h1

%0 Thesis
%A Chekol, Solomon Amsalu
%T Unveiling the relaxation dynamics of Ag/HfO<sub>2</sub> based diffusive memristors for use in neuromorphic computing
%I RWTH Aachen University
%V Dissertation
%C Aachen
%M RWTH-2023-09925
%P 1 Online-Ressource : Illustrationen, Diagramme
%D 2023
%Z Veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2024
%Z Dissertation, RWTH Aachen University, 2023
%X The rapid growth in volume and complexity of data and transfer, driven by advancements in information technologies such as artificial intelligence (AI), cloud computing, big data, and machine learning, is placing significant demands on computation power and speed. Traditional computing architectures are facing challenges in meeting these demands due to the Von Neumann bottleneck, which limits the data transfer rate between the memory and the central processing unit and causes high energy consumption. Today, neuromorphic computing (NC) concepts that mimic the structure and function of the biological brain are gaining popularity as they promise energy-efficient and scalable computing solutions. Currently, neuronal functionality is often performed using a transistor-based neuron, which is area- and energy-inefficient. Therefore, research in the "beyond von Neumann" area is aimed at novel volatile switching components with adjustable switching times, low power consumption, and high scalability, which could potentially be used as artificial neurons in NC circuits. These include threshold-switching devices that switch abruptly from the high-resistance state (HRS) to the low-resistance state (LRS) at a defined voltage. As soon as the applied voltage falls below a certain value, the cell relaxes back to the initial HRS state. In particular, diffusive memristors built from volatile electrochemical metallization (ECM) cells are attracting attention in emerging NC areas such as temporal encoding. These diffusive memristors consist of switching layers made from oxides or chalcogenides sandwiched between an electrochemically active electrode (e.g., Ag or Cu) and an inert electrode (e.g., Pt metal). The cells can be miniaturized down to the sub-micrometer range and the switching itself relies on the formation and dissolution of a metallic filament. Since the temporal behavior of diffusive memristors is their main characteristic, it is of crucial importance to understand the relaxation dynamics of these devices from a physical perspective. This is a prerequisite for optimizing and modulating the performance of diffusive memristors, especially for applications requiring precise control of switching times. Previous approaches mainly describe the relaxation time as a function of the given filament diameter while the filament growth process is not considered. In contrast, the present work takes a comprehensive approach based on the physical description of filament formation and relaxation and a possible dependence between the two. The goal is to develop a deeper understanding of the relaxation dynamics of diffusive memristors, with particular emphasis on the influence of SET parameters on the structure and diameter of the filament formed in each case. To this end, devices were fabricated from an amorphous HfO2 layer of a few nanometers thickness sandwiched between electrode layers of Ag and Pt. For material influence studies, SiO2 as an electrolyte layer and two other metals (TiN and W) as alternative counter electrodes (CEs) were also investigated. By evaluating the transient current response to applied voltage pulses, the SET kinetics and relaxation dynamics of the cells were comprehensively analyzed. This allowed the identification of different mechanisms as rate-limiting steps for filament formation. In particular, three processes, namely nucleation, electron-transfer, and mixed processes (electron-transfer and ion-migration), were identified as rate-limiting steps during the SET process. It was shown that the relaxation time strongly depends on the selected SET parameters. This is explained by the different mechanisms of filament formation during the SET event. A slow SET process at low voltage results in a relatively thick filament that takes long to self-rupture. In contrast, fast SET processes at high voltage lead to dendritic filaments with short relaxation times. This assertion is consistent with previous Monte Carlo simulations for similar ECM cells and is further supported by physics-based simulations using the "JART ECM" model developed in collaboration with IWE 2/RWTH. Furthermore, it has been demonstrated that applying a higher voltage pulse amplitude or longer duration leads to a substantial increase in relaxation time. This could be due to a small residual voltage drop across the diffusive memristor in the LRS, which causes further growth of the filament after closure. The new findings resulting from the combined analysis of SET kinetics and relaxation dynamics underline the importance of filament formation for relaxation behavior and provide important information for optimal operating conditions of these threshold switches in NC circuits. In addition, the work addresses the influence of CE material on the SET kinetics of diffusive memristors. For example, both TiN and W CEs slow down the switching speed compared to the Pt electrode. This is attributed to the different electrocatalytic activity of the different metals in the redox reaction. Oxide formation at the interface of non-noble metals (i.e., TiN and W) is also included in the discussion. An important effect for use in NC circuits is the significant influence of a series resistor on the relaxation behavior. It has been shown that the magnitude of the series resistance can strongly influence the relationship between relaxation time and pulse voltage amplitude. This can be attributed to the highly non-linear nature of the SET kinetics and the role that a residual voltage drop across the device in the ON-state plays in promoting continued growth and strengthening of the conductive filament.In summary, this work shows that diffusive memristors based on volatile ECM cells have a high potential for use as artificial neurons and further applications in the field of NC. The obtained results contribute to a deeper physical understanding of the interplay between filament formation and relaxation and can be directly transferred to the optimization of the operational conditions of diffusive memristors in neuromorphic circuits. This represents an important step towards the realization of energy-efficient NC solutions. In addition, the work classifies volatile devices in terms of different adjustable relaxation times against the background of various NC applications.
%F PUB:(DE-HGF)11
%9 Dissertation / PhD Thesis
%R 10.18154/RWTH-2023-09925
%U https://publications.rwth-aachen.de/record/971953