h1

%0 Thesis
%A Cüppers, Felix Johannes
%T Hafnium oxide based memristive devices as functional elements of neuromorphic circuits
%I Rheinisch-Westfälische Technische Hochschule Aachen
%V Dissertation
%C Aachen
%M RWTH-2023-05746
%P 1 Online-Ressource : Illustrationen, Diagramme
%D 2023
%Z Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
%Z Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023
%X Due to the approaching limit of the computational speed of classical von-Neumann architectures, data transfer-intensive cognitive applications in future information technology demand a paradigm shift. "Beyond-von Neumann" concepts such as biologically inspired neuromorphic circuits with adjustable synaptic weights promise an energy-efficient increase in computing power. In this context, novel memristive devices such as redox-based resistive random access memories (ReRAM) are investigated intensively. They combine nonvolatility, scalability and energy efficiency. Moreover, they also allow the programming of multiple different resistive states, which further increases the memory density in addition to the compact design. Due to their mixed ionic-electronic function, they differ significantly from purely electronic systems. Important criteria for the use of memristive devices in neuromorphic circuits are the operation parameters for the two switching modes abrupt and analog switching, the stochasticity of the switching processes SET and RESET, the variability of the resistance states HRS and LRS as well as the number of programmable states. In addition to the quantification of these parameters, the physical understanding of the processes taking place is crucial in order to make predictive statements about applicability and reliability in circuits. In this context, the exchange with and further development of physical models is essential. A typical filamentary ReRAM cell operating in the bipolar valence change mechanism (VCM) is composed of one or more insulating metal oxide layers and two metal electrodes, which differ in terms of work function and chemical reactivity. A preferred choice for the metal oxide layer by the industry is HfO2, since it is already available in semiconductor device fabrication lines. By intentionally introducing an additional sub-stoichiometric titanium oxide layer and using a chemically reactive titanium electrode and an inert platinum electrode, reproducible and stable switching behavior is obtained. In this work, the described switching modes are systematically analyzed on nanoscale Pt/HfO2/TiOx/Ti/Pt devices based on statistical ensembles. The devices are highly comparable to industrially available options. With the aid of compact model simulations, the results are physically interpreted to obtain a comprehensive description of the devices as a foundation for usage in future "Beyond-von Neumann" concepts. The results allow an evaluation of the HfO2-based ReRAM cells with respect to their application in novel neuromorphic circuits. ReRAM devices of atomic layer deposition (ALD)-grown 3nm thick HfO2 films were fabricated as cross-point devices with sizes ranging from 10000 nm2 to 3600 nm2. Functional devices of 1600nm2 were demonstrated as nano-plug structures. Extensive statistical characterization of the electroforming behavior and the switching stability as well as calibration of the switching properties by means of parameter variation form the basis for the differentiated analysis of the switching kinetics with rectangular voltage pulses between 100 ns and 1 s. The study of the abrupt switching kinetics in SET and RESET revealed the influence of the high-resistance state (HRS) on the switch-on process (SET) on the one hand and the influence of a series resistor on the switch-off process (RESET) on the other hand. Using the physically motivated compact model "JART v1b" developed in cooperation between IWE-II of RWTH Aachen University and PGI-7, it could be shown that, in addition to the delay time, the transition time in SET, which is difficult to access experimentally, also depends significantly on the HRS. Furthermore, if the low-resistance state (LRS) approaches the internal series resistance, a significant time delay of the RESET process is caused. By restricting HRS and LRS to a medium resistance range, the delay times can be minimized. Thus, these transition regions can be efficiently used for analog switching. Quantitative studies in this operation mode revealed that by appropriate choice of the voltage amplitudes, the behavior of the cells can be controlled to meet the requirements of neuromorphic circuits such as symmetry and programmability of intermediate conductance states. Further detailed investigations on the stochasticity of SET voltages over repeated switching operations and between different devices, performed on an extended device ensemble, allowed evaluation of parallel-connected devices as artificial synapses. The synapse was demonstrated both experimentally and in simulation using an extended version of the JART v1b model that includes device variability. The subsequent successful demonstration of synapses in a spiking neural network highlights the potential of memristive devices for neuromorphic circuits. The results illustrate that the range of applications can be further extended through a focused combination of device development and circuit design. In summary, this work shows that nanosized filamentary ReRAM devices have a high potential for use as artificial synapses in neuromorphic circuits of the future computer generation. The obtained results contribute to a deeper physical understanding of the analog and abrupt switching behavior and demonstrate the wide range of possible applications.
%F PUB:(DE-HGF)11
%9 Dissertation / PhD Thesis
%R 10.18154/RWTH-2023-05746
%U https://publications.rwth-aachen.de/record/959639