h1

TY  - THES
AU  - Hennen, Tyler Aaron
TI  - Harnessing stochasticity and negative differential resistance for unconventional computation
PB  - Rheinisch-Westfälische Technische Hochschule Aachen
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2023-07044
SP  - 1 Online-Ressource : Illustrationen, Diagramme
PY  - 2023
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
N1  - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023
AB  - Recently, there has been a resurgence of interest in materials with unusual electronic properties such as strong nonlinearity, hysteresis, and memory. This interest is due in part to the end of Moore scaling as well as the emergence of novel computing architectures. Currently, computational performance is limited by the memory bottleneck, as physical memory is not fast or large enough to feed the central processing unit (CPU) pipeline. One alternative is to introduce a new tier of memory that must be substantially faster and more scalable than existing Flash storage. Another approach is to develop schemes that take advantage of in-memory computation, as in the brain-inspired concepts of neuromorphic computing (NC). To reach their full potential, each of these strategies rely on the ability of new classes of memory technologies to exploit physical mechanisms yet to be fully harnessed on an industrial level. This dissertation contains an investigation of two such nascent nanotechnologies in the category of resistive switching (RS). The first, redox-based resistive random access memory (ReRAM), is capable of mimicking biological synapses by allowing storage of large numbers of interconnected and continuously adaptable resistance values. The second technology is based on Cr-doped V₂O₃, a correlated-electron material for which electronic control of Mott insulator-to-metal transitions potentially offers a fast and durable way to emulate the dynamical behavior of neurons. Here, we apply reimagined methods and analysis of electrical measurement to these synaptic and neuronal devices. The newly acquired data sheds further light on the nature of the resistance transitions and is used to design physically validated device models for embedding in large-scale neuromorphic simulations. The measurement circuitry developed here addresses long-standing challenges in the external stabilization of device test structures, and allows (I, V) switching curves to be captured eight orders of magnitude faster than with commercially available equipment, while causing significantly less electrical stress to the measured devices. Applying the measurement system, we introduce a new stochastic device model for solid-state synapses that is trained on a mass quantity of statistical measurement data of ReRAM. This model enables extremely fast (> 10⁸ OPS) and accurate simulations of large synaptic arrays (> 10⁹ cells) and provides a powerful new tool for statistical analysis of resistive switching data. Next, we identify an electro-thermal mechanism behind the negative differential resistance (NDR) and neuronal dynamics observed in (V₁₋ₓCrₓ)₂O₃ nanodevices. We show fast volatile switching (< 10 ns), high switching endurance (> 10¹² cycles), and favorable scaling characteristics in this promising Mott insulating material. A coexisting non-volatile (NV) mechanism is investigated and the conditional occurrence of filamentation in the devices is linked to circuit instabilities, with wider implications for NV switching in other RS materials. The (V₁₋ₓCrₓ)₂O₃ study culminates in a physical model that covers the scaling behavior and threshold adaptability, and is closely fit to observed oscillatory data.
LB  - PUB:(DE-HGF)11
DO  - DOI:10.18154/RWTH-2023-07044
UR  - https://publications.rwth-aachen.de/record/961804
ER  -