnium; FIELD EFFECT TRANSISTORS, FET (ELECTRONICS); Zirconium; tungsten oxideOrganisational unit
03925 - Luisier, Mathieu / Luisier, Mathieu
Funding
732642 - Ultra-Low Power Event-Based Camera (SBFI)
871737 - BEOL technology platform based on ferroelectric synaptic devices for advanced neuromorphic processors (EC)
More
Show all metadata
ETH Bibliography
yes
Altmetrics